On Wednesday, AMD released its highly anticipated new MI300 series of AI chips, including MI300A and MI300X chips, targeting this market dominated by NVIDIA. The new chip released by AMD has more than 150 billion transistors. The memory of the new chip is 2.4 times that of Nvidia's H100 product, and the memory bandwidth is 1.6 times that of H100. AMD has nearly doubled its forecast for the AI ​​accelerator market size in 2027 compared with August.

On Wednesday, AMD released its highly anticipated new MI300 series of AI chips, including MI300A and MI300X chips, targeting this market dominated by NVIDIA. Such chips are better than traditional computer processors at processing the large data sets involved in AI training.

This new product launch is one of the most important in AMD's 50-year history and is expected to challenge Nvidia's position in the hot artificial intelligence accelerator market.

The new chips released by AMD have more than 150 billion transistors. The MI300X accelerator supports up to 192GB of HBM3 memory.

The memory of MI300X is 2.4 times that of NVIDIA H100 products, and the memory bandwidth is 1.6 times that of H100, further improving performance.


The performance of the new MI300X chip can be improved by up to 60% compared to Nvidia's H100. In a one-to-one comparison with the H100 (Llama270B version), MI300X performance improved by up to 20%. In a one-to-one comparison with the H100 (FlashAttention2 version), performance improved by up to 20%. In an 8-to-8 server comparison with H100 (Llama270B version), performance increased by up to 40%. In an 8-on-8 server comparison with H100 (Bloom176B), performance increased by up to 60%.


AMD CEO Lisa Su said that the new chip is equivalent to the H100 in terms of its ability to train artificial intelligence software. In terms of inference, that is, the process of running the software after it is put into actual use, it is much better than competing products.

With the popularity of artificial intelligence, the market is in great demand for high-end chips. This allows chip manufacturers to target this lucrative market and accelerate the launch of high-quality AI chips.

Although the competition in the entire AI chip market is quite fierce, AMD gave a bold and shocking prediction on the future market size on Wednesday, believing that the AI ​​chip market will expand rapidly. Specifically, the artificial intelligence (AI) chip market is expected to reach more than $400 billion by 2027, nearly double the $150 billion forecast in August, highlighting the rapidly changing expectations for AI hardware.


AMD is increasingly confident that its MI300 series can win over some technology giants, which could lead to billions of dollars of spending by these companies on AMD products. AMD said that Microsoft, Oracle and Meta are all its customers.

News on the same day showed that Microsoft will evaluate the demand for AMD's AI accelerator products and evaluate the feasibility of adopting the new product. Meta will use AMD's newly launched MI300X chip products in the data center. Oracle said it will use AMD's new chips in its cloud services.

Previously, the market expected AMD's MI300 series to ship about 300,000 to 400,000 units in 2024, with its largest customers being Microsoft and Google. If it were not for the limited TSMC CoWoS capacity shortage and Nvidia's already booked more than 40% of its production capacity, AMD's shipments are expected to be revised upward.

NVIDIA's stock price fell 1.5% after the news of AMD's MI300X accelerator was released. Nvidia’s stock price has skyrocketed this year, bringing its market value to more than $1 trillion, but the biggest question is how long it can monopolize the accelerator market. AMD saw its own opportunity: Large language models require a lot of computer memory, and this is where AMD sees its advantage.

In order to consolidate its market dominance, Nvidia is also developing its own next-generation chips. The H100 will be replaced in the first half of next year by the H200, which was launched last month and offers new high-speed memory that can perform inference twice as fast as the H100 on Llama2. In addition, Nvidia will launch a new processor architecture later this year.