I recently purchases an Intel ARC B580 for $320 CAD (~$225 USD). I received also a free game (Battlefield 6) valued at $75 CAD (~$53 USD). This is an absolutely insane deal for a very reasonable GPU.
I began testing it. The first test is with llama.cpp and model Qwen3-8B-Q4_K.
The ARC B580 uses the SYCL backend on Windows, whereas the RTX 3090 uses CUDA (cuBLAS) in Ubuntu Linux
> .\llama-bench.exe -m ..\Qwen3-8B-Q4_K_M.gguf -ngl 99 --threads 8 -p 512,1024,2048 -n 128,256,512 -sm none -mg 0
Device: Intel ARC B580
| test | t/s |
| pp512 | 794.17 ± 2.06 |
| pp1024 | 769.64 ± 0.78 |
| pp2048 | 743.25 ± 0.21 |
| tg128 | 48.41 ± 0.15 |
| tg256 | 48.32 ± 0.07 |
| tg512 | 48.17 ± 0.08 |
Device: RTX 3090 (Power Limit 190W)
| test | t/s |
| --------------: | -------------------: |
| pp512 | 2517.93 ± 209.39 |
| pp1024 | 2375.25 ± 53.74 |
| pp2048 | 2180.69 ± 12.87 |
| tg128 | 68.85 ± 0.27 |
| tg256 | 67.07 ± 0.60 |
| tg512 | 65.14 ± 0.68 |
Device: RTX 3090 (Power Limit 300W)
| test | t/s |
| --------------: | -------------------: |
| pp512 | 4248.31 ± 15.44 |
| pp1024 | 3965.97 ± 31.79 |
| pp2048 | 3702.01 ± 7.88 |
| tg128 | 129.63 ± 0.22 |
| tg256 | 126.23 ± 0.14 |
| tg512 | 121.94 ± 0.29 |
I would argue that the ARC B580 is quite respectable. More testing to follow!
No comments:
Post a Comment