Intel’s “Gaudi 3” AI accelerator chip may give Nvidia’s H100 a run @ ArsTechnica
Last week Intel announced their new AI Accelator chip they call Gaudi 3 as an alternative to the Nvidia H100. Of course, as you now Nvidia showed off their Blackwell system which basically took the H100 to an entire new level.
Compared to Nvidia's H100 chip, Intel projects a 50 percent faster training time on Gaudi 3 for both OpenAI's GPT-3 175B LLM and the 7-billion parameter version of Meta's Llama 2. In terms of inference (running the trained model to get outputs), Intel claims that its new AI chip delivers 50 percent faster performance than H100 for Llama 2 and Falcon 180B, which are both open-weights models.
What I find interesting about this new chip is that it was basically created because NVIDIA held the market share in AI processing and, is something that NVIDIA held for a VERY LONG time. It wasn't until ChatGPT and freely available LLMs hit the headlines that the AI race started to heat up.
I still remember when AMD told the world that their GPUs were better at playing games because they were dedicated to visual processing. Well, joke has been on them because the focus of the NVIDIA GPU was never be only on video games ever since the launch of CUDA. Combine that with Intel trying to enter the GPU market, and failing and we start to see a real push towards AI processing, and getting their piece of the pie.
On paper the Guadi 3 looks really impressive, HBM memory + massive processing core makes the chip faster on a hardware level but, they are still years behind on the software front and it would be pretty easy for NVIDIA to lock them out. Although, that might be the point, when NVIDIA locked out AMD from using SLI, they went full bore into Crossfire development and, in some ways created the better product.
As you can imagine, investors are not as tech savvy and basically are pulling their stakes in NVIDIA stock because they are no longer a "sure bet". Combine that with "me to" style of investing and one high-level investor selling their stake and you get a literal cascade of sell offs as a result.
I guess whatever is good for the investors is good for the market. Even if this Intel chip fails to sell, the damage was done leaving NVIDIA to do what it has always done, march on and take over.
Related Web URL: https://arstechnica.com/information-technology/202...