Jim Keller Suggests Nvidia Could Have Saved Billions with Ethernet in Blackwell GPUs

Jim Keller, a revered CPU architect, and CEO of Tenstorrent, argued that network connections such as Ethernet would have been a ‘logical choice’ for the Nvidia Volta-based GPUs targeted at AI and HPC applications since these network connections are directly replacing the proprietary NVLink connections. The Semiconductor industry technical expertise Jim Keller called on Twitter recently for support of the Ethernet chip-to-chip connectivity by Nvidia’s Blackwell-based GB200 GPUs. Keller believes that this method had been well capable of saving Nvidia’s valuable hardware resources by substantial amounts and also eventually made the software migration a relatively simple procedure.

Ethernet over proprietary protocols

Nvidia’s new AI processor with integrated Gb200 GPU has a chip-to-chip communication protocol which is developed by the company: NVLink. In the quote, it is noted that for some reason open standards instead of other proprietary ones (NVLink) would have had lower costs and could have simplified the portability of the software.

Nvidia’s strategic move of using proprietary protocols like NVLink and InfiniBand for communication opens up a whole set of issues for the market holistically, in particular with the swapping of software. These solutions have likewise improved performance and user experience on Nvidia platforms, but it has made things complicated and hindered interoperability and compatibility with other platforms.

Advancements in ethernet technology

Ethernet, the once reserved and consensual protocol, is now a widespread success in areas where both hardware and software are employed, creating an appealing option to the owned protocols. Introduced innovations in Ethernet, namely, the high-speed Ultra Ethernet interconnection technology include improved performance and suitability for the specialized AI/HPC tasks.

Nvidia monopolizes data center production using the CUDA software platform and retains competitors like Unified Accelerator Foundation (UXL) that are supported in the industry. PCIe with CXL and Ultra Ethernet being two open-standards foundations contribute significantly to shifting the competitive scenario and therefore, Nvidia must think about it’s proprietary strategies.

Navigating the open-standard vs. proprietary dilemma

As a result of low-price, efficient, and easy-to-use software on port Nvidia can utilize the existent All efforts are to satisfy the data needs of a data center. It is still very early to draw commend. However, the open-standard technology’s growing use case indicates that Ethernet and some alternatives may replace Bost’s specific solutions and lead to the changing of the current strategies of the company.

As shown in Jim Keller’s recommendation to integrate Ethernet into Nvidia’s Blackwell GPUs, this has prompted the perennial quandary of members of the semiconductor industry: Will they go for the open-standard technologies or graphic card manufacturers ‘ proprietary brands? Even so, Nvidia remains steadfast in its commitment to inventing its unique solutions for the present supply and demand while the advent of Ethernet and the growing consortium initiatives might just spell some shifts in the competitive arena. Since the industry is dynamic, Nvidia and its rivals shall reciprocally be involved in the issue of private brands and bound technologies.

This article originally appeared in Tom’s Hardware.