Can Intel Gaudi 3 Help Its Comeback Efforts?

This new enterprise AI would be deemed to be open and more secure, Intel is wooing new customers, partners and collaborations across the AI space.

Intel Gaudi 3 AI accelerator deliver 50% on average better inference and 40% average better power efficiency than Nvidia H100, and this is only a fraction of the cost of H100.

Intel also announced the Gaudi 3 availability to original equipment manufacturers (OEMs), the OEMs include names like Dell Technologies, Hewlett Packard Enterprise, Lenovo and Supermicro, with this new offerings, we could see broadening the AI data center market offerings for enterprises.

But I think what I as an investor would be interested is whether this new AI accelerator could bring interest and sentiment back for Intel, how would this help Intel to come back to the AI race?

In this article, I will try to look at how market perceive this new AI accelerator, how Intel would plan to sell this?

The next questions would be how many Gaudi 3 accelerators can Intel make? How much would it cost? What can the product being shipped?

Intel which is working with Taiwan Semiconductor Manufacturing to etch and package the Gaudi 3 accelerators. We will need to look at how TSM would prioritise this production.

Gaudi 1 accelerator was launched back in July 2019 by Habana Labs which was acquired by Intel in December 2019 for $2 billion. Gaudi 1 was to compete with Nvidia’s “Volta” V100 then in a ridiculously smaller AI space.

This time, it will be different, as Intel need to be more aggressive in making and selling Gaudi 3 chips than it has been with either Gaudi 1 or Gaudi 2. We might also need to expect that Intel Gaudi would significantly be behind Nvidia when it comes to raw accelerator performance.

Intel need to seize the opportunity to sell aggressive and attract the PyTorch crowd, with Llama LLMs and a seemingly zillion other AI models available on Hugging Face. So the partnership formed up would be important.

As some of us might already know, it is simple to find out the cost of Gaudi 3 accelerators, it will be directly proportional to the performance of the Nvidia “Hopper” H100 GPU accelerator that it completes against and the street price that Hopper H100 GPUs have with 96 GB of HBM3 memory capacity and 3.9 TB/sec of bandwidth.

As the Hopper H200 variants, which have 141 GB of HBM3e memory capacity and 4.8 TB/sec of bandwidth, start shipping in a few months, the math will shift to a comparison with the H200s.

Gaudi also comes with some big-name backing, including this statement from stability.ai: “Alternatives like Intel Gaudi accelerators give us better price performance, reduced lead time and ease of use with our models taking less than a day to port over.”

With software already available on the Intel Developer Cloud, it promises developers that they can build their software for Gaudi 2 today and then “migrate it seamlessly” to Gaudi 3 on its release.

Actually the timing for the Gaudi 3 accelerator launch would be better for Intel to be in the spring of 2022 alongside with Nvidia Hopper GPU, this is because two years gap later, we could already see the demand or Nvidia GPUs is so strong that there is never a better time to sell a technology that should already be two years old.

Intel would need to sell all of Gaudi 3 and Falcon Shores that it can make in the short term. Any further delays in the production could cost more damage as Nvidia will ramps mightily and AMD starts getting traction, too, with its “Antares” Instinct MI300 series of GPU accelerators.

The Gaudi 3 accelerator is an big evolutionary jump from the Gaudi 2. But the big architectural change is coming with Falcon Shores, but one that will largely be insulated from customers using PyTorch and higher models upon that framework.

It would have been good for Habana Labs and then Intel if Meta Platforms had chosen the Gaudi accelerators as its AI engines, but clearly that did not happen and Meta Platforms is creating its own line of MTIA accelerators for AI training and inference and in the interim is buying hundreds of thousands of Nvidia GPUs.

With potential risk of missing out of sales to pytorch masses, which many would have already start building their models on Nvidia GPUs, even Meta is creating its own line of MTIA accelerators for AI training and inference and buying thousands of GPUs from Nvidia.

So the market share that Intel is trying to capture with missed timing of launch could cost them dearly. As seen from the technical analysis below, Intel is trying to come out from pullback, but it did not manage to, the best is only to stay sideway.

MACD is showing a downward trend, meaning that traders and investors are not taking much interest despite the new launch, as seen from the selling sentiment, it has been on the high side as well.

I personally think that we should look at how orders would start to form the demand for Intel new Gaudi 3, before we make any move.

Intel is facing an uphill task of trying to capture the hot AI accelerator market, with a two year gap launch than Nvidia, to catch up with the demand that Nvidia has build, is quite hard, though not impossible.

The next few weeks and months before Nvidia “Blackwell” B100 and B200 GPU accelerators start shipping later this year, how much can Intel gather in terms of demand would be important for its stock price?

I would watch out for any news of order for Intel Gaudi 3, then we shall discuss again on how Intel could fulfill the orders, that is a different discussion again.

Appreciate if you could share your thoughts in the comment section whether you think Intel could make a comeback with Gaudi 3 accelerator?

Disclaimer: The analysis and result presented does not recommend or suggest any investing in the said stock. This is purely for Analysis.