#5666 [Benchmark] RX 7900 XTX (gfx1100) / Strata 0.2.13: the Linux Docker image without the gfx1100 hipBLASLt table reads 16K prompts 35 % slower; with the table equal to 0.2.12
closed · @faf-net · 3 comments · View on GitHub
BenchmarksAMD / HIPModels & quantsDocumentationWindowsLinux
Description
### What did you measure, or what went wrong **TL;DR: without the gfx1100 hipBLASLt table, the 0.2.13 Docker image reads prompts much slower on an RX 7900 XTX: −14 % at 4K and −35 % at 16K (769 instead of 1,186 tok/s), and the first token of a 16K prompt comes after 21.3 s instead of 13.8 s.** The table ships in the image but is now off by default on Linux too, not only on Windows as the release notes say. `-e STRATA_HIPBLASLT_TABLE=on` brings the full speed back. With the table set explicitly, 0.2.13 is equal to 0.2.12. 0.2.13 on the same machine as #5321 and #5615: RX 7900 XTX 24 GB (gfx1100) in an OCuLink PCIe 4.0 x4 eGPU dock behind a UGREEN iDX6011 Pro NAS (Core Ultra 7 255H, 64 GB LPDDR5X), Linux 6.18.15, Flash-Next IQ3_XXS native pack. Engine 0.1.59 `hip-rx7900` with control 0.1.59, SHA-256 checked against `SHA256SUMS.txt`. NAS services running, my agent paused, everything in one afternoon session. **1. The Linux Docker image runs without the gfx1100 table by default.** The 0.2.13 engine in `ghcr.io/niko1221/strata:0.2.13-rocm-gfx110x` logs ``` hipBLASLt tuning table: gfx1100-hipblaslt-100200.txt is in this package and not used by default (faster prompts, but answers can differ in the last digits); STRATA_HIPBLASLT_TABLE=on uses it ``` The release notes describe this change for Windows RX 9070 / R9700 and RX 7900. In the Linux image it applies too, and on this card prompt reading drops a lot: | `standard-v1`, unchanged image, `--track default` | short decode | 4K prompt | 16K prompt | | --- | ---: | ---: | ---: | | 0.2.12 image, defaults (8 Nov) | 69.1 | 756.3 | 1,185.8 | | **0.2.13 image, defaults** | 61.9 | **649.2** (−14 %) | **769.2** (−35 %) | | 0.2.13 image, only `-e STRATA_HIPBLASLT_TABLE=on` added | 67.4 | 765.8 | 1,192.5 | All three runs complete, 21 of 21, sanity 12/12. The 16K first token takes 21.3 s by default against 13.8 s with the table. With the switch on it is back at the 0.2.12 level. If keeping the table off on Linux gfx1100 is intended (exact-digit reproducibility), a line in the release notes or the Docker README for RX 7900 users would help; if not, the Linux default may want the gfx1151-style exception for gfx1100 too. **2. 0.2.13 against 0.2.12 with my own run config: equal.** Identical run config (`--max-context 262144 --kv int8 --kv-resident 32768 --spec 4 --vram-reserve-mib 832`, conversation cache 4 slots / 9,216 MiB, `STRATA_HIPBLASLT_TUNING` set to the gfx1100 table explicitly), only engine and control swapped: | | 0.2.13 | 0.2.12 | | --- | ---: | ---: | | 3 interleaved pairs, fresh engine each: decode tok/s (63-token prompt, 900 out) | 60.0 (58.1-61.4) | 60.7 (57.4-61.9) | | my own task set, 114 tasks (102 auto-graded), temperature 0, thinking on: passed | 97 | 98 * | | same task set: wall time / decode median | 1,381 s / 75.8 | 1,430 s / 75.6 * | | `standard-v1`, package, track custom: short decode / 4K / 16K prompt | 67.7 / 798.6 / 1,203.3 | 63.6 / 788.4 / 1,201.7 * | | `tools/strata_bench.py --long`: chat / agent decode, 4K / 16K / 32K prompt | 64.1 / 73.7, 782 / 1,203 / 1,217 | 68.1 / 73.2, 781 / 1,200 / 1,217 * | \* from #5615, not the same session. - `standard-v1` on the package **complete, 21 of 21, sanity 12/12**. `tools/needle_bench.py` **7/7** (32K about 28 s, 128K 99-112 s, 250K 238 s). 8 fresh engines read ~3.8K prompts at 680-746 tok/s. - The task set differs by one task: an IBAN check that passed on 0.2.12 fails on 0.2.13 (letters mapped to one digit instead of two in the mod-97 loop). A single task flips between runs of the same version on this set as well, so I don't read a regression into it. 41 of 114 answers are identical to 0.2.12 at temperature 0 (text and tool calls). **Not run:** DFlash2 (defaults for gfx1100 unchanged in 0.2.13; my 0.2.12 results are in #5615), the concurrency stage (one slot), IQ2_XS, vision. Raw data (all three `standard-v1` folders unedited, needles, pairs, fresh engines, task-set summary per category, version chart): https://github.com/faf-net/ugreen-idx6011-egpu-llm/tree/main/messen/daten/2026-11-09-v0.2.13 ### Reproduction status Measured locally
Related on this site
Quick links to related pages.