retroSoC/docs/ip/onchip-sram.md

Configurable Native-AXI4 On-chip SRAM

Scope and maturity

The Mini SoC on-chip SRAM is a synthesis-time configurable AXI4 target backed by fixed 32-bit single-port technology wrappers. The reusable module keeps a 32-bit/ID1 compatibility mode; the Mini product instantiates native 64-bit/ID7 and stripes each beat over two wrappers. Supported capacities are 4, 16, 32, 64, and 128 KiB in compatibility mode; the product selects 32 KiB. The data aperture starts at 0x3000_0000; its generated end address and linker length match the selected physical capacity. A management-only APB4 window at 0x1001_7000 reports the configuration and saturating performance counters.

This release is a protocol and performance MVP. It is not a safety-certified, silicon-proven, or tapeout-ready memory subsystem. ECC, autonomous scrubbing, MBIST, redundancy repair, power gating, retention, and independently licensed AXI verification IP signoff remain outside the delivered scope.

Commercial reference review

The reference review was refreshed on 2026-08-24 from vendor-owned material. These products guide architecture and delivery expectations; they are not RTL dependencies and do not transfer a vendor's qualification to retroSoC.

Reference	Problem and architecture	Status and decision
Raspberry Pi RP2350	Main SRAM is split into two four-way word-striped groups. Two independent 4 KiB banks provide low-contention storage for stacks or hot data.	The official data sheet remains maintained. Reuse explicit bank ownership. Do not add low-bit striping while the Mini interconnect grants one transaction per target; it would increase switching without increasing target throughput.
STM32H7 embedded SRAM and RAMECC	AXI SRAM, AHB SRAM, TCM, backup SRAM, and power domains separate bandwidth, deterministic latency, and retention. SECDED reports failing address/data, and partial writes require read-modify-write.	STM32H7 products remain active. Reuse capability discovery, error localization, and a separate future ECC layer. Avoid claiming unchanged latency before ECC/RMW timing is verified.
NXP i.MX RT FlexRAM and RT1170 ECC guidance	Fixed banks can be assigned to OCRAM, ITCM, or DTCM. This depends on boot configuration, legal TCM sizes, linker agreement, and complete ECC initialization.	NXP continues to update RT1170 material. Reuse bank-based product configuration and generated linker consistency. Avoid runtime repartitioning and fuse/boot-order coupling for Mini.
Infineon AURIX TC3xx MTU/SSH	SRAM support hardware surrounds memories with ECC/EDC, address monitoring, error tracking, initialization, and MBIST. A central MTU exposes a uniform control plane.	AURIX documentation and products remain active. Reuse separation between functional access and reliability support. Avoid destructive runtime MBIST without a quiesce contract.
CAST SRAM-CTRL and ECC-SRAM	Commercial IP emphasizes native AXI/AHB/APB, byte accesses, low latency, lint/scan readiness, SECDED, testbenches, scripts, and integration documentation.	The products remain offered. Use the features and deliverables as an IP-readiness checklist, not as evidence for this implementation.
Synopsys STAR Memory System	Hierarchical test, diagnosis, redundancy repair, tester patterns, and physical failed-bit mapping address manufacturing yield and field diagnosis.	The product family remains published. Keep DFT/repair as a physical-delivery phase instead of embedding an unqualified March engine in the AXI datapath.

The selected architecture is static capacity configuration plus a native AXI64 datapath. The data crossbar queues up to four SRAM reads and two writes while this SRAM engine completes one transaction at a time; IDs are preserved to the terminal response.

Configuration and address contract

SRAM_SIZE_KIB is part of the reproducible configuration digest and accepts only 4, 16, 32, 64, or 128. IHP130, GF180, and SKY130 CI profiles select 32; ICS55 keeps the SRAM absent. IHP130 benchmark/CoreMark profiles retain 128. The address-map generator uses the selected value for RTL decode macros, SDK base/size/end constants, and the linker LENGTH(SRAM).

HAVE_SRAM_IF=YES marks the memory present and elaborates the bank array. When it is NO, the AXI target returns DECERR; the APB window remains readable with PRESENT=0, zero memory bytes, and zero banks.

SRAM_SIZE_KIB	4 KiB data banks	Address range
4	1	0x3000_0000..0x3000_0fff
16	4	0x3000_0000..0x3000_3fff
32	8	0x3000_0000..0x3000_7fff
64	16	0x3000_0000..0x3000_ffff
128	32	0x3000_0000..0x3001_ffff

AXI4 datapath

The product target uses 32-bit addresses, 64-bit data, seven-bit ID, and one-bit USER. It accepts one active read or write transaction, gives AR priority when AR and AW are simultaneously valid, and preserves the ID to the terminal response. The compatibility parameterization retains 32-bit data and one-bit ID for standalone/formal reuse.

Supported product requests are naturally aligned 1-, 2-, 4-, or 8-byte transfers using FIXED, INCR, or legal 2/4/8/16-beat WRAP bursts. Bursts are limited to 16 beats and one 4 KiB page. Because a bank is exactly one 4 KiB page, a legal transaction selects one striped bank pair for its lifetime.

The macro is issued directly on AR acceptance. The first response is available one cycle later. With RREADY asserted, following beats issue on preceding response cycles, giving one 64-bit beat per cycle without bubbles. Backpressure stops new reads and holds data, ID, response, and RLAST stable.

