diff --git a/Makefile b/Makefile
index c8248c45..08e45b04 100644
--- a/Makefile
+++ b/Makefile
@@ -86,6 +86,7 @@ SW_OPT               ?= APP
 SW_LTO               ?= NO
 WS2812_P3_ACCEPTANCE ?= NO
 TINY_P7_ACCEPTANCE   ?= NO
+TINY_P8_ACCEPTANCE   ?= NO
 COREMARK_MODE        ?= quick
 NPU_P5_ACCEPTANCE    ?= NO
 NPU_P6_ACCEPTANCE    ?= NO
@@ -131,6 +132,9 @@ endif
 ifeq ($(TINY_P7_ACCEPTANCE),YES)
 CONFIG_KEY_VARS += TINY_P7_ACCEPTANCE
 endif
+ifeq ($(TINY_P8_ACCEPTANCE),YES)
+CONFIG_KEY_VARS += TINY_P8_ACCEPTANCE
+endif
 VARIANT_ID := $(strip $(shell $(VCS_SHELL_PYTHON) $(ROOT_PATH)/scripts/config_key.py \
     --lock $(LOCK_FILE) --profile $(PROFILE_NAME) --timestamp $(BUILD_TIMESTAMP) \
     $(foreach var,$(CONFIG_KEY_VARS),--value $(var)=$($(var))) | tail -n 1))
@@ -292,6 +296,12 @@ $(call validate_value,SW_OPT,APP O2 O3 Os)
 $(call validate_value,SW_LTO,$(VALID_BOOL))
 $(call validate_value,WS2812_P3_ACCEPTANCE,$(VALID_BOOL))
 $(call validate_value,TINY_P7_ACCEPTANCE,$(VALID_BOOL))
+$(call validate_value,TINY_P8_ACCEPTANCE,$(VALID_BOOL))
+ifeq ($(TINY_P8_ACCEPTANCE),YES)
+ifneq ($(SOC)/$(APP)/$(HAVE_CSR)/$(TINY_P7_ACCEPTANCE),TINY/ci_smoke/YES/NO)
+$(error TINY_P8_ACCEPTANCE requires SOC=TINY APP=ci_smoke HAVE_CSR=YES TINY_P7_ACCEPTANCE=NO)
+endif
+endif
 ifeq ($(TINY_P7_ACCEPTANCE),YES)
 ifneq ($(SOC)/$(APP)/$(HAVE_CSR),TINY/ci_smoke/YES)
 $(error TINY_P7_ACCEPTANCE requires SOC=TINY APP=ci_smoke HAVE_CSR=YES)
@@ -1239,4 +1249,4 @@ regress-nightly:
 	python3 $(ROOT_PATH)/scripts/regress.py --root $(ROOT_PATH) --suite nightly $(if $(filter TINY,$(SOC)),--soc TINY --pdk $(PDK))
 
