For big AI models who can't get NVIDIA GPUs, AMD chips are here!

AMD’s “ultimate weapon” is here!

On June 13th, Eastern Time, AMD announced at the "AMD Data Center and Artificial Intelligence Technology Premiere" that it will launch the most advanced artificial intelligence GPU MI300X chip to date, which can accelerate ChatGPT and other chat robots. It uses generative artificial intelligence for processing speed and can use up to 192GB of memory.

In contrast, NVIDIA's H100 chip only supports 120GB of memory, which is generally believed to mean that NVIDIA's dominant position in this emerging market may be challenged. The MI300X will begin shipping to select customers later this year.

Although investors expect AMD to compete with NVIDIA in the field of AI chips, analysts have not followed the trend and blown away this powerful weapon of AMD. Instead, they point out very sensibly that AMD wants to compete with NVIDIA in the field of artificial intelligence chips. There is still a long way to go in the field of smart chips to challenge NVIDIA's industry leadership, and this chip alone cannot do it.

Citi chip analyst Chris Danely said in the latest report that AMD's MI300 chip seems to have achieved a huge design victory, but considering the performance limitations and history of failure, he is skeptical about the sustainability of graphics/CPU ICs. Xing expressed doubts, "While we expect AMD to continue to gain market share from Intel, its new Genoa products appear to be growing slower than expected."

On the same day, Karl Freund, founder and chief analyst of Cambrian-AI Research LLC, also wrote in Forbes that Although AMD’s newly launched chip has aroused huge interest from all parties in the market, it is not the same as Nvidia’s. Compared with the H100 chip, MI300X faces some challenges, mainly in the following four aspects:

NVIDIA began full shipping of the H100 today, and the company has by far the largest ecosystem of software and researchers in the AI ​​industry.

Second, although the MI300X chip offers 192GB of memory, Nvidia will soon catch up at this point, and may even overtake it in the same time frame, so this is not a big advantage. The MI300X will be very expensive and won't have a significant cost advantage over Nvidia's H100.

The third is the real key: MI300 does not have a Transformer Engine (a library for accelerating Transformer models on NVIDIA GPUs) like H100, which can double the performance of large language models (LLM). If it takes a year to train a new model with thousands of (NVIDIA) GPUs, then training with AMD hardware may take another 2-3 years, or invest 3 times as many GPUs to solve the problem.

Finally, AMD has not yet disclosed any benchmarks. But performance when training and running LLM depends on the system design and GPU, so I’m looking forward to seeing some comparisons with industry competitors later this year.

However, Freund also added that MI300X may become an alternative to Nvidia’s GH200 Grace Hopper super chip. Companies like OpenAI and Microsoft need such alternatives, and while he doubts AMD will give these companies an offer they can't refuse, AMD won't take much market share away from Nvidia.

AMD’s share price has risen by 94% since the beginning of this year. Yesterday, the US stock market closed down 3.61%, while its peer Nvidia closed up 3.90%, with its market value closing above the US$1 trillion mark for the first time.

In fact, Nvidia is ahead not only because of its chips, but also because of the software tools they have provided to artificial intelligence researchers for more than a decade. Moor Insights & Strategy analyst Anshel Sag said: "Even if AMD is competitive in terms of hardware performance, people are still not convinced that its software solutions can compete with Nvidia."

is expected to drive healthy competition among technology companies

American technology critic Billy Duberstein pointed out on the 12th that with the popularity of ChatGPT, AI seems to have opened up a new field, and a competition has begun among leading companies in many industries to capture market share. Whether Nvidia or AMD wins this race, every technology company will benefit.

Duberstein said potential customers are very interested in the MI300. According to a June 8 article in Digital Times that he quoted, data center customers are desperately looking for alternatives to Nvidia products. Nvidia currently occupies a dominant position in the high-profit and high-growth market of artificial intelligence GPUs, accounting for 60% to 70% of the market share in the field of AI servers.

Duberstein went on to point out that given the current high price of Nvidia H100 servers, data center operators would like to see Nvidia have a third-party competitor, which would help reduce the price of AI chips. Therefore, this is a huge advantage for AMD and a challenge for Nvidia. This can lead to good profitability for every market participant.

Last month, Morgan Stanley analyst Joseph Moore adjusted his forecast for AMD's artificial intelligence revenue, saying it could be "several times higher" than initially expected.

Chinese companies competing for the GPU AI industry chain are expected to continue to benefit from it

Currently, major technology companies around the world are competing for AI tickets. Compared with overseas giants, Chinese large technology companies are more urgent to purchase GPUs.

According to "LatePost", after the Spring Festival this year, major Chinese Internet companies with cloud computing businesses have placed large orders with NVIDIA. Byte has ordered more than $1 billion in GPUs from Nvidia this year, and another large company's order has also exceeded at least 1 billion yuan. And Byte's orders this year alone may be close to the total number of commercial GPUs sold by Nvidia in China last year.

Guojin Securities continues to be optimistic about AMD’s new moves and continues to be optimistic about the AI ​​industry chain. The agency believes that the continued popularity of generative AI has driven strong demand for AI chips. NVIDIA's second-quarter data center business guidance has exceeded expectations, and TSMC's 4nm, 5nm, and 7nm capacity utilization has increased significantly, which are the best proof.

Guojin Securities pointed out that according to industry chain research, the leading optical module-DSP chip manufacturer stated that AI-related business is expected to grow rapidly in the future; the leading CPU/GPU heat sink company stated that AI GPU experienced significant quarter-on-quarter growth in the second and third quarters. Leading companies said that in the next quarter, demand for server PCIe Retimer chips will accelerate and gradually increase. It is expected that between 2024 and 2025, new products in the AI ​​industry will promote the development of the AI ​​industry chain and bring sustained economic benefits.

The above is the detailed content of For big AI models who can't get NVIDIA GPUs, AMD chips are here!. For more information, please follow other related articles on the PHP Chinese website!