After AW acceptance the target consumes one W beat per cycle. Strobes outside the lanes selected by AWSIZE and address are illegal. A zero strobe is a no-op and does not enable a bank. WLAST must agree with AWLEN; a mismatching beat is not written and terminates with SLVERR. Earlier accepted beats are not rolled back.

Malformed protocol requests return SLVERR. An absent memory or address outside the parameterized capacity returns DECERR. Protocol-invalid requests are capped to one error response beat at the IP boundary. The SoC interconnect normally rejects them before they reach SRAM.

APB4 management ABI

All registers are read-only. Writes, unaligned accesses, and unknown offsets complete with PSLVERR. Counters are 32-bit saturating values controlled by the existing SYSCTRL PERF_CTRL enable and clear signals; no second control ABI or interrupt is introduced.

Offset	Name	Description
0x000	IP_ID	0x5352414d (SRAM)
0x004	IP_VERSION	ABI version 0x00010000
0x008	CAPABILITY	Present, native AXI4, byte-write, burst, and performance flags; max beats in 15:8 and data bytes in 23:16
0x00c	MEMORY_BYTES	Implemented capacity, or zero when absent
0x010	BANK_COUNT	Implemented data banks, or zero when absent
0x014	BANK_BYTES	Fixed value 4096
0x020	PERF_READ_REQUESTS	Accepted AR requests
0x024	PERF_WRITE_REQUESTS	Accepted AW requests
0x028	PERF_READ_BEATS	Accepted R beats
0x02c	PERF_WRITE_BEATS	Accepted W beats, including drained error traffic
0x030	PERF_STALL_CYCLES	Cycles with any AXI VALID blocked by READY
0x034	PERF_ERROR_RESPONSES	Terminal non-OKAY responses

The SDK exposes rs_onchip_sram_probe() and rs_onchip_sram_read_performance(). RTL and C constants remain handwritten; the register parity test prevents drift without adding a generator.

Technology mapping

tc_sram_1024x32 and tc_sram_4096x32 remain the stable technology boundaries. Behavioral simulation uses the Common byte-mask RAM model. IHP130 maps to RM_IHPSG13_1P_1024x32_c2_bm_bist. GF180 composes four byte lanes across two gf180mcu_fd_ip_sram__sram512x8m8wm1 depth halves and registers the address high bit for synchronous read selection. SKY130 maps to the OpenRAM-generated sky130_sram_4kbyte_1rw_32x1024_8. Its Liberty-compatible logical boundary exports 32 data bits and hides OpenRAM's required physical spare column; a future physical adapter must tie the spare enable and spare input low. The required spare row expands the macro address port to 11 bits, so the wrapper ties its high bit low and exposes exactly logical addresses 0 through 1023. ICS55 pairs two 4096x32 macros for the 32 KiB AXI64 product target, but its committed profiles keep the macro disabled. Bank count, rather than technology depth or width, implements every product size.

The wrapper preserves byte masks and one-cycle synchronous reads. PDK timing, LEF/GDS placement, power hookup, BIST pins, and foundry signoff remain physical-flow responsibilities.

Verification and performance evidence

tests/test_onchip_ram.py elaborates all five 32-bit compatibility capacities and covers bank boundaries, first/last words, byte/half/word masks, FIXED, INCR, WRAP, 16-beat continuous traffic, backpressure, AR/AW priority, absent-memory and malformed accesses, counters, all four PDK wrapper polarities and masks, and the GF180 511/512 depth boundary. The setup tests independently validate the SKY130 artifact manifest and generated-file hashes. Full Verilator regression firmware exercises the first and last locations through the SoC interconnect. The native AXI64 test additionally covers seven-bit IDs, continuous 64-bit bursts, and upper/lower 32-bit byte strobes.

make CONFIG=configs/ci/ihp130.mk formal-onchip-ram runs a 20-cycle bounded proof and cover set for AXI stability, channel exclusion, response bounds, physical-write qualification, and reachable read/write/error completions. The directed performance comparison measures 16 response cycles for a 16-word native read and 18 through the compatibility bridge. This is an 11.1 percent total-latency reduction and a complete removal of the bridge's two non-payload cycles; 16 cycles is the minimum possible on the 32-bit response channel. The SRAM-resident four-iteration CoreMark run completes in 4,662,868 cycles against the recorded 7,620,324-cycle compatibility baseline, a 38.8 percent cycle reduction. It remains an unqualified regression workload rather than a public CoreMark score.

Reliability and performance roadmap

The next phase should add SECDED outside the stable data macro. For every four data banks, one 1024 x 32 bank can store four independent ECC bytes per word address. A full write computes one ECC byte and writes its byte lane; a partial write reads and corrects the old word, merges bytes, then writes data and ECC. The management ABI would add first-error address/syndrome, corrected and uncorrectable counters, injection, sticky status, and an uncorrectable-error interrupt. All locations must be initialized with valid data/ECC before reads.

MBIST follows ECC. A future support wrapper must quiesce AXI, exclude CPUs, DMA, and debuggers, run a reviewed startup March test or qualified nondestructive runtime test, record the failing bank/address, and hand technology BIST/redundancy pins to the physical flow. That phase requires DFT and tester-ready evidence, not only RTL simulation.

Further throughput work is conditional on the fabric. Independent read/write target ownership, more IDs, or per-bank queues are useful only after concurrent transactions can reach SRAM. RP2350-style striping, bank-aware arbitration, QoS, and power gating should then be evaluated from measured contention.