 sim-asm: asm
-	$(MAKE) SIM_FIRMWARE_NAME=$(ASM_FIRMWARE_NAME) sim
\ No newline at end of file
+	$(MAKE) SIM_FIRMWARE_NAME=$(ASM_FIRMWARE_NAME) sim
diff --git a/app/apps/ci_smoke/tiny.mk b/app/apps/ci_smoke/tiny.mk
index 4478e0e2..f6995fac 100644
--- a/app/apps/ci_smoke/tiny.mk
+++ b/app/apps/ci_smoke/tiny.mk
@@ -4,4 +4,7 @@ APP_SRCS   += $(ROOT_PATH)/app/apps/ci_smoke/mcu_acceptance.S
 APP_SRCS   += $(ROOT_PATH)/app/apps/ci_smoke/tiny_r2_p6.c
 ifeq ($(TINY_P7_ACCEPTANCE),YES)
 APP_SRCS := $(ROOT_PATH)/app/apps/ci_smoke/tiny_r2_p7.c
-endif
\ No newline at end of file
+endif
+ifeq ($(TINY_P8_ACCEPTANCE),YES)
+APP_SRCS := $(ROOT_PATH)/app/apps/ci_smoke/tiny_r2_p8.c
+endif
diff --git a/app/apps/ci_smoke/tiny_r2_p8.c b/app/apps/ci_smoke/tiny_r2_p8.c
new file mode 100644
index 00000000..78613cae
--- /dev/null
+++ b/app/apps/ci_smoke/tiny_r2_p8.c
@@ -0,0 +1,180 @@
+#include <retrosoc/arch/riscv/system_base.h>
+#include <retrosoc/hal/dma.h>
+#include <retrosoc/hal/gpio.h>
+#include <retrosoc/hal/rcu.h>
+#include <retrosoc/hal/sysctrl.h>
+#include <retrosoc/hal/uart.h>
+#include <retrosoc/hal/xpi_psram.h>
+#include <retrosoc/lib/printf.h>
+
+#define RS_P8_WORDS   UINT32_C(256)
+#define RS_P8_ADDRESS UINT32_C(0x54000c00)
+static _Alignas(64) uint32_t rs_p8_source[RS_P8_WORDS];
+static _Alignas(64) uint32_t rs_p8_readback[RS_P8_WORDS];
+static volatile uint32_t rs_p8_cpu_progress;
+
+typedef struct {
+    uint32_t sys_hz, mem_hz;
+    uint32_t isolated_cycles, contended_cycles, progress;
+} rs_p8_measurement_t;
+static rs_p8_measurement_t rs_p8_results[3];
+
+static bool rs_p8_dma(rs_dma_session_t *session, uintptr_t source, uintptr_t destination,
+                      bool contend, uint32_t *cycles) {
+    const rs_dma_config_t config = {.kind = RS_DMA_KIND_MM_TO_MM,
+                                    .request = RS_DMA_REQUEST_SOFTWARE,
+                                    .source = source,
+                                    .destination = destination,
+                                    .byte_count = sizeof(rs_p8_source),
+                                    .width = RS_DMA_WIDTH_32,
+                                    .source_increment = true,
+                                    .destination_increment = true,
+                                    .burst_beats = RS_XPI_PSRAM_BURST_BEATS};
+    rs_dma_status_t state = {0};
+    rs_timeout_t timeout = RS_TIMEOUT_DEFAULT;
+    uint64_t start;
+    if (rs_dma_session_configure(session, &config) != RS_OK ||
+        rs_sysctrl_set_perf_control(true, true, false) != RS_OK) {
+        return false;
+    }
+    __FENCE(rw, rw);
+    start = __get_rv_cycle();
+    if (rs_dma_session_start(session) != RS_OK) {
+        return false;
+    }
+    do {
+        if (contend) {
+            /* Defined competing SRAM work, with no race on the DMA buffers. */
+            rs_p8_cpu_progress++;
+        }
+        if (timeout-- == 0U || rs_dma_get_status(session->channel, &state) != RS_OK) {
+            return false;
+        }
+    } while (state.busy);
+    *cycles = (uint32_t)(__get_rv_cycle() - start);
+    __FENCE(rw, rw);
+    if (rs_sysctrl_set_perf_control(false, false, false) != RS_OK) {
+        return false;
+    }
+    return state.done && !state.error && !state.aborted &&
+           state.bytes_done == sizeof(rs_p8_source) && state.remaining == 0U;
+}
+
+static bool rs_p8_profile(rs_tiny_rcu_profile_t profile, rs_p8_measurement_t *measurement) {
+    rs_tiny_rcu_snapshot_t clocks;
+    rs_xpi_psram_t device;
+    rs_dma_session_t session;
+    volatile uint32_t *const memory = (volatile uint32_t *)(uintptr_t)RS_P8_ADDRESS;
+    volatile uint32_t *const first = (volatile uint32_t *)(uintptr_t)RS_XPI_PSRAM_BASE;
+    volatile uint32_t *const last = (volatile uint32_t *)(uintptr_t)UINT32_C(0x547ffffc);
+    bool pass = true;
+    uint32_t unused_cycles;
+    if (rs_tiny_rcu_request_profile(profile, RS_TIMEOUT_DEFAULT) != RS_OK ||
+        rs_tiny_rcu_get_snapshot(&clocks) != RS_OK ||
+        rs_xpi_psram_initialize(clocks.mem_hz, RS_TIMEOUT_DEFAULT, &device) != RS_OK ||
+        rs_xpi_psram_validate_dma(&device, RS_P8_ADDRESS, sizeof(rs_p8_source), 4U) != RS_OK) {
+        return false;
+    }
+    measurement->sys_hz = clocks.sys_hz;
+    measurement->mem_hz = clocks.mem_hz;
+    *first = 0U;
+    *(volatile uint8_t *)(uintptr_t)(RS_XPI_PSRAM_BASE + 1U) = UINT8_C(0xa5);
+    *(volatile uint16_t *)(uintptr_t)(RS_XPI_PSRAM_BASE + 2U) = UINT16_C(0xbeef);
+    *last = UINT32_C(0x2468ace0);
+    if (*first != UINT32_C(0xbeefa500) || *last != UINT32_C(0x2468ace0)) {
+        return false;
+    }
+    if (rs_dma_session_acquire(&session, RS_DMA_CHANNEL_BULK) != RS_OK) {
+        return false;
+    }
+    for (uint32_t round = 0U; round < 2U && pass; round++) {
+        for (uint32_t i = 0U; i < RS_P8_WORDS; i++) {
+            rs_p8_source[i] = UINT32_C(0x13579bdf) ^ (i * UINT32_C(0x01010101)) ^ round;
+            rs_p8_readback[i] = 0U;
+        }
+        *(volatile uint32_t *)(uintptr_t)(RS_P8_ADDRESS - 4U) = UINT32_C(0x55aa55aa);
+        memory[RS_P8_WORDS] = UINT32_C(0xaa55aa55);
+        rs_p8_cpu_progress = 0U;
+        pass =
+            rs_p8_dma(&session, (uintptr_t)rs_p8_source, RS_P8_ADDRESS, round != 0U,
+                      round == 0U ? &measurement->isolated_cycles : &measurement->contended_cycles);
+        measurement->progress = rs_p8_cpu_progress;
+        for (uint32_t i = 0U; i < RS_P8_WORDS && pass; i++) {
+            pass = memory[i] == rs_p8_source[i];
+        }
+        if (pass) {
+            pass = rs_p8_dma(&session, RS_P8_ADDRESS, (uintptr_t)rs_p8_readback, false,
+                             &unused_cycles);
+        }
+        for (uint32_t i = 0U; i < RS_P8_WORDS && pass; i++) {
+            pass = rs_p8_readback[i] == rs_p8_source[i];
+        }
+        pass = pass &&
+               (*(volatile uint32_t *)(uintptr_t)(RS_P8_ADDRESS - 4U) == UINT32_C(0x55aa55aa)) &&
+               (memory[RS_P8_WORDS] == UINT32_C(0xaa55aa55));
+    }
+    if (!pass && rs_dma_abort_wait(session.channel, RS_TIMEOUT_DEFAULT) != RS_OK) {
+        return false; /* Static buffers remain owned; terminal failure follows. */
+    }
+    return rs_dma_session_release(&session) == RS_OK && pass && measurement->progress != 0U;
+}
+
+int main(void) {
+    const rs_gpio_config_t pin = {.mode = RS_GPIO_MODE_ALT0,
+                                  .pull = RS_GPIO_PULL_NONE,
+                                  .trigger = RS_GPIO_TRIGGER_NONE,
+                                  .output_high = true};
+    rs_tiny_rcu_snapshot_t clocks;
+    rs_xpi_psram_t device;
+    rs_status_t status;
+    const uint32_t nor_word = *(volatile const uint32_t *)(uintptr_t)UINT32_C(0x50000000);
+    uint32_t count = 0U;
+    uint8_t code = 0U;
+    __disable_irq();
+    if (rs_gpio_configure(29U, &pin) != RS_OK || rs_tiny_rcu_get_snapshot(&clocks) != RS_OK) {
+        code = 1U;
+    } else {
+        status = rs_xpi_psram_initialize(clocks.mem_hz, RS_TIMEOUT_DEFAULT, &device);
+        if (status == RS_ENOTSUP && !device.initialized && device.capacity_bytes == 0U) {
+            (void)rs_uart_init(24000000U, 921600U);
+            printf("P8_ABSENT bounded optional-device rejection\n");
+            code =
+                (*(volatile const uint32_t *)(uintptr_t)UINT32_C(0x50000000) == nor_word) ? 0U : 2U;
+        } else if (status != RS_OK) {
+            code = 3U;
+        } else {
+            if (!rs_p8_profile(RS_TINY_RCU_PROFILE_SAFE24, &rs_p8_results[count++])) {
+                code = 4U;
+            }
+            if (code == 0U && (clocks.capability & 12U) == 12U) {
+                if (!rs_p8_profile(RS_TINY_RCU_PROFILE_PLL192, &rs_p8_results[count++]) ||
+                    !rs_p8_profile(RS_TINY_RCU_PROFILE_PLL240, &rs_p8_results[count++])) {
+                    code = 5U;
+                }
+            } else if (code == 0U && clocks.xin_hz != 24000000U) {
+                if (!rs_p8_profile(RS_TINY_RCU_PROFILE_XIN, &rs_p8_results[count++])) {
+                    code = 6U;
+                }
+            }
+            if (code == 0U && rs_tiny_rcu_request_profile(RS_TINY_RCU_PROFILE_SAFE24,
+                                                          RS_TIMEOUT_DEFAULT) != RS_OK) {
+                code = 7U;
+            }
+            if (code == 0U &&
+                *(volatile const uint32_t *)(uintptr_t)UINT32_C(0x50000000) != nor_word) {
+                code = 8U;
+            }
+        }
+    }
+    (void)rs_uart_init(24000000U, 921600U);
+    for (uint32_t i = 0U; i < count; i++) {
+        printf("P8_TRANSPORT sys_hz=%u mem_hz=%u bytes=1024 burst=4 isolated_cycles=%u "
+               "contended_cycles=%u progress=%u\n",
+               rs_p8_results[i].sys_hz, rs_p8_results[i].mem_hz, rs_p8_results[i].isolated_cycles,
+               rs_p8_results[i].contended_cycles, rs_p8_results[i].progress);
+    }
+    printf("P8_RESULT code=%u profiles=%u\n", code, count);
+    rs_sysctrl_write_test_status(code == 0U, code);
+    for (;;) {
+    }
+}
diff --git a/crt/include/retrosoc/hal/xpi_psram.h b/crt/include/retrosoc/hal/xpi_psram.h
new file mode 100644
index 00000000..c267ce19
--- /dev/null
+++ b/crt/include/retrosoc/hal/xpi_psram.h
@@ -0,0 +1,29 @@
+#ifndef RETROSOC_HAL_XPI_PSRAM_H
+#define RETROSOC_HAL_XPI_PSRAM_H
+
+#include <retrosoc/hal/xpi.h>
+
+#define RS_XPI_PSRAM_BASE        UINT32_C(0x54000000)
+#define RS_XPI_PSRAM_BYTES       UINT32_C(0x00800000)
+#define RS_XPI_PSRAM_BURST_BEATS UINT8_C(4)
+
+typedef struct {
+    bool initialized;
+    uint32_t memory_clock_hz;
+    uint32_t capacity_bytes;
+    uint8_t id[6];
+} rs_xpi_psram_t;
+
+/* ESP-PSRAM64H reference profile: slot 1, LUT 8..11, QPI SDR, SCK=MEM/2.
+ * Caller owns XPI exclusively, routes NSS1, and executes code/IRQs from SRAM.
+ * Clock changes/reset invalidate this software session; explicitly initialize
+ * again before using memory. This is not physical part/rate qualification. */
+rs_status_t rs_xpi_psram_timing(uint32_t memory_clock_hz, rs_xpi_slot_config_t *config);
+rs_status_t rs_xpi_psram_initialize(uint32_t memory_clock_hz, rs_timeout_t timeout,
