Skip to content

[Engine] No Windows HIP engine for gfx1032 (RX 6600/6600 XT/6650 XT) — the app's own message says "Ask for the gfx1032 build" #4455

Open

Description

@zxtemplar

Area

[Engine] (speed, kernels, models, the server and API)

Desktop app, Terminal or source?

Desktop app (Windows, Linux or Mac)

What happened

Strata starts, detects my GPU (AMD Radeon RX 6600, 8 GB), but cannot pick any
engine for it: the engine update list contains no Windows HIP build with code
for gfx1032. The app itself suggests asking for one, so here it is.

Full startup output:

Welcome to Strata
Strata found your graphics card:
AMD Radeon RX 6600, 8 GB

Strata can’t pick an engine for this PC from its update list.

This build of Strata is for AMD RX 6000 and RX 7000 and RX 9000 and Ryzen AI Max cards.
This computer has an AMD gfx1032 card (gfx1032). Ask for the gfx1032 build.

0.1.52 linux-x86_64-cuda-cuda12 is for linux-x86_64, this PC is windows-x86_64
0.1.52 linux-x86_64-cuda-cuda13 is for linux-x86_64, this PC is windows-x86_64
0.1.52 linux-x86_64-hip-halo is for linux-x86_64, this PC is windows-x86_64
0.1.52 linux-x86_64-hip-rx6000 is for linux-x86_64, this PC is windows-x86_64
0.1.52 linux-x86_64-hip-rx7900 is for linux-x86_64, this PC is windows-x86_64
0.1.52 linux-x86_64-hip-rx9070 is for linux-x86_64, this PC is windows-x86_64
0.1.52 linux-x86_64-sycl is for linux-x86_64, this PC is windows-x86_64
0.1.52 macos-arm64-metal is for macos-arm64, this PC is windows-x86_64
0.1.52 windows-x86_64-cuda-cuda12 is a cuda engine, this PC needs hip
0.1.52 windows-x86_64-cuda-cuda13 is a cuda engine, this PC needs hip
0.1.52 windows-x86_64-hip-rdna2 has code for gfx1030, gfx1031, this PC's GPU is gfx1032
0.1.52 windows-x86_64-hip-rdna3 has code for gfx1100, gfx1101, gfx1102, gfx1151, this PC's GPU is gfx1032
0.1.52 windows-x86_64-hip-rx9070 has code for gfx1200, gfx1201, this PC's GPU is gfx1032

What I expected instead: since windows-x86_64-hip-rdna2 already ships
gfx1030 and gfx1031, and RX 6600 / RX 6600 XT / RX 6650 XT (all gfx1032) fall
under the "RX 6000" umbrella the build advertises, I'd expect either a
gfx1032-capable Windows HIP engine (dedicated build, or gfx1032 added to the
rdna2 fatbin), or a documented fallback instead of a dead end.

Context that may help: gfx1032 is same-ISA RDNA2 as gfx1030. The community
has shipped Windows gfx1032 HIP libraries before (e.g. brknsoul/ROCmLibs —
"Prebuilt Windows ROCm Libs for gfx1031 and gfx1032", commonly used to get
Ollama/ComfyUI working on RX 6600-class cards on Windows), so a working
gfx1032 engine on Windows looks feasible even though AMD's official HIP SDK
for Windows doesn't ship gfx1032 codegen.

Smallest steps or command that reproduce it

  1. Install Strata 0.2.7 desktop app on Windows with an RX 6600 (gfx1032).
  2. Launch it.

The GPU is detected, the engine picker prints the mismatch table above, and
no engine can be installed. (Also HSA_OVERRIDE_GFX_VERSION=10.3.0 is not
set — noted since that override is the usual Linux workaround.)

Reproduction status

Reproduced locally

Strata version

Strata 0.2.7 (desktop app) · engine update list 0.1.52 (no engine installable on this PC)

Graphics card, memory and system

AMD Radeon RX 6600 8 GB (gfx1032) · 32 GB RAM · Windows 11 Pro (10.0.26200)

Model, size and settings

N/A — the engine picker blocks at startup, before any model can be downloaded
or started.

Logs

The startup output above is the complete message shown. The app's
Settings > Engine > Diagnostics > View logs file can be attached on request.

strata-issue-gfx1032.md

Activity

  1. Niko1221 commented on Oct 29, 2026

    @Niko1221
    Owner

    You are right: the Windows rdna2 package holds code for gfx1030 and gfx1031 only (the Linux rx6000 one too), so there is no engine for gfx1032 (RX 6600, 6600 XT, 6650 XT). A gfx1032 build needs a card to test it on, and we have none.

