GMKtec EVO-X3

Strix Halo LLM upgrades: what got faster and what failed

By Darren Soothill · Soot / Silicon

Source published

Aug 11, 2026

Test performed

Aug 11, 2026

Added to MiniPC.guru

Sep 17, 2026

Source reviewed

Sep 17, 2026

Tested hardware

AMD Ryzen AI Max+ 395 · Radeon 8060S, gfx1151 · 128 GB unified memory · not applicable: token generation does not read from disk once the model is resident. The measured unit carried the 4 TB drive; the figure is shown here because the protocol does not use the disk, not because a 2 TB unit was tested

Software / version / OS

llama.cpp, ROCm backend, built with GGML_HIP_GRAPHS=ON and run with HIP_LAUNCH_BLOCKING=1 · llama.cpp b10362, commit 4801e3c567d5; ROCm 7.14.60850 · Ubuntu 24.04.4, kernel 6.17.0-40

Power profile

not reported by source: the author documents elsewhere in the same series that the chassis power mode cannot be read from Linux

Test conditions

DeepSeek V4 Flash UD-IQ3_XXS, 97.05 GiB across four shards. Flags -ngl 99 -fa 1 -p 2048 -n 64 -r 1 -mmp 0. This is the shallow case, with no added context depth: 14.10 tok/s generated, 145.01 tok/s prefill. The same build compiled with HIP graphs disabled gave 10.35 tok/s, so the 26 per cent difference is a compile option rather than a version change

Editorial review checks the source and documentation; it does not certify the measurement or mean we tested the device ourselves.

Strix Halo LLM upgrades: what got faster and what failed

By Darren Soothill · Soot / Silicon

Source published

Aug 11, 2026

Test performed

Aug 11, 2026

Added to MiniPC.guru

Sep 17, 2026

Source reviewed

Sep 17, 2026

Tested hardware

AMD Ryzen AI Max+ 395 · Radeon 8060S, gfx1151 · 128 GB unified memory · not applicable: token generation does not read from disk once the model is resident. The measured unit carried the 4 TB drive; the figure is shown here because the protocol does not use the disk, not because a 2 TB unit was tested

Software / version / OS

llama.cpp, ROCm backend, built with GGML_HIP_GRAPHS=ON and run with HIP_LAUNCH_BLOCKING=1 · llama.cpp b10362, commit 4801e3c567d5; ROCm 7.14.60850 · Ubuntu 24.04.4, kernel 6.17.0-40

Power profile

not reported by source: the author documents elsewhere in the same series that the chassis power mode cannot be read from Linux

Test conditions

DeepSeek V4 Flash UD-IQ3_XXS, 97.05 GiB across four shards. Flags -ngl 99 -fa 1 -p 2048 -n 64 -r 1 -mmp 0 plus -d 24576, the deep case: 10.77 tok/s generated, 88.38 tok/s prefill

Editorial review checks the source and documentation; it does not certify the measurement or mean we tested the device ourselves.

Strix Halo LLM upgrades: what got faster and what failed

By Darren Soothill · Soot / Silicon

Source published

Aug 11, 2026

Test performed

Aug 11, 2026

Added to MiniPC.guru

Sep 17, 2026

Source reviewed

Sep 17, 2026

Tested hardware

AMD Ryzen AI Max+ 395 · Radeon 8060S, gfx1151 · 128 GB unified memory · not applicable: token generation does not read from disk once the model is resident. The measured unit carried the 4 TB drive; the figure is shown here because the protocol does not use the disk, not because a 2 TB unit was tested

Software / version / OS

vLLM serving the completions API · vLLM 0.27.0, commit 4bdc8a788d2e, wheel built locally for Python 3.12 against ROCm 7.14, SHA-256 020fac1682b2e500bbb327466d01a552fe12a3e750843ded5cd599155b0ced50 · Ubuntu 24.04.4, kernel 6.17.0-40

Power profile

not reported by source: the author documents elsewhere in the same series that the chassis power mode cannot be read from Linux

Test conditions

Qwen3.5-122B-A10B GPTQ Int4 snapshot on ROCm 7.14 with the tuned mixture-of-experts registers, a 4096-token server limit, two warm-up waves and ten measured waves, 64 output tokens and seed 12345. With one concurrent client: 10.25 completion tok/s, mean time to first token 1.115 ms, mean end-to-end 6.244 ms; all thirty measured requests completed

Editorial review checks the source and documentation; it does not certify the measurement or mean we tested the device ourselves.