+                                    rs_xpi_psram_t *device);
+/* Direct DMA uses SOFTWARE request and mapped addresses, not XPI TX/RX FIFO.
+ * Align bases to 16 bytes for four-beat bursts; the DMA splits at 4 KiB. */
+rs_status_t rs_xpi_psram_validate_dma(const rs_xpi_psram_t *device, uint32_t address,
+                                      uint32_t bytes, uint8_t burst_beats);
+
+#endif
diff --git a/crt/src/hal/xpi_psram.c b/crt/src/hal/xpi_psram.c
new file mode 100644
index 00000000..d64f3c0c
--- /dev/null
+++ b/crt/src/hal/xpi_psram.c
@@ -0,0 +1,147 @@
+#include <retrosoc/hal/clint.h>
+#include <retrosoc/hal/xpi_psram.h>
+
+#define RS_XPI_PSRAM_SLOT             UINT8_C(1)
+#define RS_XPI_PSRAM_COMMAND_SEQUENCE UINT8_C(10)
+#define RS_XPI_PSRAM_ID_SEQUENCE      UINT8_C(11)
+
+static rs_status_t rs_xpi_psram_delay(uint32_t microseconds, rs_timeout_t timeout) {
+    uint64_t first, now;
+    const uint64_t ticks =
+        ((uint64_t)microseconds * RS_CLINT_TIMEBASE_HZ + UINT64_C(999999)) / UINT64_C(1000000);
+    if (rs_clint_get_time(&first) != RS_OK) {
+        return RS_EIO;
+    }
+    while (timeout-- != 0U) {
+        if (rs_clint_get_time(&now) != RS_OK) {
+            return RS_EIO;
+        }
+        if (now - first >= ticks) {
+            return RS_OK;
+        }
+    }
+    return RS_ETIMEOUT;
+}
+
+static rs_status_t rs_xpi_psram_command(uint8_t opcode, rs_xpi_pads_t pads, rs_timeout_t timeout) {
+    const uint16_t lut[] = {rs_xpi_instruction(RS_XPI_INSTR_COMMAND, pads, opcode), 0U};
+    const rs_xpi_transfer_t transfer = {.slot = RS_XPI_PSRAM_SLOT,
+                                        .sequence = RS_XPI_PSRAM_COMMAND_SEQUENCE};
+    rs_status_t status = rs_xpi_enable(false, timeout);
+    if (status == RS_OK) {
+        status = rs_xpi_lut_write(RS_XPI_PSRAM_COMMAND_SEQUENCE, lut, 2U, timeout);
+    }
+    if (status == RS_OK) {
+        status = rs_xpi_enable(true, timeout);
+    }
+    if (status == RS_OK) {
+        status = rs_xpi_transfer(&transfer, timeout);
+    }
+    return status;
+}
+
+rs_status_t rs_xpi_psram_initialize(uint32_t memory_clock_hz, rs_timeout_t timeout,
+                                    rs_xpi_psram_t *device) {
+    const uint16_t read_lut[] = {UINT16_C(0x18eb), UINT16_C(0x2818), UINT16_C(0x4006),
+                                 UINT16_C(0x6800), 0U};
+    const uint16_t write_lut[] = {UINT16_C(0x1838), UINT16_C(0x2818), UINT16_C(0x5800), 0U};
+    const uint16_t id_lut[] = {UINT16_C(0x109f), UINT16_C(0x2018), UINT16_C(0x6000), 0U};
+    rs_xpi_slot_config_t config;
+    rs_xpi_transfer_t transfer = {
+        .slot = RS_XPI_PSRAM_SLOT, .sequence = RS_XPI_PSRAM_ID_SEQUENCE, .byte_count = 3U};
+    rs_status_t status;
+    if (device == NULL) {
+        return RS_EINVAL;
+    }
+    *device = (rs_xpi_psram_t){0};
+    status = rs_xpi_psram_timing(memory_clock_hz, &config);
+    if (status != RS_OK) {
+        return status;
+    }
+    status = rs_xpi_probe();
+    if (status != RS_OK) {
+        return status;
+    }
+    /* Reject locks before issuing any potentially faulting configuration write. */
+    if ((RS_XPI_REG(RS_XPI_REG_CONFIG_LOCK) &
+         (RS_XPI_CONFIG_LOCK_GLOBAL | RS_XPI_CONFIG_LOCK_SLOT1 | RS_XPI_CONFIG_LOCK_LUT)) != 0U) {
+        return RS_EIO;
+    }
+    status = rs_xpi_enable(false, timeout);
+    if (status != RS_OK) {
+        return status;
+    }
+    status = rs_xpi_slot_configure(RS_XPI_PSRAM_SLOT, &config, timeout);
+    if (status == RS_OK) {
+        status = rs_xpi_lut_write(8U, read_lut, 5U, timeout);
+    }
+    if (status == RS_OK) {
+        status = rs_xpi_lut_write(9U, write_lut, 4U, timeout);
+    }
+    if (status == RS_OK) {
+        status = rs_xpi_lut_write(RS_XPI_PSRAM_ID_SEQUENCE, id_lut, 4U, timeout);
+    }
+    if (status == RS_OK) {
+        status = rs_xpi_enable(true, timeout);
+    }
+    if (status == RS_OK) {
+        status = rs_xpi_psram_delay(150U, timeout);
+    }
+    /* Exit QPI is harmless when the device is already in SPI reset mode. */
+    if (status == RS_OK) {
+        status = rs_xpi_psram_command(UINT8_C(0xf5), RS_XPI_PADS_4, timeout);
+    }
+    if (status == RS_OK) {
+        status = rs_xpi_psram_command(UINT8_C(0x66), RS_XPI_PADS_1, timeout);
+    }
+    if (status == RS_OK) {
+        status = rs_xpi_psram_command(UINT8_C(0x99), RS_XPI_PADS_1, timeout);
+    }
+    if (status == RS_OK) {
+        status = rs_xpi_psram_delay(1U, timeout);
+    }
+    if (status == RS_OK) {
+        transfer.rx_data = device->id;
+        status = rs_xpi_transfer(&transfer, timeout);
+    }
+    if ((status == RS_OK) &&
+        ((device->id[0] != UINT8_C(0x0d)) || (device->id[1] != UINT8_C(0x5d)) ||
+         (device->id[2] != UINT8_C(0x26)))) {
+        status = RS_ENOTSUP;
+    }
+    if (status == RS_OK) {
+        status = rs_xpi_psram_command(UINT8_C(0x35), RS_XPI_PADS_1, timeout);
+    }
+    if (status == RS_OK) {
+        status = rs_xpi_enable(false, timeout);
+    }
+    if (status == RS_OK) {
+        config.memory_read_enable = true;
+        config.memory_write_enable = true;
+        status = rs_xpi_slot_configure(RS_XPI_PSRAM_SLOT, &config, timeout);
+    }
+    if (status != RS_OK) {
+        /* Keep the first error snapshot. Never publish capacity after a failed
+         * probe, and leave mapped NSS1 disabled while restoring NOR access. */
+        rs_status_t cleanup = rs_xpi_abort(timeout);
+        if (cleanup == RS_OK) {
+            cleanup = rs_xpi_enable(false, timeout);
+        }
+        config.memory_read_enable = false;
+        config.memory_write_enable = false;
+        if (cleanup == RS_OK) {
+            cleanup = rs_xpi_slot_configure(RS_XPI_PSRAM_SLOT, &config, timeout);
+        }
+        if (cleanup == RS_OK) {
+            cleanup = rs_xpi_enable(true, timeout);
+        }
+        return (cleanup == RS_OK) ? status : cleanup;
+    }
+    status = rs_xpi_enable(true, timeout);
+    if (status == RS_OK) {
+        device->capacity_bytes = RS_XPI_PSRAM_BYTES;
+        device->memory_clock_hz = memory_clock_hz;
+        device->initialized = true;
+    }
+    return status;
+}
diff --git a/crt/src/hal/xpi_psram_math.c b/crt/src/hal/xpi_psram_math.c
new file mode 100644
index 00000000..66a2da7c
--- /dev/null
+++ b/crt/src/hal/xpi_psram_math.c
@@ -0,0 +1,67 @@
+#include <retrosoc/hal/xpi_psram.h>
+
+rs_status_t rs_xpi_psram_timing(uint32_t memory_clock_hz, rs_xpi_slot_config_t *config) {
+    if ((config == NULL) ||
+        ((memory_clock_hz != UINT32_C(24000000)) && (memory_clock_hz != UINT32_C(48000000)) &&
+         (memory_clock_hz != UINT32_C(96000000)) && (memory_clock_hz != UINT32_C(120000000)))) {
+        return RS_EINVAL;
+    }
+    config->enable = true;
+    config->memory_read_enable = false;
+    config->memory_write_enable = false;
+    config->mode3 = false;
+    config->device_size = RS_XPI_PSRAM_BYTES;
+    config->read_sequence = UINT8_C(8);
+    config->write_sequence = UINT8_C(9);
+    config->clock_divider = UINT8_C(0);
+    config->cs_setup_cycles = UINT8_C(1);
+    config->cs_hold_cycles =
+        (uint8_t)(((uint64_t)memory_clock_hz * UINT64_C(20) + UINT64_C(999999999)) /
+                  UINT64_C(1000000000));
+    config->cs_high_cycles =
+        (uint8_t)(((uint64_t)memory_clock_hz * UINT64_C(50) + UINT64_C(999999999)) /
+                  UINT64_C(1000000000));
+    /* Includes core setup/hold/high and consumer stalls. Leave two MEM edges
+     * of margin below the reference device's 8 us active-CS limit. */
+    config->timeout_cycles = (memory_clock_hz / UINT32_C(125000)) - UINT32_C(2);
+    config->burst_boundary = UINT32_C(1024);
+    return RS_OK;
+}
+
+rs_status_t rs_xpi_psram_validate_dma(const rs_xpi_psram_t *device, uint32_t address,
+                                      uint32_t bytes, uint8_t burst_beats) {
+    uint32_t offset;
+    uint32_t cursor;
+    const uint32_t burst_bytes = (uint32_t)burst_beats * UINT32_C(4);
+    if (device == NULL) {
+        return RS_EINVAL;
+    }
+    if (!device->initialized || (device->capacity_bytes != RS_XPI_PSRAM_BYTES)) {
+        return RS_ENOTSUP;
+    }
+    if ((address < RS_XPI_PSRAM_BASE) || (bytes == 0U) || ((bytes & 3U) != 0U) ||
+        ((burst_beats != 1U) && (burst_beats != 2U) && (burst_beats != 4U)) ||
+        ((address & ((uint32_t)burst_beats * 4U - 1U)) != 0U)) {
+        return RS_EINVAL;
+    }
+    offset = address - RS_XPI_PSRAM_BASE;
+    if ((offset >= device->capacity_bytes) || (bytes > device->capacity_bytes - offset)) {
+        return RS_EINVAL;
+    }
+    /* The XPI slot rejects a serial-boundary crossing command, and the DMA
+     * itself cannot cross a 4 KiB AXI page. Validate every emitted burst before
+     * admitting the session; do not rely on the target to fail mid-frame. */
+    for (cursor = 0U; cursor < bytes; cursor += burst_bytes) {
+        uint32_t length = bytes - cursor;
+        const uint32_t serial_offset = (offset + cursor) & UINT32_C(0x3ff);
+        const uint32_t page_offset = (offset + cursor) & UINT32_C(0xfff);
+        if (length > burst_bytes) {
+            length = burst_bytes;
+        }
+        if ((length > UINT32_C(0x400) - serial_offset) ||
+            (length > UINT32_C(0x1000) - page_offset)) {
+            return RS_EINVAL;
+        }
+    }
+    return RS_OK;
+}
diff --git a/rtl/ip/storage/xpi_mm.sv b/rtl/ip/storage/xpi_mm.sv
index ffa13707..9194e542 100644
--- a/rtl/ip/storage/xpi_mm.sv
+++ b/rtl/ip/storage/xpi_mm.sv
@@ -83,6 +83,7 @@ module xpi_mm (
   logic [1:0] s_resp_d, s_resp_q;
   logic [31:0] s_rdata_d, s_rdata_q;
   logic s_rvalid_d, s_rvalid_q;
+  logic s_word_complete_d, s_word_complete_q;
   logic [2:0] s_byte_idx_d, s_byte_idx_q;
   logic [6:0] s_cursor_d, s_cursor_q;
   logic [6:0] s_run_end_d, s_run_end_q;
@@ -198,7 +199,7 @@ module xpi_mm (
       axi4.awburst, axi4.awlen
   ) && !axi4.awlock && (axi4.awsize <= `AXI4_BURST_SIZE_4BYTES) &&
       ((axi4.awaddr & ((32'd1 << axi4.awsize) - 1'b1)) == 32'd0) &&
-      (axi4.awaddr[31:12] == s_write_last_addr[31:12]);
+      (axi4.awaddr[31:12] == s_write_last_addr[31:12]) && s_write_boundary_ok;
 
   assign s_legal_read = address_supported(
       axi4.araddr
@@ -216,7 +217,7 @@ module xpi_mm (
       axi4.arburst, axi4.arlen
   ) && !axi4.arlock && (axi4.arsize <= `AXI4_BURST_SIZE_4BYTES) &&
       ((axi4.araddr & ((32'd1 << axi4.arsize) - 1'b1)) == 32'd0) &&
-      (axi4.araddr[31:12] == s_read_last_addr[31:12]);
+      (axi4.araddr[31:12] == s_read_last_addr[31:12]) && s_read_boundary_ok;
 
   assign axi4.awready = (s_state_q == Idle) && accept_enable_i && !axi4.arvalid;
   assign axi4.arready = (s_state_q == Idle) && accept_enable_i;
@@ -247,7 +248,7 @@ module xpi_mm (
                            (16'd1 << s_size_q));
 
   assign core_tx_valid_o = s_state_q == WriteStream;
-  assign core_rx_ready_o = (s_state_q == ReadData) && !s_rvalid_q;
+  assign core_rx_ready_o = (s_state_q == ReadData) && !s_word_complete_q;
   assign core_tx_data_o = s_write_buffer_q[s_tx_beat][(s_tx_lane*8)+:8];
 
   assign s_err_code_q = xpi_error_e'(s_err_code_bits_q);
@@ -306,8 +307,13 @@ module xpi_mm (
   end
 
   always_comb begin
-    s_write_lane_mask = (4'd1 << s_size_q) - 1'b1;
-    s_read_lane       = {1'b0, s_addr_q[1:0]} + s_byte_idx_q;
+    unique case (s_size_q)
+      3'd0:    s_write_lane_mask = 4'b0001;
+      3'd1:    s_write_lane_mask = 4'b0011;
+      3'd2:    s_write_lane_mask = 4'b1111;
+      default: s_write_lane_mask = 4'b0000;
+    endcase
+    s_read_lane = {1'b0, s_addr_q[1:0]} + s_byte_idx_q;
   end
 
