Comments on: With Gaudi 3, Intel Can Sell AI Accelerators To The PyTorch Masses https://www.nextplatform.com/2024/04/09/with-gaudi-3-intel-can-sell-ai-accelerators-to-the-pytorch-masses/ In-depth coverage of high-end computing at large enterprises, supercomputing centers, hyperscale data centers, and public clouds. Wed, 22 May 2024 21:36:55 +0000 hourly 1 https://wordpress.org/?v=6.5.5 By: mulcas https://www.nextplatform.com/2024/04/09/with-gaudi-3-intel-can-sell-ai-accelerators-to-the-pytorch-masses/#comment-224666 Wed, 22 May 2024 21:36:55 +0000 https://www.nextplatform.com/?p=143942#comment-224666 In reply to Calamity Jim.

Gaudi 3 is a purposed build AI accelerator, not a GPU as Nvidia H200 or MI300X. The target of Gaudi is purely Deep Learning and GenAI, they are not looking to reach HPC markets, so FP64 is not required. This could be considered a disadvantage, but if one thinks merely on AI, then Gaudi is quite competitive. For HPC, we should wait for Intel’s next gen.

]]> By: Harish https://www.nextplatform.com/2024/04/09/with-gaudi-3-intel-can-sell-ai-accelerators-to-the-pytorch-masses/#comment-223112 Sat, 13 Apr 2024 20:11:39 +0000 https://www.nextplatform.com/?p=143942#comment-223112 In reply to Calamity Jim.

The absence of Gaudi4,5,6 in the roadmap may work against Gaudi3 to some extent. It is too risky for Intel to abandon Gaudi line for unknown ‘Shores’

]]> By: Calamity Jim https://www.nextplatform.com/2024/04/09/with-gaudi-3-intel-can-sell-ai-accelerators-to-the-pytorch-masses/#comment-222960 Tue, 09 Apr 2024 18:38:11 +0000 https://www.nextplatform.com/?p=143942#comment-222960 Seeing how capable this Gaudi 3 is (a bit more than twice Gaudi2, or 4x Gaudi1), I got to wonder what Intel still needs Falcon Shores for (why not Gaudi 4, 5, 6, …)? But then, reading this article, it dawned on me: it doesn’t have FP64! Merging the Xᵉ GPU Max’s FP64 vector engines with Gaudi’s FP32-FP16-FP8 tensor and matrix units will likely be challenging, but is needed to fully compete with the MI300s and GB200s of the world. The result should be most interesting IMHO.

]]>