Issues / #3639
#3639 [Benchmark] 2x RX 6700 XT (gfx1031) on Ubuntu 26.04, Qwen IQ2_XS - works, 43 tok/s decode
open · @drtobiasarnold-tech · 2 comments · View on GitHub
BenchmarksSetup & installMulti-GPUAMD / HIPModels & quants
Description
Report from an RX 6700 XT owner, since the rx6000 package is listed as not yet tested on real RX 6000 hardware. **System** - GPUs: 2x Radeon RX 6700 XT 12 GB (gfx1031): Sapphire Nitro+ in PCIe 4.0 x16, PowerColor in PCIe 3.0 x4 (chipset slot) - CPU: Ryzen 7 5800X, RAM: 64 GB DDR4-3200 - OS: Ubuntu 26.04.1 LTS, kernel 7.0.0-34-generic, amdgpu (in-kernel) - Strata: [VERSION from `strata 0.2.2 (build 0.1.47; python 3.12.15`], installed via the Terminal (`strata launch`) - Model: qwen IQ2_XS, context 32K, both cards used **Result: works out of the box.** - Decode: 42-46 tok/s on longer answers (43.6 tok/s on a 2,580-token reply) - Prefill: 127-189 tok/s - VRAM: 11.5 + 11.8 GB, experts in VRAM 14.5 GB, VRAM hit rate ~90% - System RAM in use: 53 of 61 GB - No crashes or hangs over six requests so far **Quality note:** For German medical writing (turning dictated notes into a discharge summary), this was the most accurate of five local models I tested: no invented facts, correct order of follow-up steps, fluent prose. Thanks for supporting the RX 6000 series.
Related on this site
Quick links to related pages.