   for (genvar write_index = 0; write_index < 16; write_index++) begin : write_buffer_block
@@ -339,6 +345,7 @@ module xpi_mm (
     s_resp_d                = s_resp_q;
     s_rdata_d               = s_rdata_q;
     s_rvalid_d              = s_rvalid_q;
+    s_word_complete_d       = s_word_complete_q;
     s_byte_idx_d            = s_byte_idx_q;
     s_cursor_d              = s_cursor_q;
     s_run_end_d             = s_run_end_q;
@@ -362,6 +369,7 @@ module xpi_mm (
         s_resp_d            = `AXI4_RESP_OKAY;
         s_rdata_d           = '0;
         s_rvalid_d          = 1'b0;
+        s_word_complete_d   = 1'b0;
         s_byte_idx_d        = '0;
         s_cursor_d          = '0;
         s_frame_done_d      = 1'b0;
@@ -423,37 +431,45 @@ module xpi_mm (
           s_rdata_d[(s_read_lane*8)+:8] = core_rx_data_i;
           perf_read_byte_event_o        = 1'b1;
           if (s_byte_idx_q == ((3'd1 << s_size_q) - 1'b1)) begin
-            s_byte_idx_d = '0;
-            s_rvalid_d   = 1'b1;
+            s_byte_idx_d      = '0;
+            s_word_complete_d = 1'b1;
+            // Hold the physical command's last beat until STOP/CS completion
+            // so a late sequence error cannot follow a successful AXI end.
+            s_rvalid_d        = s_coalesce_q && (s_beat_q != s_len_q);
           end else begin
             s_byte_idx_d = s_byte_idx_q + 1'b1;
           end
         end
         if (core_done_i) begin
           s_frame_done_d = 1'b1;
-          if (core_error_i) begin
-            s_resp_d            = `AXI4_RESP_SLAVE_ERROR;
+          if (core_error_i || !s_word_complete_q || (s_coalesce_q && (s_beat_q != s_len_q))) begin
             s_transaction_err_d = 1'b1;
-            s_err_code_d        = core_error_code_i;
+            s_err_code_d        = core_error_i ? core_error_code_i : XpiErrorSequence;
             s_err_addr_d        = s_addr_q;
             s_err_slot_d        = s_slot_q;
             s_err_pc_d          = core_error_pc_i;
             axi_error_event_o   = 1'b1;
             if (!s_rvalid_q) begin
-              s_rdata_d  = '0;
-              s_rvalid_d = 1'b1;
+              s_resp_d          = `AXI4_RESP_SLAVE_ERROR;
+              s_rdata_d         = '0;
+              s_rvalid_d        = 1'b1;
+              s_word_complete_d = 1'b1;
             end
+          end else begin
+            s_rvalid_d = 1'b1;
           end
         end
         if (s_rvalid_q && axi4.rready) begin
-          s_rvalid_d = 1'b0;
-          s_rdata_d  = '0;
+          s_rvalid_d        = 1'b0;
+          s_word_complete_d = 1'b0;
+          s_rdata_d         = '0;
           if (s_beat_q == s_len_q) begin
             s_state_d = (s_frame_done_q || core_done_i) ? Idle : ReadFinish;
           end else begin
             s_addr_d = s_next_addr;
             s_beat_d = s_beat_q + 1'b1;
-            if (s_transaction_err_q || core_error_i) begin
+            if (s_transaction_err_d) begin
+              s_resp_d  = `AXI4_RESP_SLAVE_ERROR;
               s_state_d = ReadErrorResponse;
             end else if (s_coalesce_q) begin
               s_state_d = ReadData;
@@ -516,10 +532,12 @@ module xpi_mm (
             axi_error_event_o = 1'b1;
             s_state_d         = axi4.wlast ? WriteResponse : WriteDrain;
           end else if (s_beat_q == s_len_q) begin
+            s_addr_d   = s_start_addr_q;
             s_beat_d   = '0;
             s_cursor_d = '0;
             s_state_d  = WriteScan;
           end else begin
+            s_addr_d = s_next_addr;
             s_beat_d = s_beat_q + 1'b1;
           end
         end
@@ -549,7 +567,17 @@ module xpi_mm (
       end
 
       WriteStream: begin
-        if (core_tx_valid_o && core_tx_ready_i) begin
+        // The core can time out or abort before consuming all buffered bytes.
+        // AXI W is already drained; complete B rather than wait for a dead core.
+        if (core_done_i) begin
+          s_resp_d          = `AXI4_RESP_SLAVE_ERROR;
+          s_err_code_d      = core_error_i ? core_error_code_i : XpiErrorSequence;
+          s_err_addr_d      = s_addr_q;
+          s_err_slot_d      = s_slot_q;
+          s_err_pc_d        = core_error_pc_i;
+          axi_error_event_o = 1'b1;
+          s_state_d         = WriteResponse;
+        end else if (core_tx_valid_o && core_tx_ready_i) begin
           perf_write_byte_event_o = 1'b1;
           s_cursor_d              = s_cursor_q + 1'b1;
           if ((s_cursor_q + 1'b1) >= s_run_end_q) begin
@@ -609,6 +637,14 @@ module xpi_mm (
       .dat_i  (s_state_d),
       .dat_o  (s_state_bits_q)
   );
+  dffr #(
+      .DATA_WIDTH(1)
+  ) u_word_complete_dffr (
+      .clk_i  (clk_i),
+      .rst_n_i(rst_n_i),
+      .dat_i  (s_word_complete_d),
+      .dat_o  (s_word_complete_q)
+  );
   dffr #(
       .DATA_WIDTH(32)
   ) u_addr_dffr (
diff --git a/rtl/mini/filelist/tb.fl b/rtl/mini/filelist/tb.fl
index fe254d66..c4064997 100644
--- a/rtl/mini/filelist/tb.fl
+++ b/rtl/mini/filelist/tb.fl
@@ -3,7 +3,7 @@
 /model/uart_model.sv
 /device_model/mic.sv
 /device_model/w25q128jvxim.sv
-/model/ESP_PSRAM64H.sv
+/device_model/ESP_PSRAM64H.sv
 /model/DVP_CAMERA.sv
 /third_party/model/sdram/sdr.v
 /tb/retrosoc_tb.sv
diff --git a/rtl/mini/mk/verilator.mk b/rtl/mini/mk/verilator.mk
index 3889b9b2..0feea6f4 100644
--- a/rtl/mini/mk/verilator.mk
+++ b/rtl/mini/mk/verilator.mk
@@ -11,7 +11,7 @@ SOC_VSRC_HOME    += $(RTL_PATH)/dv/verilator/rtl
 SOC_COMPILE_HOME := $(BUILD_DIR)/emu_compile
 
 SOC_VXXFILES      := $(RTL_FLIST)
-SOC_VXXFILES      += $(RTL_PATH)/dv/model/ESP_PSRAM64H.sv
+SOC_VXXFILES      += $(ROOT_PATH)/rtl/model/ESP_PSRAM64H.sv
 SOC_VXXFILES      += $(RTL_PATH)/dv/verilator/rtl/flash_read_binder.sv
 SOC_VXXFILES      += $(RTL_PATH)/dv/verilator/rtl/flash_read_byte_binder.sv
 SOC_VXXFILES      += $(RTL_PATH)/dv/verilator/rtl/QSPIFlash.sv
@@ -51,7 +51,7 @@ VERILATOR_DEPFILE       := $(BUILD_DIR)/verilate.d
 VERILATOR_EMU           := $(BUILD_DIR)/emu
 RTL_LINT_STAMP          := $(RTL_LINT_DIR)/rtl-lint.stamp
 RTL_LINT_DEPFILE        := $(RTL_LINT_DIR)/rtl-lint.d
-VERILATOR_EXTRA_SOURCES := $(RTL_PATH)/dv/model/ESP_PSRAM64H.sv \
+VERILATOR_EXTRA_SOURCES := $(ROOT_PATH)/rtl/model/ESP_PSRAM64H.sv \
                            $(RTL_PATH)/dv/verilator/rtl/flash_read_binder.sv \
                            $(RTL_PATH)/dv/verilator/rtl/flash_read_byte_binder.sv \
                            $(RTL_PATH)/dv/verilator/rtl/QSPIFlash.sv \
@@ -144,4 +144,4 @@ clean:
 	python3 $(ROOT_PATH)/scripts/clean.py --root $(ROOT_PATH) --path $(BUILD_DIR)
 	python3 $(ROOT_PATH)/scripts/clean.py --root $(ROOT_PATH) --path $(RTL_LINT_DIR)
 
-.PHONY: lint rtl-lint check-rtl-lint comp sim debug-sim clean
\ No newline at end of file
+.PHONY: lint rtl-lint check-rtl-lint comp sim debug-sim clean
diff --git a/rtl/mk/software.mk b/rtl/mk/software.mk
index 741dc220..b2e76588 100644
--- a/rtl/mk/software.mk
+++ b/rtl/mk/software.mk
@@ -168,6 +168,8 @@ CRT_SRCS := $(ROOT_PATH)/crt/arch/riscv/startup.S \
             $(ROOT_PATH)/crt/src/hal/spisd_math.c \
             $(ROOT_PATH)/crt/src/hal/spisd.c \
             $(ROOT_PATH)/crt/src/hal/xpi.c \
+            $(ROOT_PATH)/crt/src/hal/xpi_psram_math.c \
+            $(ROOT_PATH)/crt/src/hal/xpi_psram.c \
             $(ROOT_PATH)/crt/src/hal/fabric_monitor.c \
             $(ROOT_PATH)/crt/src/hal/resource.c \
             $(ROOT_PATH)/crt/src/hal/extension.c \
@@ -180,7 +182,7 @@ CFLAGS         += -DRS_SOC_TINY
 TINY_CRT_NAMES := startup.S clzsi2.c divdi3.c ffssi2.c udivdi3.c umoddi3.c \
     stdlib.c string.c console.c printf.c archinfo.c clint.c uart_math.c uart.c \
     gpio_math.c gpio.c timer_math.c timer.c pwm.c rtc.c wdg.c watchdog.c i2c_math.c i2c.c \
-    dma_math.c dma.c onchip_sram.c xpi.c rng.c crc.c ws2812_math.c ws2812.c \
+    dma_math.c dma.c onchip_sram.c xpi.c xpi_psram.c xpi_psram_math.c rng.c crc.c ws2812_math.c ws2812.c \
     crypto.c crypto_lifecycle.c i2s_math.c i2s.c sdio_math.c sdio.c rcu.c
 CRT_SRCS       := $(foreach source,$(CRT_SRCS),$(if $(filter $(notdir $(source)),$(TINY_CRT_NAMES)),$(source)))
 CRT_SRCS       += $(ROOT_PATH)/crt/src/hal/sysctrl.c
@@ -305,4 +307,4 @@ firmware: $(FIRMWARE_ELF)
 
 firmware asm: | manifest
 
-.PHONY: FORCE_VERSION
\ No newline at end of file
+.PHONY: FORCE_VERSION
diff --git a/rtl/mini/dv/model/ESP_PSRAM64H.sv b/rtl/model/ESP_PSRAM64H.sv
similarity index 99%
rename from rtl/mini/dv/model/ESP_PSRAM64H.sv
rename to rtl/model/ESP_PSRAM64H.sv
index 7ad94fb7..72f0162a 100644
--- a/rtl/mini/dv/model/ESP_PSRAM64H.sv
+++ b/rtl/model/ESP_PSRAM64H.sv
@@ -8,7 +8,7 @@
 // MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE.
 // See the Mulan PSL v2 for more details.
 //
-// Behavioral model for the external ESP-PSRAM64H 64-Mbit PSRAM.
+// Shared behavioral model for the external ESP-PSRAM64H 64-Mbit PSRAM.
 // Only SDR SPI and QPI transactions are modeled.  The model intentionally
 // does not initialize the memory array unless INITIALIZE_MEMORY is enabled.
 
diff --git a/rtl/tiny/dv/retrosoc_tiny_tb.sv b/rtl/tiny/dv/retrosoc_tiny_tb.sv
index 1cef5b6f..72a9c2f3 100644
--- a/rtl/tiny/dv/retrosoc_tiny_tb.sv
+++ b/rtl/tiny/dv/retrosoc_tiny_tb.sv
@@ -23,13 +23,28 @@ module retrosoc_tiny_tb;
   wire  s_jtag_trst_n = s_rst_n;
   wire  s_jtag_tdo;
   `include "tiny_tb_gpio.svh"
-  wire          s_xpi_sck;
-  wire    [3:0] s_xpi_nss;
-  tri     [3:0] s_xpi_data;
-  integer       s_cycles;
+  wire s_xpi_sck;
+  wire [3:0] s_xpi_nss;
+  tri1 [3:0] s_xpi_data;
+  // Optional board population; the unchanged default has no PSRAM dependency.
+  wire s_psram_csn = ($test$plusargs(
+      "tiny_p8_psram"
+  ) && !$test$plusargs(
+      "p8_absent"
+  )) ? s_gpio[29] : 1'b1;
+  ESP_PSRAM64H #(
+      .TIMING_VIOLATION_FATAL(1),
+      .MIN_CS_HIGH_NS        (50)
+  ) u_psram (
+      .sclk(s_xpi_sck),
+      .csn (s_psram_csn),
+      .sio (s_xpi_data)
+  );
+  integer s_cycles;
 `ifndef RETROSOC_SOC__TINY_NETLIST
   `include "tiny_p6_devices.svh"
   `include "tiny_p7_observers.svh"
+  `include "tiny_p8_observers.svh"
 `endif
   integer       s_max_cycles = 20000000;
   integer       s_done_cycles = 0;
diff --git a/rtl/tiny/dv/tiny_p8_observers.svh b/rtl/tiny/dv/tiny_p8_observers.svh
new file mode 100644
index 00000000..649629f0
--- /dev/null
+++ b/rtl/tiny/dv/tiny_p8_observers.svh
@@ -0,0 +1,63 @@
+// Verification-only MEM-clock observations. No DUT state is driven.
+`define _P8_AXI u_dut.u_soc.u_xpi_mem_axi4_if
+logic s_p8_previous_enable = 0;
+logic s_p8_finished = 0;
+longint unsigned s_p8_cycles = 0, s_p8_write_bytes = 0, s_p8_read_bytes = 0;
+longint unsigned s_p8_stall_run[5], s_p8_stall_max[5];
+longint unsigned s_p8_all_writes = 0;
+integer s_p8_window = 0;
+wire [4:0] s_p8_stalls = {
+  `_P8_AXI.awvalid && !`_P8_AXI.awready,
+  `_P8_AXI.wvalid && !`_P8_AXI.wready,
+  `_P8_AXI.arvalid && !`_P8_AXI.arready,
+  `_P8_AXI.bvalid && !`_P8_AXI.bready,
+  `_P8_AXI.rvalid && !`_P8_AXI.rready
+};
+always @(posedge u_dut.u_soc.s_backend_mem_clk) begin
+  if ($test$plusargs("tiny_p8_psram") && u_dut.u_soc.s_xpi_rst_n) begin
+    if (`_P8_AXI.wvalid && `_P8_AXI.wready) s_p8_all_writes++;
+    if (u_dut.u_soc.s_perf_enable) begin
+      if (!s_p8_previous_enable) begin
+        s_p8_cycles      = 0;
+        s_p8_write_bytes = 0;
+        s_p8_read_bytes  = 0;
+        for (int i = 0; i < 5; i++) begin
+          s_p8_stall_run[i] = 0;
+          s_p8_stall_max[i] = 0;
+        end
+      end
+      s_p8_cycles++;
+      if (`_P8_AXI.wvalid && `_P8_AXI.wready) s_p8_write_bytes += $countones(`_P8_AXI.wstrb);
+      if (`_P8_AXI.rvalid && `_P8_AXI.rready) s_p8_read_bytes += 4;
+      for (int i = 0; i < 5; i++) begin
+        s_p8_stall_run[i] = s_p8_stalls[i] ? s_p8_stall_run[i] + 1 : 0;
+        if (s_p8_stall_run[i] > s_p8_stall_max[i]) s_p8_stall_max[i] = s_p8_stall_run[i];
+      end
+    end else if (s_p8_previous_enable) begin
+      $display(
+          "P8_XPI_WINDOW index=%0d mem_cycles=%0d write_bytes=%0d read_bytes=%0d cs_max=%0d aw_max=%0d w_max=%0d ar_max=%0d b_max=%0d r_max=%0d",
+          s_p8_window, s_p8_cycles, s_p8_write_bytes, s_p8_read_bytes, s_p8_cs_max,
+          s_p8_stall_max[0], s_p8_stall_max[1], s_p8_stall_max[2], s_p8_stall_max[3],
+          s_p8_stall_max[4]);
+      s_p8_window++;
+    end
+    s_p8_previous_enable = u_dut.u_soc.s_perf_enable;
+    if (u_dut.s_test_done && !s_p8_finished) begin
+      s_p8_finished = 1;
+      if (!$test$plusargs("p8_absent") && (s_p8_window < 4 || s_p8_all_writes == 0))
+        $fatal(1, "SIM_TEST_FAIL P8 missing mapped transport evidence");
+      if ($test$plusargs("p8_absent") && s_p8_all_writes != 0)
+        $fatal(1, "SIM_TEST_FAIL P8 absent device exposed mapped writes");
+    end
+  end
+end
+`undef _P8_AXI
+
+time s_p8_cs_start, s_p8_cs_max = 0;
+always @(negedge s_psram_csn) s_p8_cs_start = $time;
+always @(posedge s_psram_csn) begin
+  if ($test$plusargs("tiny_p8_psram") && s_p8_cs_start != 0) begin
+    if ($time - s_p8_cs_start > 8000) $fatal(1, "SIM_TEST_FAIL P8 active CS exceeds 8 us");
+    if ($time - s_p8_cs_start > s_p8_cs_max) s_p8_cs_max = $time - s_p8_cs_start;
+  end
+end
diff --git a/rtl/tiny/filelist/tb.fl b/rtl/tiny/filelist/tb.fl
index da7d4e7d..37be89bf 100644
--- a/rtl/tiny/filelist/tb.fl
+++ b/rtl/tiny/filelist/tb.fl
@@ -1,3 +1,4 @@
 /dv/tiny_qspi_nor.sv
+/device_model/ESP_PSRAM64H.sv
 /tests/rtl/sdio_native_model.sv
 /dv/retrosoc_tiny_tb.sv
diff --git a/scripts/generate_tiny.py b/scripts/generate_tiny.py
index 22d04c40..18af8a3b 100644
--- a/scripts/generate_tiny.py
+++ b/scripts/generate_tiny.py
@@ -225,7 +225,8 @@ def generate_filelists(output: Path, generated: Path, defines: list[str],
     prefixes = {"top": TINY / "top", "dv": TINY / "dv", "ip": ROOT / "rtl/ip",
                 "tech": ROOT / "rtl/tech", "clusterip": ROOT / "rtl/managed/clusterip",
                 "hazard3": ROOT / "rtl/managed/hazard3/hdl", "physical": ROOT / "physical",
-                "cache": ROOT / ".cache", "tests": ROOT / "tests"}
+                "cache": ROOT / ".cache", "tests": ROOT / "tests",
+                "device_model": ROOT / "rtl/model"}
     suffix = "_yosys" if synthesis == "YOSYS" else "_verilator" if simulator == "VERILATOR" else ""
     pdk_name = "pdk_ihp130.fl" if pdk == "IHP130" else f"pdk_ics55{suffix}.fl"
     templates = [*sorted((TINY / "filelist").glob("*.fl")),
diff --git a/scripts/run_c_tests.py b/scripts/run_c_tests.py
index fe896652..942c7148 100644
--- a/scripts/run_c_tests.py
+++ b/scripts/run_c_tests.py
@@ -27,6 +27,7 @@ TEST_SOURCES = (
     "crt/src/hal/uart_math.c",
     "crt/src/hal/timer_math.c",
     "crt/src/hal/psram_math.c",
+    "crt/src/hal/xpi_psram_math.c",
     "crt/src/hal/sdram_math.c",
     "crt/src/hal/sdio_math.c",
     "crt/src/hal/usb2_math.c",
diff --git a/scripts/tiny_r2_p8.py b/scripts/tiny_r2_p8.py
new file mode 100644
index 00000000..91697271
--- /dev/null
+++ b/scripts/tiny_r2_p8.py
@@ -0,0 +1,158 @@
+#!/usr/bin/env python3
+"""Retain P8 inputs and validate CPU/DMA pin-level workload evidence."""
+from __future__ import annotations
+
+import argparse
+from datetime import datetime, timezone
+import re
+import sys
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[1]
+sys.path.insert(0, str(ROOT))
+from scripts import tiny_r2_baseline as evidence  # noqa: E402
+from scripts.check_simulation import DEFAULT_FAILURE  # noqa: E402
+
+PHASE = "TINY-R2-P8"
+
+
+def measurements(content: str, config: dict, absent: bool = False) -> dict:
+    if DEFAULT_FAILURE.search(content) or "SIM_TEST_PASS Tiny cycles=" not in content:
+        raise ValueError("missing strict Tiny terminal success or forbidden error marker")
+    rows = [dict((k, int(v)) for k, v in re.findall(r"(\w+)=(\d+)", line))
+            for line in content.splitlines() if line.startswith("P8_TRANSPORT ")]
+    windows = [dict((k, int(v)) for k, v in re.findall(r"(\w+)=(\d+)", line))
+               for line in content.splitlines() if line.startswith("P8_XPI_WINDOW ")]
+    terminal = re.findall(r"P8_RESULT code=(\d+) profiles=(\d+)", content)
+    if len(terminal) != 1 or terminal[0] != ("0", str(len(rows))):
+        raise ValueError("missing or inconsistent firmware workload result")
+    if absent:
+        if rows or windows or "P8_ABSENT bounded optional-device rejection" not in content:
+            raise ValueError("absent-device run must reject initialization and perform no mapped DMA")
+        return {"mode": "absent", "profiles": [], "windows": []}
+    expected = [(24000000, 24000000)]
+    if config["PDK"] == "ICS55":
+        expected += [(192000000, 96000000), (240000000, 120000000)]
+    elif int(config["EXT_CLK_HZ"]) != 24000000:
+        expected += [(int(config["EXT_CLK_HZ"]), int(config["EXT_CLK_HZ"]))]
+    if [(r.get("sys_hz"), r.get("mem_hz")) for r in rows] != expected:
+        raise ValueError("missing, duplicated or unexpected clock profiles")
+    if len(windows) != 4 * len(rows):
+        raise ValueError("missing isolated/contended write and readback observation windows")
+    for index, row in enumerate(rows):
+        if row.get("bytes") != 1024 or row.get("burst") != 4 or row.get("progress", 0) <= 0:
+            raise ValueError("incorrect workload size/burst or no concurrent CPU progress")
+        for name in ("isolated", "contended"):
+            cycles = row.get(f"{name}_cycles", 0)
+            if cycles <= 0:
+                raise ValueError("invalid measured cycle count")
+            row[f"{name}_bytes_per_second"] = row["bytes"] * row["sys_hz"] / cycles
+        for offset, window in enumerate(windows[index * 4:index * 4 + 4]):
+            counts = (window.get("write_bytes"), window.get("read_bytes"))
+            if (counts != ((1024, 0) if offset % 2 == 0 else (0, 1024)) or
+                    window.get("index") != index * 4 + offset or window.get("mem_cycles", 0) <= 0):
+                raise ValueError("wrong mapped-path byte accounting")
+            for channel in ("aw", "w", "ar", "b", "r"):
+                stall = window.get(f"{channel}_max", -1)
+                if not 0 <= stall <= window["mem_cycles"]:
+                    raise ValueError("missing or impossible backpressure measurement")
+                window[f"{channel}_max_ns"] = stall * 1e9 / row["mem_hz"]
+    return {"mode": "present", "profiles": rows, "windows": windows,
+            "limits": "Digital model only; measured stalls apply only to this finite workload. "
+                      "No physical frequency, unrestricted contention or camera rate qualification."}
+
+
+def source_inputs() -> dict:
+    inputs = evidence.source_inputs(ROOT, "ICS55") | evidence.source_inputs(ROOT, "IHP130")
+    for folder in ("rtl/model", "tests", "configs/ci"):
+        for path in (ROOT / folder).rglob("*"):
+            if path.is_file() and path.suffix in (".sv", ".svh", ".py", ".c", ".h", ".mk"):
+                inputs[str(path.relative_to(ROOT))] = evidence.sha256(path)
+    return inputs
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("action", choices=("capture", "report"))
+    parser.add_argument("--output", type=Path, required=True)
+    parser.add_argument("--source", type=Path)
+    parser.add_argument("--variant-root", type=Path)
+    parser.add_argument("--simulator", choices=("verilator", "iverilog"))
+    parser.add_argument("--attempt", type=Path)
+    parser.add_argument("--absent", action="store_true")
+    args = parser.parse_args()
+    if args.variant_root:
+        args.variant_root = args.variant_root.resolve()
+    try:
+        if args.output.exists():
+            raise ValueError("evidence already exists; select a new attempt")
+        inputs = source_inputs()
+        if args.action == "capture":
+            if not all((args.variant_root, args.simulator)):
+                raise ValueError("capture requires a built --variant-root and --simulator")
+            runtime = evidence.runtime_inputs(ROOT, args.variant_root, args.simulator)
+            for path in (args.variant_root / "sw/firmware",
+                         args.variant_root / "sw/retrosoc_fw.hex",
+                         args.variant_root / "sim" / args.simulator /
+                         ("emu" if args.simulator == "verilator" else "simv")):
+                runtime[str(path.resolve().relative_to(ROOT))] = evidence.sha256(path)
+            evidence.dump(args.output, {
+                "phase": PHASE, "repository": evidence.git_info(ROOT),
+                "captured_at": datetime.now(timezone.utc).isoformat(),
+                "variant": str(args.variant_root.resolve()), "simulator": args.simulator,
+                "runtime_inputs": runtime,
+                "specification": evidence.artifact(ROOT / "docs/ip/tiny-soc.md"),
+                "source_inputs": inputs, "source_digest": evidence.digest_map(inputs),
+                "tools": evidence.tools_identity(ROOT),
+                "snapshot": evidence.snapshot(ROOT, args.output.parent / "source", inputs),
+            })
+        else:
+            if not all((args.source, args.variant_root, args.simulator, args.attempt)):
+                raise ValueError("report requires --source --variant-root --simulator --attempt")
+            source = evidence.read_json(args.source)
+            if source.get("phase") != PHASE or source.get("source_inputs") != inputs:
+                raise ValueError("source changed since the retained snapshot")
+            if (source.get("variant") != str(args.variant_root.resolve()) or
+                    source.get("simulator") != args.simulator):
+                raise ValueError("runtime snapshot belongs to another variant or simulator")
+            for name, digest in source["runtime_inputs"].items():
+                if evidence.sha256(ROOT / name) != digest:
+                    raise ValueError(f"runtime input changed after capture: {name}")
+            manifest = evidence.read_json(args.variant_root / "meta/manifest.json")
+            config = manifest["configuration"]
+            if (config.get("SOC") != "TINY" or config.get("TINY_P8_ACCEPTANCE") != "YES" or
+                    config.get("PDK") not in ("ICS55", "IHP130")):
+                raise ValueError("not an explicit Tiny P8 configuration")
+            result = evidence.read_json(args.attempt / "result-sim.json")
+            log = args.attempt / "sim.log"
+            if (result.get("status") != "passed" or result.get("exit_code") != 0 or
+                    Path(result.get("log", "")).resolve() != log.resolve()):
+                raise ValueError("simulation command did not complete successfully with this log")
+            if datetime.fromisoformat(result["started_at"]) < datetime.fromisoformat(source["captured_at"]):
+                raise ValueError("simulation predates the source/runtime snapshot")
+            command = result.get("command", [])
+            binary = args.variant_root / "sim" / args.simulator / (
+                "emu" if args.simulator == "verilator" else "simv")
+            firmware_arg = "+firmware=" + str((args.variant_root / "sw/retrosoc_fw.hex").resolve())
+            if (str(binary.resolve()) not in command or firmware_arg not in command or
+                    "+tiny_p8_psram" not in command or (("+p8_absent" in command) != args.absent)):
+                raise ValueError("simulation command does not match captured binary/firmware/mode")
+            report = measurements(log.read_text(), config, args.absent)
+            artifacts = [log, args.attempt / "result-sim.json", args.source,
+                         args.variant_root / "sw/firmware",
+                         args.variant_root / "sw/retrosoc_fw.hex",
+                         args.variant_root / "sim" / args.simulator /
+                         ("emu" if args.simulator == "verilator" else "simv")]
+            evidence.dump(args.output, {"phase": PHASE, "status": "passed",
+                "configuration": config, "simulator": args.simulator, "measurements": report,
+                "source_digest": source["source_digest"],
+                "artifacts": [evidence.artifact(p) for p in artifacts]})
+        print(f"{PHASE} {args.action}: PASS ({args.output})")
+        return 0
+    except (OSError, ValueError, KeyError) as error:
+        print(f"{PHASE}: {error}", file=sys.stderr)
+        return 1
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/tests/c/test_runtime.c b/tests/c/test_runtime.c
index 664778e6..c2c4d37b 100644
--- a/tests/c/test_runtime.c
+++ b/tests/c/test_runtime.c
@@ -25,6 +25,7 @@
 #include <retrosoc/hal/memory.h>
 #include <retrosoc/hal/npu.h>
 #include <retrosoc/hal/psram.h>
+#include <retrosoc/hal/xpi_psram.h>
 #include <retrosoc/hal/resource.h>
 #include <retrosoc/hal/sdram.h>
 #include <retrosoc/hal/sdio.h>
@@ -638,6 +639,36 @@ static int test_sdram_helpers(void) {
     return 0;
 }
 
