On April 14th, Jim Keller, not only renowned for his chip design prowess but also as a staunch advocate for open technology, expressed his disdain for closed technologies once again. Naturally, NVIDIA found itself in his crosshairs.
Recently, Keller proposed that NVIDIA's latest Blackwell GPU should not use the proprietary NVLink standard protocol for multi-chip and network interconnects. Instead, he suggests adopting the open Ethernet standard, potentially saving NVIDIA billions of dollars.
He further argues that NVIDIA should not use its proprietary InfiniBand solution in data center networks, advocating for a shift to Ethernet.
While NVIDIA's InfiniBand network boasts low latency and high bandwidth, reaching up to 200GbE, Ethernet can achieve speeds of 400GbE or even 800GbE.
Major players like AMD, Broadcom, Intel, Meta, Microsoft, and Oracle are collaborating on the development of the next-generation Ultra Ethernet, offering even higher throughput and better suitability for AI and HPC applications.
Additionally, Keller has long been critical of NVIDIA's closed ecosystem around CUDA, likening it to a swamp rather than a moat.
Arm, Intel, Qualcomm, Samsung, among others, have formed the Unified Accelerator Foundation (UXF) with the aim of supplanting NVIDIA's solutions.
However, for Jensen Huang, whether it's NVLink or CUDA, both represent the fruits of his company's multi-billion-dollar investments over the years and are crucial tools for safeguarding its interests. Thus, abandoning them is not a decision to be made lightly.
