Benchmarks from the latest eBay special: W6800 (modded V620)
Summary
A user benchmarks a modded AMD V620 GPU flashed with W6800 firmware and a custom blower fan for running LLMs via Vulkan and ROCm backends, comparing performance on Qwen2.5-27B at various quantization levels.
Recently there was a guy selling modded V620s on eBay for a slight markup with two major changes: * Flashed with W6800 firmware, which enables a mini-displayport output. Unfortunately that disables some compute cores, although the W6800 has higher boost clocks. * Blower fan with custom 3D printed ABS shroud. There is no fan control built into the card but you can plug the fan into a motherboard fan header or external fan controller. I decided to pick one up because I have a spare micro atx PC lying around. This PC has no integrated graphics and can really only fit one card, so it would have been challenging to get a headless datacenter card running. The V620 is probably a better deal if you can run it as it has more compute, and the Tesla V100s are still the best deal if you want to stay in the CUDA ecosystem. Having said that, here are the benchmarks. Qwen 3.6 27B @ Q6\_K Vulkan (official llama.cpp build) ggml_vulkan: Found 1 Vulkan devices: ggml_vulkan: 0 = AMD Radeon Pro W6800 (RADV NAVI21) (radv) | uma: 0 | fp16: 1 | bf16: 0 | warp size: 32 | shared memory: 65536 | int dot: 1 | matrix cores: none | model | size | params | backend | ngl | test | t/s | | ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: | | qwen35 27B Q6_K | 21.30 GiB | 27.32 B | Vulkan | -1 | pp1024 | 297.94 ± 0.30 | | qwen35 27B Q6_K | 21.30 GiB | 27.32 B | Vulkan | -1 | tg256 | 20.35 ± 0.00 | | qwen35 27B Q6_K | 21.30 GiB | 27.32 B | Vulkan | -1 | pp1024 @ d8192 | 232.40 ± 0.32 | | qwen35 27B Q6_K | 21.30 GiB | 27.32 B | Vulkan | -1 | tg256 @ d8192 | 19.70 ± 0.00 | | qwen35 27B Q6_K | 21.30 GiB | 27.32 B | Vulkan | -1 | pp1024 @ d16384 | 185.07 ± 0.12 | | qwen35 27B Q6_K | 21.30 GiB | 27.32 B | Vulkan | -1 | tg256 @ d16384 | 19.18 ± 0.00 | ROCm (lemonade ROCm nightly build) ggml_cuda_init: found 1 ROCm devices (Total VRAM: 32095 MiB): Device 0: AMD Radeon Pro W6800, gfx1030 (0x1030), VMM: no, Wave Size: 32, VRAM: 32095 MiB | model | size | params | backend | ngl | test | t/s | | ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: | | qwen35 27B Q6_K | 21.30 GiB | 27.32 B | ROCm | -1 | pp1024 | 265.71 ± 0.02 | | qwen35 27B Q6_K | 21.30 GiB | 27.32 B | ROCm | -1 | tg256 | 18.69 ± 0.01 | | qwen35 27B Q6_K | 21.30 GiB | 27.32 B | ROCm | -1 | pp1024 @ d8192 | 246.81 ± 0.03 | | qwen35 27B Q6_K | 21.30 GiB | 27.32 B | ROCm | -1 | tg256 @ d8192 | 18.15 ± 0.02 | | qwen35 27B Q6_K | 21.30 GiB | 27.32 B | ROCm | -1 | pp1024 @ d16384 | 230.19 ± 0.06 | | qwen35 27B Q6_K | 21.30 GiB | 27.32 B | ROCm | -1 | tg256 @ d16384 | 17.64 ± 0.02 | This probably wont be a surprise to anyone who runs AMD, but Vulkan is faster at TG while ROCm is faster at PP, particularly over long context depths. Now for some Q4 benchmarks for more of a comparison to the 24GB VRAM class. Qwen 3.6 27B @ Q4\_K\_XL Vulkan (official llama.cpp build) ggml_vulkan: Found 1 Vulkan devices: ggml_vulkan: 0 = AMD Radeon Pro W6800 (RADV NAVI21) (radv) | uma: 0 | fp16: 1 | bf16: 0 | warp size: 32 | shared memory: 65536 | int dot: 1 | matrix cores: none | model | size | params | backend | ngl | test | t/s | | ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: | | qwen35 27B Q4_K - Medium | 16.67 GiB | 27.32 B | Vulkan | -1 | pp1024 | 353.85 ± 0.04 | | qwen35 27B Q4_K - Medium | 16.67 GiB | 27.32 B | Vulkan | -1 | tg256 | 24.73 ± 0.00 | | qwen35 27B Q4_K - Medium | 16.67 GiB | 27.32 B | Vulkan | -1 | pp1024 @ d8192 | 265.14 ± 0.34 | | qwen35 27B Q4_K - Medium | 16.67 GiB | 27.32 B | Vulkan | -1 | tg256 @ d8192 | 23.77 ± 0.00 | | qwen35 27B Q4_K - Medium | 16.67 GiB | 27.32 B | Vulkan | -1 | pp1024 @ d16384 | 205.36 ± 0.67 | | qwen35 27B Q4_K - Medium | 16.67 GiB | 27.32 B | Vulkan | -1 | tg256 @ d16384 | 23.03 ± 0.00 | ROCm (lemonade ROCm nightly build) ggml_cuda_init: found 1 ROCm devices (Total VRAM: 32095 MiB): Device 0: AMD Radeon Pro W6800, gfx1030 (0x1030), VMM: no, Wave Size: 32, VRAM: 32095 MiB | model | size | params | backend | ngl | test | t/s | | ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: | | qwen35 27B Q4_K - Medium | 16.67 GiB | 27.32 B | ROCm | -1 | pp1024 | 328.96 ± 0.09 | | qwen35 27B Q4_K - Medium | 16.67 GiB | 27.32 B | ROCm | -1 | tg256 | 21.40 ± 0.01 | | qwen35 27B Q4_K - Medium | 16.67 GiB | 27.32 B | ROCm | -1 | pp1024 @ d8192 | 298.96 ± 0.01 | | qwen35 27B Q4_K - Medium | 16.67 GiB | 27.32 B | ROCm | -1 | tg256 @ d8192 | 20.68 ± 0.03 | | qwen35 27B Q4_K - Medium | 16.67 GiB | 27.32 B | ROCm | -1 | pp1024 @ d16384 | 275.02 ± 0.03 | | qwen35 27B Q4_K - Medium | 16.67 GiB | 27.32 B | ROCm | -1 | tg256 @ d16384 | 20.02 ± 0.03 | Unfortunately llama-bench massively lags behind in features compared to llama-server so I can't use it to benchmark MTP, but it is a massive boost! Like 75-100% TG increase. That makes this card VERY usable. Curious to know how this compares to a single MI50 now that V620 is the better deal. Every benchmark I found was for at least 2 x MI50 though.
Similar Articles
Running Qwen 3.6 35B A3B-Q8_0 gguf on a cheap radeon 7600 at 18 token/s * update increased to 21 t/s
User reports running Qwen 3.6 35B A3B-Q8_0 gguf on a Radeon 7600 with llama.cpp and ROCm, achieving 21 tokens per second after VRAM overclocking, with a note about a display-related performance bug.
2× Radeon R9700 — Qwen 3.6 27B Q8 MTP on llama.cpp
Technical report on running Qwen 3.6 27B Q8 model on a dual AMD Radeon R9700 setup using llama.cpp with ROCm, including performance benchmarks and configuration details.
Testing Gemma 4 & Qwen 3.6 MoE on AMD 6800H (iGPU/UMA) - Performance Breakdown
Benchmarks of Gemma 4 and Qwen 3.6 MoE models on AMD 6800H iGPU using llama.cpp Vulkan backend show that MoE models and lower-bit quantization (Q4_0) provide the best performance, while Q8_0 is too slow for large models.
7900XTX 24GB vram, can finally fit Q6K+MTP with Qwen 3.6 27B at 131k context
A guide on optimizing VRAM usage on an AMD 7900XTX to run a 27B Qwen model with Q6K quantization and 131k context by compiling llama.cpp with OpenBLAS and CUDA_FA_ALL_QUANTS, and using kvcache quantization at q5_0/q4_0.
I benched quad 5060Tis for code generation with Qwen3.6-27B so you don't have to (it's really good)
The author details benching a quad 5060Ti setup for running Qwen3.6-27B at Q8 with FP16 KV and MTP for code generation, concluding it offers good performance per dollar compared to alternatives like dual 3090s or modded 3080s.