+static int test_xpi_psram_helpers(void) {
+    rs_xpi_slot_config_t config;
+    rs_xpi_psram_t device = {.initialized = true, .capacity_bytes = RS_XPI_PSRAM_BYTES};
+    const uint32_t rates[] = {24000000U, 48000000U, 96000000U, 120000000U};
+    for (uint32_t i = 0U; i < 4U; i++) {
+        if (rs_xpi_psram_timing(rates[i], &config) != RS_OK || config.memory_write_enable ||
+            config.memory_read_enable || config.device_size != RS_XPI_PSRAM_BYTES ||
+            config.timeout_cycles >= rates[i] / 125000U || config.burst_boundary != 1024U ||
+            (uint64_t)config.cs_high_cycles * 1000000000U < (uint64_t)rates[i] * 50U) {
+            return 1;
+        }
+    }
+    if (rs_xpi_psram_timing(UINT32_MAX, &config) != RS_EINVAL ||
+        rs_xpi_psram_timing(24000000U, NULL) != RS_EINVAL ||
+        rs_xpi_psram_validate_dma(&device, RS_XPI_PSRAM_BASE, RS_XPI_PSRAM_BYTES, 4U) != RS_OK ||
+        rs_xpi_psram_validate_dma(&device, 0x54000c00U, 1024U, 4U) != RS_OK ||
+        rs_xpi_psram_validate_dma(&device, 0x547ffffcU, 4U, 1U) != RS_OK ||
+        rs_xpi_psram_validate_dma(&device, 0x547ffffcU, 8U, 1U) != RS_EINVAL ||
+        rs_xpi_psram_validate_dma(&device, 0x540003f4U, 64U, 4U) != RS_EINVAL ||
+        rs_xpi_psram_validate_dma(&device, 0x54000ff4U, 16U, 4U) != RS_EINVAL ||
+        rs_xpi_psram_validate_dma(&device, 0x540003fcU, 64U, 4U) != RS_EINVAL ||
+        rs_xpi_psram_validate_dma(&device, RS_XPI_PSRAM_BASE, UINT32_MAX, 4U) != RS_EINVAL ||
+        rs_xpi_psram_validate_dma(&device, RS_XPI_PSRAM_BASE, 64U, 16U) != RS_EINVAL ||
+        rs_xpi_psram_validate_dma(&device, RS_XPI_PSRAM_BASE, 0U, 4U) != RS_EINVAL) {
+        return 2;
+    }
+    device.initialized = false;
+    return rs_xpi_psram_validate_dma(&device, RS_XPI_PSRAM_BASE, 4U, 1U) == RS_ENOTSUP ? 0 : 3;
+}
+
 static int test_psram_helpers(void) {
     rs_psram_timing_t timing;
 