  2. cdolm121 commented on Oct 29, 2026

    @cdolm121

    Update: a gfx1032 build now runs on a real gfx1032 card.

    The build needed a card to test on, so I made one: RX 6600 (Navi 23, gfx1032), 8.0 GiB,
    wave32, driver 32.0.21045.16010, Windows 11 Pro 26H1 build 28000, Ryzen AI 9 HX 370 / 23.6 GB.

    Build: Strata 0.2.7 source, CMAKE_HIP_ARCHITECTURES=gfx1032, ROCm 10.2.0a20260930 +
    rocm_sdk_device_gfx1032, engine version 0.1.52 (the same version the rdna2 package carries).

    Engine log:

    strata generate: GPU 0: AMD Radeon RX 6600 (gfx1032)
    strata generate: HIP runtime ...\amdhip64_7.dll
    strata generate: 300 native projection matrices, 1935.94 MiB of weights, in 9.9 s
    strata generate: expert cache 866 slots, 0.85 GiB of VRAM
    strata generate: PCIe probe: 2.8 GB/s host->device -> pcie_frac 0.08
    strata generate: warning: Hadamard-folded experts are validated end to end on gfx1151 among AMD
                     GPUs, not yet on gfx1032
    strata serve: session is up (engine 0.1.52)
    

    Measured on the card:

    decode 5.8 – 7.2 tok/s
    prompt read 8 – 14 tok/s (first chunk), 25 – 28 tok/s (warm)
    expert slots 866 (0.85 GiB VRAM)
    PCIe host->device 2.8 GB/s
    weights resident 1936 MiB (300 native projection matrices)

    The desktop app loads it and answers (32K context, Gyro-S, drafting off):

    done: tokensIn 122, tokensOut 50, tps 5.8, promptTps 8, ttftMs 688, finishReason stop
    "RDNA2 是 AMD 的第二代显卡 GPU 架构,用于提升图形性能、能效和光线追踪能力。"
    

    Two notes, both about the card rather than the engine:

    • the GPU's WDDM budget gives the process 6302 of the card's 8176 MiB, so the engine reports
      ~645 MiB free — the same VRAM-headroom warning as on other 8 GB cards;
    • on a 32K prompt the engine plans ~676 GB of SSD reads (the 24.2 GiB expert set does not fit in
      the page cache), so long prompts are slow here.

    Full report material (AMD_HIP.md's "Reporting a Windows AMD run"):

    GPU          AMD Radeon RX 6600 (gfx1032), 8.0 GiB, wave32
    iGPU         AMD Radeon 890M (gfx1150) — off
    Driver       32.0.21045.16010
    OS           Windows 11 Pro 26H1 build 28000
    CPU/RAM      Ryzen AI 9 HX 370, 12C/24T, 23.6 GB
    App/Control  Strata 0.2.7 / control 0.1.53
    engine       0.1.52, Windows, hip, gfx1032 (self-built)
    

    So a gfx1032 package is buildable and testable now — this card can serve as the reference card
    for it, the same way rdna2 covers gfx1030 and gfx1031.

  3. yangweijie commented on Oct 30, 2026

    @yangweijie

    A pull request for the setup and Terminal side of this report: #4650.

    It adds gfx1032 to setup as an opt-in (STRATA_EXPERIMENTAL_GFX1032=1, the pattern gfx1103 and gfx1150
    already use), makes setup skip the ready-made Windows package for this arch (the rdna2 zip carries
    gfx1030/gfx1031 code only), and accepts an engine the owner builds with tools\hip\build_windows.bat by
    reading engine\BUILD.json. The Terminal gets 0x73FF, so strata launch can use a component package made
    from that build. The card report is docs/RX6600M.md. It is not #2784: the card stays opt-in until the feed
    carries a package for the arch.

    Measured on an RX 6600M (8 GiB, gfx1032), Windows 11 10.0.26100, engine 0.1.53 built from 46c152a1 with
    ROCm 10.2.0a20260930, Coder IQ1_M, 8,192-token context. The cache held 1,165 of the model's 12,288
    (layer, expert) pairs, so most routed experts ran on the CPU pool. tools/strata_bench.py, 3 runs per test:

    Test Result (median, range)
    short chat, 100 new tokens decode 8.10 tok/s (7.50-10.70, n=9)
    fresh 4K prompt, 4,019 tokens 46.0 tok/s (45.7-50.2, n=3)

    0 failed requests. Raw result and per-request data:
    bench/results/2026-10-30-community-rx-6600m-qwen3-8-flash-next-coder-iq1-m/ (the 16K prompt and the agent
    session skip themselves at an 8K context). The desktop RX 6600 report above is the second machine; this one
    is the 6600M.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions