Running DeepSeek-R1 14B (Q4_K_M) at 33.5 tokens/s on an Intel Arc B580! For many people, it could be an alternative way to access OpenAI's paid subscription.@Intel pic.twitter.com/X4JilAos0O
— Xiao Yang (@XYang2023) February 8, 2025
Array ( [content] => [params] => Array ( [0] => /forum/threads/intel-arc-b580.21735/page-4 ) [addOns] => Array ( [DL6/MLTP] => 13 [Hampel/TimeZoneDebug] => 1000070 [SV/ChangePostDate] => 2010200 [SemiWiki/Newsletter] => 1000010 [SemiWiki/WPMenu] => 1000010 [SemiWiki/XPressExtend] => 1000010 [ThemeHouse/XLink] => 1000970 [ThemeHouse/XPress] => 1010570 [XF] => 2021770 [XFI] => 1050270 ) [wordpress] => /var/www/html )
Running DeepSeek-R1 14B (Q4_K_M) at 33.5 tokens/s on an Intel Arc B580! For many people, it could be an alternative way to access OpenAI's paid subscription.@Intel pic.twitter.com/X4JilAos0O
— Xiao Yang (@XYang2023) February 8, 2025
I used the Intel Arc B580 for an LLM training run, which took 6.5 minutes to complete. The same training code took 0.86 minutes on the Nvidia A100 and 1.83 minutes on the Nvidia L4. This further demonstrates that the B580 is a viable option for running such tasks on a budget,… pic.twitter.com/orZSOXdXtx
— Xiao Yang (@XYang2023) February 8, 2025
I think B580 Lacked the VRAM that held it's compute back otherwise it's 1/3 the compute of A100I used the Intel Arc B580 for an LLM training run, which took 6.5 minutes to complete. The same training code took 0.86 minutes on the Nvidia A100 and 1.83 minutes on the Nvidia L4. This further demonstrates that the B580 is a viable option for running such tasks on a budget,… pic.twitter.com/orZSOXdXtx
— Xiao Yang (@XYang2023) February 8, 2025
I think B580 Lacked the VRAM that held it's compute back otherwise it's 1/3 the compute of A100
Using the same model and prompt (llama3.2:1B and 'write js code hello world'), Alex got 212 tokens/s on an RTX 5080, while I got 170.11 tokens/s on an Intel Arc B580. Looking forward to seeing the Intel Arc B770, which has 32 Xe cores compared to 20 on the B580.@Intel
— Xiao Yang (@XYang2023) February 9, 2025
It is not that far from A100!I think B580 Lacked the VRAM that held it's compute back otherwise it's 1/3 the compute of A100
My latest update. In short, I reduced the training time from 6.5 minutes to 1.35 minutes, just 0.5 minutes behind Nvidia A100! The Youtube video covers it in details.https://t.co/PEjtwY2uGj
— Xiao Yang (@XYang2023) February 10, 2025
That is what I felt about it after going through the exercise. They should be confident about their products and start competing aggressively.So what, pray tell, are Intel doing with all this Falcon/Jaguar Shores reshuffling if their consumer GPUs are already this competent? I’ve heard it’s to “refocus Jaguar Shores on rack-scale system abilities” so is that the big delay? The core GPU IP and overall system seem quite competent now… gosh, I wish this company could just get it together. It’s all at their fingertips if only they would execute on a good plan for once.
Tested @intel Arc B580 GPU with PyTorch 2.7.0.dev20250211+xpu (nightly) for LLM training/fine-tuning. Seamless transition from NVIDIA—just change device value from "cuda" to "xpu". Check it out:https://t.co/PEjtwY1WQL
— Xiao Yang (@XYang2023) February 12, 2025
RAG with Deepseek R1 14B on an @intel Arc B580 GPU. I really don't think you need a $2K AMD-based Framework desktop for AI-related tasks. An all-Intel machine costing around $1K should do the job. Maybe I can create a video on how to configure such a machine. pic.twitter.com/P56AF207rj
— Xiao Yang (@XYang2023) February 26, 2025
Finally finished the video. IMO, @intel should address competition head-on by leveraging its diverse product portfolio.@MJHolthaus
— Xiao Yang (@XYang2023) March 4, 2025
Performant AI Workstation Setup for Around $1,100 with Intel Core Ultra 7 265K/265 & Intel Arc B580https://t.co/b6QaGTV1Qe
I think this is an interesting slide for comparing the @intel Arc B580 and AMD 9070 for realistic AI tasks. The 12GB VRAM is not a disadvantage here, as tasks can be offloaded to the integrated GPU in the Intel Core Ultra 7 265K. Please check out my video for a detailed… pic.twitter.com/5Xz8evj2Lp
— Xiao Yang (@XYang2023) March 4, 2025