@@ -3193,13 +3224,11 @@ static int test_tiny_rcu_boundaries(void) {
         rs_tiny_rcu_get_snapshot(NULL) != RS_EINVAL ||
         rs_tiny_rcu_get_snapshot(&snapshot) != RS_OK || snapshot.xin_hz != hz ||
         rs_tiny_rcu_set_clkout(5U, 1U) != RS_EINVAL ||
-        rs_tiny_rcu_set_clkout(1U, 0U) != RS_EINVAL ||
-        rs_tiny_rcu_clear_fault(32U) != RS_EINVAL ||
+        rs_tiny_rcu_set_clkout(1U, 0U) != RS_EINVAL || rs_tiny_rcu_clear_fault(32U) != RS_EINVAL ||
         rs_tiny_rcu_clear_reset_cause(64U) != RS_EINVAL) {
         return 6;
     }
-    if (rs_tiny_rcu_set_clkout(1U, 1U) != RS_OK ||
-        rs_tiny_rcu_set_clkout(1U, 2U) != RS_EBUSY ||
+    if (rs_tiny_rcu_set_clkout(1U, 1U) != RS_OK || rs_tiny_rcu_set_clkout(1U, 2U) != RS_EBUSY ||
         rs_tiny_rcu_set_clkout(0U, 0U) != RS_OK) {
         return 7;
     }
@@ -3222,6 +3251,7 @@ int main(void) {
         test_ws2812_completion_race(),
         test_timer_helpers(),
         test_psram_helpers(),
+        test_xpi_psram_helpers(),
         test_sdram_helpers(),
         test_uart_helpers(),
         test_i2s_helpers(),
diff --git a/tests/rtl/tiny_p8_xpi_tb.sv b/tests/rtl/tiny_p8_xpi_tb.sv
new file mode 100644
index 00000000..d7a55e1f
--- /dev/null
+++ b/tests/rtl/tiny_p8_xpi_tb.sv
@@ -0,0 +1,303 @@
+`timescale 1ns / 1ps
+
+// Production APB/AXI wrapper and real serial device. No memory backdoor.
+module tiny_p8_xpi_tb;
+  import xpi_pkg::*;
+  logic s_clk = 0, s_rst_n = 0;
+  apb4_if apb (
+      s_clk,
+      s_rst_n
+  );
+  axi4_if #(
+      .ADDR_WIDTH(32),
+      .DATA_WIDTH(32),
+      .ID_WIDTH  (1),
+      .USER_WIDTH(1)
+  ) axi ();
+  xpi_if pins ();
+  tri1 [3:0] s_data;
+  logic s_idle, s_tx_stall, s_rx_stall, s_irq;
+  logic   [31:0] s_value;
+  integer        s_commands = 0;
+
+  always #20.833 s_clk = ~s_clk;
+  for (genvar i = 0; i < 4; i++) begin : gen_pads
+    assign s_data[i] = pins.io_oe_o[i] ? pins.io_do_o[i] : 1'bz;
+  end
+  assign pins.io_di_i = s_data;
+  always @(negedge pins.nss_o[1]) s_commands++;
+
+  apb4_xpi u_dut (
+      .clk_i           (s_clk),
+      .rst_n_i         (s_rst_n),
+      .dma_xfer_done_i (1'b0),
+      .lifecycle_idle_o(s_idle),
+      .dma_tx_stall_o  (s_tx_stall),
+      .dma_rx_stall_o  (s_rx_stall),
+      .irq_o           (s_irq),
+      .apb4            (apb),
+      .mem_axi4        (axi),
+      .xpi             (pins)
+  );
+  ESP_PSRAM64H #(
+      .POWER_UP_CHECK        (1),
+      .TIMING_VIOLATION_FATAL(1)
+  ) u_ram (
+      .sclk(pins.sck_o),
+      .csn (pins.nss_o[1]),
+      .sio (s_data)
+  );
+
+  task automatic access (input bit write, input logic [11:0] addr, input logic [31:0] data,
+                         output logic [31:0] result);
+    @(negedge s_clk);
+    apb.psel    = 1;
+    apb.penable = 0;
+    apb.pwrite  = write;
+    apb.paddr   = {20'd0, addr};
+    apb.pwdata  = data;
+    apb.pstrb   = 4'hf;
+    @(negedge s_clk);
+    apb.penable = 1;
+    do @(posedge s_clk); while (!apb.pready);
+    if (apb.pslverr) $fatal(1, "SIM_TEST_FAIL APB addr=%h", addr);
+    result = apb.prdata;
+    @(negedge s_clk);
+    apb.psel    = 0;
+    apb.penable = 0;
+  endtask
+  task automatic wr(input logic [11:0] addr, input logic [31:0] data);
+    logic [31:0] unused;
+    access (1, addr, data, unused);
+  endtask
+  task automatic rd(input logic [11:0] addr, output logic [31:0] data);
+    access (0, addr, 0, data);
+  endtask
+  task automatic idle;
+    logic [31:0] status;
+    do rd(12'h010, status); while ((status & 32'h0e) != 0);
+  endtask
+  task automatic program_lut(input int seq, input logic [15:0] a, b, c, d);
+    wr(12'(12'h200 + seq * 16), {b, a});
+    wr(12'(12'h204 + seq * 16), {d, c});
+    wr(12'(12'h208 + seq * 16), 0);
+    wr(12'(12'h20c + seq * 16), 0);
+  endtask
+  task automatic command(input logic [7:0] opcode, input logic [1:0] pads);
+    idle();
+    wr(12'h00c, 0);
+    program_lut(3, xpi_instr(XpiInstrCommand, pads, opcode), 0, 0, 0);
+    wr(12'h00c, 1);
+    wr(12'h024, 255);
+    wr(12'h048, 0);
+    wr(12'h04c, 0);
+    wr(12'h050, 32'h31);
+    wr(12'h014, 1);
+    idle();
+    rd(12'h024, s_value);
+    if ((s_value & 32'h61) != 1) $fatal(1, "SIM_TEST_FAIL command %h status=%h", opcode, s_value);
+  endtask
+
+  task automatic write_burst(input logic [31:0] addr, input int beats, input logic [2:0] size,
+                             input logic [1:0] expected, input logic [31:0] pattern,
+                             input logic [3:0] strobes);
+    @(negedge s_clk);
+    axi.awaddr  = addr;
+    axi.awlen   = 8'(beats - 1);
+    axi.awsize  = size;
+    axi.awvalid = 1;
+    do @(posedge s_clk); while (!axi.awready);
+    @(negedge s_clk);
+    axi.awvalid = 0;
+    for (int b = 0; b < beats; b++) begin
+      axi.wdata  = pattern + 32'(b);
+      axi.wstrb  = size < 2 ? 4'(((1 << (1 << size)) - 1) << ((addr + (b << size)) & 3)) : strobes;
+      axi.wlast  = (b == beats - 1);
+      axi.wvalid = 1;
+      do @(posedge s_clk); while (!axi.wready);
+      @(negedge s_clk);
+      axi.wvalid = 0;
+    end
+    do @(posedge s_clk); while (!axi.bvalid);
+    if (axi.bresp !== expected)
+      $fatal(1, "SIM_TEST_FAIL write addr=%h resp=%h expected=%h", addr, axi.bresp, expected);
+    @(negedge s_clk);
+  endtask
+  task automatic read_burst(input logic [31:0] addr, input int beats, input logic [1:0] expected,
+                            input logic [31:0] pattern);
+    @(negedge s_clk);
+    axi.araddr  = addr;
+    axi.arlen   = 8'(beats - 1);
+    axi.arvalid = 1;
+    do @(posedge s_clk); while (!axi.arready);
+    @(negedge s_clk);
+    axi.arvalid = 0;
+    for (int b = 0; b < beats; b++) begin
+      do @(posedge s_clk); while (!axi.rvalid);
+      if (axi.rresp !== expected || axi.rlast !== (b == beats - 1))
+        $fatal(1, "SIM_TEST_FAIL read response addr=%h beat=%0d resp=%h", addr, b, axi.rresp);
+      if (expected == 0 && axi.rdata !== pattern + 32'(b))
+        $fatal(
+            1,
+            "SIM_TEST_FAIL readback addr=%h beat=%0d got=%h expected=%h",
+            addr,
+            b,
+            axi.rdata,
+            pattern + 32'(b)
+        );
+      @(negedge s_clk);
+    end
+    idle();
+  endtask
+
+  task automatic held_read;
+    @(negedge s_clk);
+    axi.araddr  = 32'h54000000;
+    axi.arlen   = 1;
+    axi.arvalid = 1;
+    axi.rready  = 0;
+    do @(posedge s_clk); while (!axi.arready);
+    @(negedge s_clk);
+    axi.arvalid = 0;
+    do @(posedge s_clk); while (!axi.rvalid);
+    for (int i = 0; i < 250; i++) begin
+      @(negedge s_clk);
+      if (!axi.rvalid || axi.rresp !== 0 || axi.rdata !== 32'h12345678 || axi.rlast)
+        $fatal(1, "SIM_TEST_FAIL held read changed across core timeout");
+    end
+    axi.rready = 1;
+    @(posedge s_clk);
+    @(negedge s_clk);
+    do @(posedge s_clk); while (!axi.rvalid);
+    if (axi.rresp !== 2 || !axi.rlast) $fatal(1, "SIM_TEST_FAIL timeout tail");
+    @(negedge s_clk);
+    idle();
+  endtask
+
+  initial begin
+    #2000000;
+    $fatal(1, "SIM_TEST_TIMEOUT Tiny P8 XPI");
+  end
+  initial begin
+    apb.psel     = 0;
+    apb.penable  = 0;
+    apb.pwrite   = 0;
+    apb.paddr    = 0;
+    apb.pwdata   = 0;
+    apb.pstrb    = 0;
+    apb.pprot    = 0;
+    axi.awid     = 0;
+    axi.awaddr   = 0;
+    axi.awlen    = 0;
+    axi.awsize   = 2;
+    axi.awburst  = 1;
+    axi.awlock   = 0;
+    axi.awcache  = 0;
+    axi.awprot   = 0;
+    axi.awqos    = 0;
+    axi.awregion = 0;
+    axi.awuser   = 0;
+    axi.awvalid  = 0;
+    axi.wdata    = 0;
+    axi.wstrb    = 0;
+    axi.wlast    = 0;
+    axi.wuser    = 0;
+    axi.wvalid   = 0;
+    axi.bready   = 1;
+    axi.arid     = 0;
+    axi.araddr   = 0;
+    axi.arlen    = 0;
+    axi.arsize   = 2;
+    axi.arburst  = 1;
+    axi.arlock   = 0;
+    axi.arcache  = 0;
+    axi.arprot   = 0;
+    axi.arqos    = 0;
+    axi.arregion = 0;
+    axi.aruser   = 0;
+    axi.arvalid  = 0;
+    axi.rready   = 1;
+    repeat (5) @(negedge s_clk);
+    s_rst_n = 1;
+    #150000;
+    wr(12'h00c, 0);
+    wr(12'h120, 1);
+    wr(12'h124, 32'h800000);
+    wr(12'h128, 32'h21);
+    wr(12'h12c, 32'h02010100);
+    wr(12'h130, 192);
+    wr(12'h134, 1024);
+    program_lut(1, 16'h18eb, 16'h2818, 16'h4006, 16'h6800);
+    program_lut(2, 16'h1838, 16'h2818, 16'h5800, 0);
+    // Read ID: SPI 9f, 24 address bits, six bytes, no dummy clocks.
+    program_lut(4, 16'h109f, 16'h2018, 16'h6000, 0);
+    wr(12'h00c, 1);
+    command(8'h66, 0);
+    command(8'h99, 0);
+    #100;
+    wr(12'h024, 255);
+    wr(12'h048, 0);
+    wr(12'h04c, 6);
+    wr(12'h050, 32'h41);
+    wr(12'h014, 1);
+    idle();
+    rd(12'h044, s_value);
+    if (s_value[23:0] !== 24'h265d0d) $fatal(1, "SIM_TEST_FAIL PSRAM identity %h", s_value);
+    rd(12'h044, s_value);
+    command(8'h35, 0);
+    idle();
+    wr(12'h00c, 0);
+    wr(12'h120, 7);
+    wr(12'h00c, 1);
+
+    if ($test$plusargs("boundary")) begin
+      read_burst(32'h540003fc, 2, 2, 0);
+      read_burst(32'h54000ffc, 2, 2, 0);
+      read_burst(32'h54800000, 1, 2, 0);
+      write_burst(32'h540003fc, 2, 2, 2, 0, 4'hf);
+      write_burst(32'h54000ffc, 2, 2, 2, 0, 4'hf);
+      write_burst(32'h54800000, 1, 2, 2, 0, 4'hf);
+      write_burst(32'h00000000, 1, 2, 2, 0, 4'hf);
+    end else if ($test$plusargs("lanes")) begin
+      write_burst(32'h54000200, 1, 2, 0, 32'h00000000, 4'hf);
+      write_burst(32'h54000200, 4, 0, 0, 32'h44332211, 4'hf);
+      read_burst(32'h54000200, 1, 0, 32'h44332211);
+      write_burst(32'h54000200, 2, 1, 0, 32'h88776655, 4'hf);
+      read_burst(32'h54000200, 1, 0, 32'h88776655);
+      write_burst(32'h54000200, 1, 2, 0, 32'h01020304, 4'h5);
+      read_burst(32'h54000200, 1, 0, 32'h88026604);
+      write_burst(32'h54000200, 1, 2, 0, 32'hffffffff, 4'h0);
+      read_burst(32'h54000200, 1, 0, 32'h88026604);
+    end else if ($test$plusargs("timeout")) begin
+      // A mapped TX must retire with error even if the core times out
+      // before accepting its first payload byte.
+      idle();
+      wr(12'h00c, 0);
+      wr(12'h130, 1);
+      wr(12'h00c, 1);
+      write_burst(32'h54000000, 1, 2, 2, 32'h12345678, 4'h1);
+    end else if ($test$plusargs("short_read")) begin
+      idle();
+      wr(12'h00c, 0);
+      program_lut(1, 0, 0, 0, 0);
+      wr(12'h00c, 1);
+      read_burst(32'h54000000, 1, 2, 0);
+    end else if (!$test$plusargs("identity")) begin
+      write_burst(32'h54000000, 1, 2, 0, 32'h12345678, 4'hf);
+      if ($test$plusargs("held_read")) held_read();
+      else if ($test$plusargs("late_error")) begin
+        idle();
+        wr(12'h00c, 0);
+        wr(12'h218, 32'h0000f000);
+        wr(12'h00c, 1);
+        read_burst(32'h54000000, 1, 2, 0);
+      end else begin
+        read_burst(32'h54000000, 1, 0, 32'h12345678);
+        write_burst(32'h547ffffc, 1, 2, 0, 32'h89abcdef, 4'hf);
+        read_burst(32'h547ffffc, 1, 0, 32'h89abcdef);
+      end
+    end
+    $display("SIM_TEST_PASS Tiny P8 XPI pins");
+    $finish;
+  end
+endmodule
diff --git a/tests/test_psram.py b/tests/test_psram.py
index 4f2d8aad..637eb1a7 100644
--- a/tests/test_psram.py
+++ b/tests/test_psram.py
@@ -59,7 +59,7 @@ def test_psram_controller_data_integrity_and_fault_isolation(tmp_path: Path) ->
             str(ROOT / "rtl/ip/memory/psram_phy.sv"),
             str(ROOT / "rtl/ip/memory/psram_reg.sv"),
             str(ROOT / "rtl/ip/memory/apb4_psram.sv"),
-            str(ROOT / "rtl/mini/dv/model/ESP_PSRAM64H.sv"),
+            str(ROOT / "rtl/model/ESP_PSRAM64H.sv"),
             str(ROOT / "tests/rtl/psram_tb.sv"),
             "-Mdir",
             str(tmp_path / "obj"),
diff --git a/tests/test_tiny.py b/tests/test_tiny.py
index 330520f4..c39b742d 100644
--- a/tests/test_tiny.py
+++ b/tests/test_tiny.py
@@ -49,6 +49,9 @@ def test_tiny_sources_and_abi_are_independent(tmp_path: Path) -> None:
     for source in resolved.files:
         assert "/rtl/mini/" not in str(source)
         assert not any(token in source.name.lower() for token in ("rib", "vexii", "radio"))
+    devices = parse_filelists([filelists / "tb.fl"])
+    assert ROOT / "rtl/model/ESP_PSRAM64H.sv" in devices.files
+    assert all("/rtl/mini/" not in str(source) for source in devices.files)
     rtl = (output / "memory_map/rtl/mmap_define.svh").read_text()
     sdk = (output / "memory_map/include/retrosoc/generated/memory_map.h").read_text()
     assert "RIB" not in rtl
diff --git a/tests/test_tiny_r2_p8.py b/tests/test_tiny_r2_p8.py
new file mode 100644
index 00000000..bb4e3001
--- /dev/null
+++ b/tests/test_tiny_r2_p8.py
@@ -0,0 +1,34 @@
+"""Pin-level XPI PSRAM transport; no fast NOR or separate RAM controller."""
+import subprocess
+
+import pytest
+
+from test_tiny_r2_p7 import COMMON, ROOT, run_rtl
+from scripts.check_simulation import DEFAULT_FAILURE
+
+
+@pytest.fixture(scope="module", params=["iverilog", "verilator"])
+def xpi_simulator(request, tmp_path_factory):
+    simulator = request.param
+    tmp_path = tmp_path_factory.mktemp(f"p8-{simulator}")
+    run_rtl(tmp_path, simulator, "tiny_p8_xpi_tb", [
+        COMMON / "utils/fifo.sv", COMMON / "interface/axi4_addr_gen.sv",
+        *[ROOT / "rtl/ip/storage" / name for name in (
+            "xpi_if.sv", "xpi_reg.sv", "xpi_mm.sv", "xpi_indirect.sv",
+            "xpi_clkgen.sv", "xpi_core.sv", "apb4_xpi.sv")],
+        ROOT / "rtl/model/ESP_PSRAM64H.sv",
+    ], defines=[f"-I{ROOT / 'rtl/ip/storage'}", "-DRETROSOC_SOC__TINY"],
+        packages=[ROOT / "rtl/ip/storage/xpi_pkg.sv"], plusargs=["+identity"])
+    launch = [str(tmp_path / "sim")] if simulator == "verilator" else ["vvp", str(tmp_path / "sim")]
+    return launch, tmp_path
+
+
+@pytest.mark.parametrize("scenario", ["mapped", "boundary", "lanes", "timeout", "held_read",
+                                      "short_read", "late_error"])
+def test_xpi_psram_pins(xpi_simulator, scenario: str) -> None:
+    launch, directory = xpi_simulator
+    result = subprocess.run([*launch, f"+{scenario}"], capture_output=True, text=True, timeout=30)
+    log = result.stdout + result.stderr
+    (directory / f"{scenario}.log").write_text(log)
+    assert result.returncode == 0 and not DEFAULT_FAILURE.search(log), log
+    assert "SIM_TEST_PASS Tiny P8 XPI pins" in log, log
diff --git a/tests/test_tiny_r2_p8_evidence.py b/tests/test_tiny_r2_p8_evidence.py
new file mode 100644
index 00000000..907f63f2
--- /dev/null
+++ b/tests/test_tiny_r2_p8_evidence.py
@@ -0,0 +1,63 @@
+"""Reject incomplete P8 transport/optional-device evidence."""
+import pytest
+from pathlib import Path
+
+from scripts.tiny_r2_p8 import measurements
+
+CONFIG = {"PDK": "IHP130", "EXT_CLK_HZ": "24000000"}
+
+
+@pytest.fixture
+def valid_log():
+    return "\n".join([
+        "P8_TRANSPORT sys_hz=24000000 mem_hz=24000000 bytes=1024 burst=4 "
+        "isolated_cycles=24000 contended_cycles=48000 progress=10",
+        *[f"P8_XPI_WINDOW index={i} mem_cycles=24000 "
+          f"write_bytes={1024 if i % 2 == 0 else 0} read_bytes={0 if i % 2 == 0 else 1024} "
+          "aw_max=10 w_max=20 ar_max=30 b_max=40 r_max=50" for i in range(4)],
+        "P8_RESULT code=0 profiles=1", "SIM_TEST_PASS Tiny cycles=100000",
+    ])
+
+
+def test_p8_measurement_units(valid_log):
+    result = measurements(valid_log, CONFIG)
+    assert result["profiles"][0]["isolated_bytes_per_second"] == 1024000
+    assert result["profiles"][0]["contended_bytes_per_second"] == 512000
+    assert result["windows"][0]["b_max_ns"] == pytest.approx(40e9 / 24e6)
+
+
+@pytest.mark.parametrize("old,new", [
+    ("SIM_TEST_PASS Tiny cycles=100000", "Hello retroSoC!"),
+    ("profiles=1", "profiles=0"), ("code=0", "code=3"),
+    ("write_bytes=1024", "write_bytes=1020"), ("progress=10", "progress=0"),
+    ("aw_max=10", "aw_max=25000"), ("isolated_cycles=24000", "isolated_cycles=0"),
+])
+def test_p8_rejects_incomplete_measurements(valid_log, old, new):
+    with pytest.raises(ValueError):
+        measurements(valid_log.replace(old, new), CONFIG)
+
+
+def test_p8_does_not_promote_missing_profiles_or_errors(valid_log):
+    with pytest.raises(ValueError):
+        measurements(valid_log, {"PDK": "ICS55", "EXT_CLK_HZ": "24000000"})
+    with pytest.raises(ValueError):
+        measurements(valid_log + "\nSIM_TEST_FAIL", CONFIG)
+    with pytest.raises(ValueError):
+        measurements(valid_log, CONFIG, absent=True)
+
+
+def test_p8_optional_absence_has_separate_verdict():
+    log = ("P8_ABSENT bounded optional-device rejection\nP8_RESULT code=0 profiles=0\n"
+           "SIM_TEST_PASS Tiny cycles=100000\n")
+    assert measurements(log, CONFIG, absent=True)["mode"] == "absent"
+    with pytest.raises(ValueError):
+        measurements(log, CONFIG)
+
+
+def test_p8_observer_labels_match_axi_channels():
+    source = (Path(__file__).resolve().parents[1] /
+              "rtl/tiny/dv/tiny_p8_observers.svh").read_text()
+    order = source.split("wire [4:0] s_p8_stalls = {", 1)[1].split("};", 1)[0]
+    channels = [line.split("`_P8_AXI.", 1)[1].split("valid", 1)[0]
+                for line in order.splitlines() if "`_P8_AXI." in line]
+    assert channels == ["aw", "w", "ar", "b", "r"]