News
AMD Launches MI325X AI Accelerator, Claims to Beat NVIDIA, Stocks Rise Then Fall
On Thursday, October 10th, AMD, which is in direct competition with NVIDIA in the field of artificial intelligence accelerators, launched new products, claiming that their AI chips will soon be shipped and hit the market, and will surpass NVIDIA in performance.
However, after erasing an early loss of 1.7% and briefly turning positive, AMD's stock price turned down again and fell by more than 4% to a one-week low. Intel's stock price fell by more than 1% after noon, while NVIDIA maintained a gain of more than 1%.
Some analysts believe that in the AI accelerator market, AMD is far behind the industry leader NVIDIA and is in second place, and this situation will not change in the short term. Due to the demand for AI chips far exceeding the supply, AMD's new products are unlikely to affect NVIDIA's data center revenue. It is expected that AMD's data center revenue this year will be $12.83 billion, while NVIDIA's similar revenue will be $110.36 billion.
What was released? Server CPUs, AI GPU accelerators, and several network chips.
Advertisement
At the "Advancing AI Event" data center event held in San Francisco, California, AMD launched the new fifth-generation Epyc data center CPU, code-named "Turin", and computers equipped with this processor will soon be on the market.
The new chip has up to 192 processor cores, priced at $14,813, using the Zen 5 architecture, which can provide up to a 37% increase in speed for advanced AI data processing, and is expected to further enhance AMD's leading position in the data center field.
AMD CEO Su姿丰 said that the performance of the Turin chip will surpass the latest Intel chips. Since its launch, AMD's EPYC (Xiaolong) CPU has occupied 34% of the server market share, far higher than the 2% in 2018.
At the same time, AMD launched the Instinct MI325X AI GPU accelerator, directly targeting NVIDIA's Blackwell. Su姿丰 said that the new AI chip's performance in inference exceeds NVIDIA's similar products, and the new type of memory used will speed up AI computing.
She also pointed out that the demand for AI accelerators has exceeded expectations. It is estimated that by 2027, the market size of AI accelerators will reach $400 billion, and will further rise to $500 billion in 2028.Some analyses refer to the MI325X as an "interim upgrade" in the cycle, still employing the same basic design and CDMA 3 GPU architecture as last year's MI300X launch, but the MI325X is "the first AI GPU to feature 256GB HBM3e memory." In the second half of 2025, AMD still plans to introduce the MI355X chip with 288GB HBM3e memory and the CDNA 4 architecture, with a 35-fold increase in inference performance.
It is reported that the MI325X will offer 1.8 times the capabilities of NVIDIA's H200 GPU and 1.3 times the bandwidth, with theoretical peak FP16 and FP8 computational performance that is also 1.3 times that of the H200 chip. The first batch of MI325X is expected to be mass-produced and shipped in the fourth quarter of this year, and will be more widely available in the first quarter of next year through server manufacturers such as Dell, Eviden, Gigabyte, HPE, Lenovo, and Supermicro.
Additionally, AMD has released several network chips that can accelerate data movement between chips and systems within data centers. For instance, the company has launched the Pensando Salina DPU and Pensando Pollara 400, the latter being the industry's first AI NIC to support the Ultra Ethernet Consortium (UEC). On the software front, AMD has introduced the latest ROCm 6.2 ecosystem, which offers an average performance improvement of 2.4 times, can increase AI workloads in the inference domain by up to 2.8 times, and raises training performance by an average of 2.4 times.
According to AMD's presentation, the MI 400 accelerator chip will be introduced in 2026, featuring a more advanced CDNA Next architecture. Some commentators have noted that AMD is attempting to break NVIDIA's dominance in the field of artificial intelligence accelerators, and like NVIDIA, AMD is committed to launching a new accelerator each year to accelerate the pace of innovation.
Forrest Norrod, a senior executive in charge of data center solutions, stated: "AMD will continue to fulfill its product roadmap, providing customers with the performance they need and the choices they want, to bring AI infrastructure to market faster on a large scale: 'With the new AMD Instinct accelerators, EPYC processors, and AMD Pensando networking engines, the ongoing development of the open software ecosystem, and the ability to integrate all of this into AI infrastructure optimization, AMD highlights the key expertise in building and deploying world-class AI solutions.'"
Lisa Su also mentioned that there are currently no plans to hire foundries other than TSMC, but the company is open to the prospect of cooperating with other foundries in the future, hoping for a diversified geographical distribution among its production partners.
How does Wall Street view this? The new product launch was disappointing, but the MI300X may ship 500,000 units this year.
Analysts generally recognize AMD's efforts to catch up with NVIDIA in the lucrative AI GPU market and to continue surpassing Intel in the server CPU domain. Moreover, in the development of AI accelerators, AMD's response has been far better than Intel's, and its market value has also surpassed that of Intel.However, from the stock performance perspective, investors seem to be quite disappointed with AMD's AI chip product launch event on Thursday. Lynx Equity Strategies believes this may be due to the MI325X's inability to significantly change the status quo and effectively challenge NVIDIA's dominant position:
"As investors anticipated clear details of the MI325 and its follow-up products, which led to a slight increase in AMD's stock price over the past week, we are skeptical about AMD's ability to narrow the gap with NVIDIA unless it can demonstrate a clear driving force for market share growth. Otherwise, the stock may retest its annual lows.
In recent months, AMD's customer appeal has been relatively weak. Although the company's management may publicize that its chip's benchmark test results are superior to NVIDIA's Hopper series, we believe that the MI300X still lags behind NVIDIA's H100 when running actual workloads. In addition to the excellent HBM memory density, the software stack, backplane network, and power consumption are equally important."
The aforementioned institution also mentioned that Meta, which supported the launch of the MI300X last year, may have reduced its commitment. Meta may replace NVIDIA's products with internally developed chips, and Amazon AWS may have the same intention. Therefore, there is doubt whether AMD can be considered a viable alternative to NVIDIA, and the MI series chips and their software stack may not be the "plug-and-play" solution that AMD claims.
However, John Vinh, an analyst at KeyBanc, believes that even if AMD is far behind NVIDIA, its development trajectory is obviously optimistic. It is expected to ship 500,000 Instinct MI300X AI accelerators this year, occupying an increasing share in the AI market:
"Thanks to its excellent performance and cost-effectiveness, AMD's Instinct MI300X AI GPU is a 'hot' product in this field. Praise from partners such as Lenovo also helps to bring in new business.
AMD is more competitive in the AI market than it was a few months ago. The company has implemented a thorough business transformation, such as unifying its data center and consumer GPU architecture, and expanding its AI business through cooperation with companies like Samsung, Microsoft, and Oracle.
Thursday's AI event will reshape AMD's AI business, as the company will launch a new architecture and showcase its future plans, which may intensify industry competition."
AMD will announce its third-quarter results in a few weeks. In July, it raised its guidance for this year's AI accelerator chip sales from the previous $4 billion to $4.5 billion, representing rapid growth. Previously, due to enterprises' enthusiasm for developing and deploying generative AI products, the demand for AMD's MI300X chips launched last year surged.
Before Thursday's event, Wall Street gave AMD 13 "buy" ratings and 3 "hold" ratings, with an average target price of $190.56, representing about an 11% upside. AMD has risen nearly 11% this year and set a new high in March.
Leave A Comments