AI chips AI Showdown at Computex: AMD vs. Nvidia – and a bit of Intel

Related Vendors

NVIDIA GmbH

Intel Corp.

AMD Advanced Micro Devices GmbH

FAULHABER Drive Systems

Wuxi InfiMotion Technology Co., Ltd.

On June 3, 2024, AMD and Nvidia executives provided AI news in their keynote speeches for visitors to Computex. AMD is introducing the latest generation of AI chips and accelerating its schedule; Nvidia already has the successors to Blackwell and Grace in sight. And Intel also wants to join in.

The Computex in Taipei is one of the most important computer trade fairs in the world after CES and in Taiwan, the number one topic is AI. This is hardly surprising, as the development and supply of data centers with high-performance AI chips is proving to be a real gold mine of our time, as recently underscored by AI market leader Nvidia with its business figures for Q1 of fiscal year 2025.

Accordingly, the major manufacturers in the field of artificial intelligence chips took the spotlight at Computex and presented their medium-term plans to the listeners. And they are ambitious.

Nvidia announces successors to Grace and Blackwell

Just in March 2024, Nvidia CEO Jensen Huang introduced the Blackwell GPU, which is currently in production. Now, as part of the Nvidia keynote at Computex, he added even more with the successor to Blackwell, which is supposed to go into mass production in 2026, as well as a look at the 3-year roadmap for Nvidia.

By 2025 the Blackwell-Ultra-GPU, an even improved version of Blackwell, will arrive with twelve stacks of HBM3e for about 50 percent more memory capacity. 2026 will see Nvidia introduce the Rubin-GPU as a direct successor, then 2027 will have the Rubin-Ultra-GPU on the menu. In other words: A new version of AI technology is expected to be available every year; each one faster and more energy-efficient than the previous one.

While the name of the Rubin platform may be set, Huang has not really revealed any details about the specs yet. The graphics processor will, at least this is known, be equipped with eight stacks of HBM4; for the Ultra-variant, twelve stacks of HBM4 are expected to be used. HBM4 will also be used for the Vera CPU, which succeeds Grace. Nvidia's future superchip team will therefore bear the name Vera Rubin and replace Grace Hopper. In addition, in the same year, Nvidia plans to double the throughput of the NVLink switch to 3.6 TB/s (NVLink 6 vs. NVLink 5).

AMD counters with MI325X as a competitor to Hopper

Also, AMD CEO Dr. Lisa Su did not miss the opportunity to present at Computex with her own keynote. In this regard, she introduced the improved version of the AI accelerator MI300X – namely MI325X. The new variant is equipped with HBM3e and therefore offers more memory capacity (up to 288 GB) and a slightly increased memory transfer rate (from 5.3 to 6 TB per second).

The comparison to the current Hopper chip from Nvidia is obvious: The original version H100 still uses HBM3, the H200 version is only available with 141 GB of HBM3e memory. MI325X is expected to be available in the 4th quarter of 2024, so around the time Blackwell also hits the market. Blackwell has been announced with 192GB HBM3e with 8TB per second.

To adapt to the rapidly evolving AI market, AMD now also wants to follow Nvidia's lead and adopt a yearly product cycle for the Instinct series. In 2025, it starts with MI350 based on the CDNA4 architecture, which supports the AI data types FP4 and FP6 and also uses 288 GB of HBM3e memory. In addition, 3-nm process technology will be used, possibly coming from TSMC. At least AMD expressed satisfaction with the contract manufacturer. The MI400 series with a new CDNA architecture is planned for 2026.

And Intel?

While Nvidia dominates a large portion of the AI market and AMD tries to increase its own share, it is still up for debate what Intel is doing with regard to artificial intelligence. CEO Pat Gelsinger primarily talked about one thing during his keynote: money. Intel wants to make a significant impact on the wallet with the Gaudi-2 and Gaudi-3 AI accelerator kits and emphasizes that the costs for the combination of Xeon, Ultra-Core and Gaudi platform are up to two thirds lower than comparable competition offers.

The price tag for a standard rack with eight Gaudi-2 accelerators with a universal baseboard is 65,000 US dollars; a standard rack with eight Gaudi-3 accelerators with UBB comes in at 125,000 US dollars.

Intel also highlights the potential savings with the Xeon-6 processors. "With the acceleration of digital transformation, businesses are facing increasing pressure to renew their outdated data center systems in order to achieve cost savings, meet sustainability goals, maximize physical space in the racks, and create brand new digital capabilities throughout the enterprise. The entire Xeon-6 platform and processor family was specifically designed for these challenges and includes both E-Core (Efficiency) and P-Core (Performance) SKUs to handle a wide range of use cases and workloads, from AI and other high-performance computing requirements to scalable cloud-native applications. Both E-Cores and P-Cores are based on a compatible architecture with a common software stack and an open ecosystem of hardware and software providers," it says.

With the Xeon-6 series, Intel aims to score not only in terms of performance, but also energy efficiency and space requirements. The efficiency Xeon-6 processors with the code name Sierra Forest are available now and the performance Xeon-6 processors with the code name Granite Rapids are expected to be available in the next quarter, i.e. Q3 2024. (sb)