[Vop3p]

Results only (no engine change). Nine runs of strata bench run on one Windows 11 PC with one RX 6900 XT, all complete, every scheduled request kept. The 0.2.3 notes call the Windows rdna2 package "built but untried"; these are runs of it on a real card, plus two runs of main with #4005. The same card under Linux is in #3790.

PC: Windows 11 (10.0.26300), i9-10900K, 64 GB DDR4-3600, RX 6900 XT 16 GB on PCIe 3.0 x16 (driver 32.0.21045.16010), model files on an NVMe SSD, 64 GB fixed page file. Qwen3.8-Flash-Next GSQ-RCO IQ3_S, 64K context, 8-bit KV, --spec 4 --spec-min-p 0.5, one slot. Engine: the official windows-x86_64-hip-rdna2 package 0.1.48 unchanged, fetched by strata launch from the update feed, except the last two rows.

run	track	what differs from the row above	prompt 4K	prompt 16K	decode short / 4K / 16K	first token (short)	sanity
…-687e512e	default	setup defaults (reserve 2,560 MiB; a 5120x2160 monitor on the card)	451.8	479.3	28.8 / 28.3 / 28.4	2.81 s	12/12
…-7dc83265	custom	the monitor moved to the iGPU, dwm.exe restarted	445.2	470.0	29.2 / 29.1 / 30.1	2.64 s	12/12
…-8f3712c4	custom	defaults with --vram-reserve-mib 700, the monitor on the card	460.4	487.2	17.2 / 18.7 / 24.8	2.52 s	0/12
…-84b6d25c	custom	--vram-reserve-mib 1536 --kv-resident 32768, rebooted with the monitor on the iGPU	467.1	488.9	36.6 / 33.3 / 36.3	2.44 s	12/12
…-1daf1da0	custom	+ the opt-in prompt switches without STRATA_SH_STREAM (list below)	591.1	797.9	35.2 / 35.4 / 37.3	1.14 s	12/12
…-6886b54a	custom	+ a rocBLAS table tuned on this PC with tools/hip/tune_rocblas	594.6	799.0	38.0 / 35.5 / 35.9	0.97 s	12/12
…-0e04f3ea	custom	+ --pcie-frac 0.2 --prefill 8192	585.6	1,075.5	44.5 / 42.3 / 41.3	1.11 s	12/12
…-c97e2108	custom	main 7dd5cf1 + #4005 built on this PC (build_windows.bat), its switches on, --prefill auto (6144)	793.1	978.9	48.3 / 42.6 / 44.1	1.06 s	12/12
…-9b72e2b1	custom	the same build and settings with --prefill 8192 (added 2026-10-25)	808.5	1,125.1	43.9 / 45.5 / 45.3	1.11 s	12/12

Prompt and decode in tok/s, medians of three, the engine's own timings. Switches: STRATA_HIP_PROMPT_F16=1 STRATA_HIP_ADAPT_KERNEL_COPY=1 STRATA_PREFILL_CPU_SHARE=auto STRATA_HC_PERSIST=1 STRATA_PIPED_CONCURRENT=1 STRATA_PREFILL_STREAM_MIN=2048 STRATA_IO_THREADS=128. Each run's config.json has every setting.

What the rows show

• The package runs on gfx1030 under Windows. strata launch picked 0.1.48 windows-x86_64-hip-rdna2 by itself, the test start passed, and every run below finished its 21 required requests.
• Pictures (not part of the protocol, so not in any run directory; checked by hand with the settings of …-0e04f3ea, one cold start each):
  • The package's CPU reader (--vision cpu) answers a test image (an invoice number, an amount, two shapes) and a slide (six facts, bar colours) correctly; big pictures are cut to 300 image tokens.
  • The package's GPU reader has code for gfx1200/gfx1201 only, so --vision gpu quietly falls back to the CPU reader on this card ([Engine] Windows rdna2 package 0.1.48: the GPU picture reader (strata-vision-hip.exe) is built for gfx1200/gfx1201 only, so --vision gpu falls back to the CPU on RX 6000 #4017).
  • Built from the package's own commit (6f8800aefe27, tools/vision unchanged) for gfx1030/gfx1031, the GPU reader runs next to the official engine: 4/4 answers correct, the slide read at 1,059 image tokens in 7.6 s (2.2 s asked again), the small image in 3.5 s (1.1 s). It takes about 1.3 GB of VRAM, which the engine gives up from the expert cache by itself: 3,479 slots instead of 4,166; Strata's benchmark.py 4K prompt 582 / decode 44.4 tok/s against 589 / 44.5 with pictures off.
• VRAM reserve. What is left after the load is the reserve minus about 340 MiB at every value tried (2,560 / 1,536 / 1,280 / 1,024 / 700). Every run that ended below about 1 GiB paged; with the monitor on the card, 700 left nothing and every sanity answer read the prompt as empty (…-8f3712c4, 0/12, kept here as measured). 1,536 is the lowest stable value found.
• Decode: the PCIe share, not the cache. The start-up rule gives this link (12.6 GB/s probed) a share of 0.35; a sweep of --pcie-frac 0 / 0.1 / 0.2 / 0.3 / 0.35 (benchmark.py, two runs each) peaked at 0.2, decode +29%, and Strata's own refinement (STRATA_PCIE_FRAC_DEFAULT=auto) arrives at 0.20 too. Details and a Linux check in [Engine] HIP gfx103x: the start-up pcie_frac (bandwidth only) is too high when the CPU pool is fast; the PcieAuto refinement (opt-in there) finds 0.20 — enable it by default on gfx103x? #4007.
• Long prompts: the chunk. --prefill auto stops at 6,144 on Windows HIP because of Windows HIP (RX 7900 XTX): a long prompt with --prefill 8192 / auto makes the commit charge jump ~70 GB in one second, the engine dies at the commit limit (0xC0000409); 7500 and below stay flat #1630 (7900 XTX). On this card 8,192 ran clean in six cold starts, including three at --max-context 262144 with 27K and 100K prompts (Windows HIP (RX 7900 XTX): a long prompt with --prefill 8192 / auto makes the commit charge jump ~70 GB in one second, the engine dies at the commit limit (0xC0000409); 7500 and below stay flat #1630's setting), and reads 16K in two chunks instead of three (…-0e04f3ea: 798 → 1,076).
• Short prompts: copy events. Under WDDM every event record and cross-stream wait is a submission of its own; [Engine] prefill: opt-in sparse copied events and deduplicated slot waits (Windows HIP prompt reads +36% at 4K, +15% at 32K) #4005 removes 15 of 16 of them from the prompt path. …-c97e2108 reads 4K at 793 tok/s with it; the same-build comparison (10 interleaved pairs, switches off vs on) is in [Engine] prefill: opt-in sparse copied events and deduplicated slot waits (Windows HIP prompt reads +36% at 4K, +15% at 32K) #4005: 4K +36%, 32K +15%. …-c97e2108 ran at the auto chunk (6,144), so its 16K is below …-0e04f3ea; …-9b72e2b1 is the same build at 8,192: 4K 808 / 16K 1,125 against the package's 586 / 1,076 at the same chunk (+38% / +5%; the copy-event gain shrinks at 8,192 because fewer chunks mean fewer waits).
• STRATA_SH_STREAM is left out of every run: on Windows it lowered decode and, with KV streaming, stopped the engine ([Docs] AMD_HIP: why STRATA_SH_STREAM goes both ways, and leave it off on Windows #3895).
• A rocBLAS table made under Linux is refused on Windows (status 11 for every row, same rocBLAS version string); the table in …-6886b54a was tuned on this PC and helps only prompts under ~1K new tokens.

Failed, skipped or not measured

• …-687e512e carries ENGINE-NOT-STOPPED.txt: run from a Scheduled Task, the bench's final taskkill was denied and the engine kept the card until stopped by hand; the measurements are complete. Later runs were started from an interactive session.
• …-8f3712c4 failed its sanity checks (0/12) because of VRAM paging, as above.
• …-c97e2108 and …-9b72e2b1 are a local build of main + [Engine] prefill: opt-in sparse copied events and deduplicated slot waits (Windows HIP prompt reads +36% at 4K, +15% at 32K) #4005, not a release; its engine reports 0.1.49.2. The two differ only in --prefill (config.json).
• Not measured: background work, other RX 6000 cards (6800 / 6700 XT), a fresh PC from zero, the pictures under the protocol (only by hand, above), the strata desktop app path.

results.json sha256:

e190c4ee3d72bd9a09df94274b982394df57ed81c68cf8d21ca9b73dd02b1af1  2026-10-23-standard-1x-rx-6900xt-windows-687e512e
faa37b892080e989fd275471fe5c4267aa727b24f830fa43b40bc912fc0bedc5  2026-10-23-standard-1x-rx-6900xt-windows-7dc83265
d099f5fbc39af4838f0352b8200b232958e8f41ead0ead12832d03433fe95573  2026-10-23-standard-1x-rx-6900xt-windows-84b6d25c
4ad2eba8f325d700bfaa3f852fdc96a6119c3ef1d7322249284594054d335e12  2026-10-23-standard-1x-rx-6900xt-windows-8f3712c4
68fec9079f106a777f5940bf41ca3169455ae5c27a4cc9a75a6e7cd5ba4cae0c  2026-10-24-standard-1x-rx-6900xt-windows-0e04f3ea
288c9520cf7792e857707cbb46a49d4387fe81c3e2e33924d77ebc56b7293c3f  2026-10-24-standard-1x-rx-6900xt-windows-1daf1da0
dc1d8d026eed20802e6e6bcda33cfe6e2bdeb8c0ad0490ab0074c7e2b3b5e839  2026-10-24-standard-1x-rx-6900xt-windows-6886b54a
cb9c268fd78e66a16b63f0c00765815e3477598c9cbda596d8d06a245f10cc8e  2026-10-24-standard-1x-rx-6900xt-windows-c97e2108
965b401ea30b183f203f6b15d263e09f1d9217e2da07f48d3921a4803b18981e  2026-10-25-standard-1x-rx-6900xt-windows-9b72e2b1

Engine logs are not included (they carry local paths); available on request.

Developed with an AI coding assistant; every number above was measured on the PC described.

🤖 Generated with Claude Code

[Niko1221]

Thank you for measuring this and writing it up. We are building a Strata website with a page for community benchmarks, and results like these will be added there when it launches, so we are closing this pull request here; the numbers are not lost. If your numbers change with a new version, you are welcome to send them again.