Banner reading XPU dot pub
Enable images in email to see Jensen Huang signing a woman's dress

Nvidia and Intel Talk Data-Center-AI Economics at Computex 2024

Ramenez la coupe à la maison

Some of the biggest product news came from AMD at Computex 2024, but Nvidia and Intel also delivered keynotes. Nvidia’s CEO Jensen Huang talked mostly about data-center AI, where the company is the undisputed leader. Testifying to Nvidia’s gravitational pull, he disclosed AMD and Intel are supporting Nvidia’s MGX modular reference design with compatible CPU modules.

Putting those companies’ AI businesses on notice, he revealed plans to release new data-center AI accelerators (GPUs/NPUs) and Spectrum-X networking gear every year. To this end, he revealed (i.e., confirmed the name of) Blackwell’s successor. Called Rubin, it will coincide with the forthcoming Arm-based Vera processor that will replace Grace. In the meantime, Nvidia must ship Blackwell as well as the Blackwell Ultra upgraded-memory midlife kicker.

For those worried about the power and cost of Nvidia’s data-center chips, Huang highlighted how Nvidia’s GPUs have raised throughput (TOPS) 25× faster than they have increased power. The degree to which real-world applications can achieve new GPUs’ theoretical performance gains is debatable, but no one questions whether data centers’ energy bills will go up. Likewise, peak TOPS per Nvidia chip may be climbing but so too is cost.

Intel Breaks Convention to Discloses Gaudi Prices

Seizing on the latter point, Intel CEO Pat Gelsinger disclosed prices for the Intel Gaudi 3 and Gaudi 2 systems. Unusual as a discussion of price is, especially for infrastructure equipment, it’s the main weakness of Nvidia-based hardware. Competitors’ claims of marginally superior throughput have done little to blunt Nvidia’s ascendance, but acquisition cost could emerge as a decisive issue and is far less nebulous than product performance.

Gelsinger also formally announced the all-E-core Xeon 6, better known as Sierra Forest—a processor Intel has talked about for years. Intel touts how one Sierra system can replace multiple servers based on older Xeons. However, the same is true for AMD’s Epyc 9700 Bergamo processors released last year. Without naming names, both x86 suppliers offer these dense processors to fend off in-house processors based on Arm Neoverse and challengers like Ampere.

Nvidia Tells Its AI-PC Story

We previously reported on Intel Lunar Lake, disclosed in Gelsinger’s keynote. Not to be shut out of AI-PC mania, Nvidia’s keynote also discussed how PCs integrating its discrete RTX GPUs can handle AI processing. Their throughput will dwarf that of NPUs in the all-in-one chips from AMD, Intel, and Qualcomm, but their added cost and power won’t expand Nvidia’s serviceable market beyond gaming laptops and desktops that benefit from the GPU’s graphics performance.

Bottom Line

In summary, Computex 2024 was a big show for chipmakers. AMD refreshed broad swaths of its product line, Nvidia enraptured audiences with its data-center AI success, and Intel sought to stay relevant in the data-center and PC markets.




error: Unable to select