Intel Brings Gaudi2 AI Accelerators To China As An Alternative To NVIDIA A800 Chips

By: Hassan Mujtaba

Intel has announced that the company is bringing its Habana Gaudi2 accelerators for AI & HPC to the Chinese market.

Intel Announces Habana Gaudi2 Accelerator Launch For China As AI Demand Surges Amidst Potential “Extended” Chip Ban

The Habana Labs Gaudi2 accelerator was launched last year and the company claimed that the new chip offered training performance faster than the NVIDIA Ampere A100 GPU. The company even put out some internal performance estimates and stated that the Gaudi2 is the one & only alternative to NVIDIA GPUs for LLM Training. This was showcased within MLPerf benchmarks that showed up to a 69% gain in performance over the A100.

As the US seeks to tighten export controls on AI chips to China, Intel held a press conference in Beijing on the July 11, announcing the launch of the deep learning processor Habana Gaudi 2 in the Chinese market. This processor can be used to accelerate AI training and inference tasks. Chinese AI server giants, such as Inspur, New H3C, and xFusion, are expected to release server products featuring the Gaudi 2 processor.

via DigiTimes Asia

Intel has also joined hands with BCG to utilize the same Gaudi2 accelerators within Generative AI workloads. Considering how popular Generative AI has become within a year, this has raised the demand for GPUs that can handle AI well. NVIDIA is the undisputed leader within the AI segment but this has also caused supply issues across the globe. The massive demand for NVIDIA’s GPUs that are capable of doing AI tasks has led to an opportunity that Intel and AMD want to take a slice from.

We just reported how a potential “Extended” chip ban on China can further increase the prices of the already super-expensive NVIDIA AI GPUs within the country. The GPUs cost more than what it costs to acquire them in the US and these aren’t even the full variants. For China, NVIDIA made a cut-down design under the guise of A800 and H800 and we have also heard that orders which are placed now can take up to a year to be fulfilled.

Considering all of this, Intel announced during a press conference in Beijing, China, that the blue team is bringing its Gaudi2 AI accelerators to the country through its partners such as Inspur, H3C, and X-Fusion. Just like NVIDIA’s custom solution for China markets to meet US regulations, the Guadi2 accelerator will also get a custom version that will specifically be optimized for LLMs (large language models).

Just two of the HPC companies, Inspur & H3C, hold around 50% of the market share in China and Intel working with them could make a big dent in NVIDIA’s AI GPU demand if the performance is there but real-market reports show that Gaudi2 hasn’t been a very popular AI product within the year of its launch with no major deals with large customers. The BCG deal was the one instance & that too came a year later after the chip’s announcement. But the opportunity is there and Intel is also working to bring its next-generation Gaudi3 accelerators along with GPU-based AI solutions by 2025.

Written by Hassan Mujtaba