Sunday, June 2, 2024

NVIDIA scales its Spectrum-X Ethernet networking platform

At COMPUTEX 2024 in Taiwan, NVIDIA's Jensen Huang unveiled a roadmap for new semiconductors that will arrive on a one-year rhythm. 

The Rubin platform will succeed the upcoming Blackwell platform, featuring new GPUs, a new Arm-based CPU — Vera — and advanced networking with NVLink 6, CX9 SuperNIC and the X1600 converged InfiniBand/Ethernet switch.

“Our company has a one-year rhythm. Our basic philosophy is very simple: build the entire data center scale, disaggregate and sell to you parts on a one-year rhythm, and push everything to technology limits,” Huang explained.

Spectrum-X features the NVIDIA Spectrum SN5600 Ethernet switch and the NVIDIA BlueField -3 SuperNIC. The platform leverages adaptive routing and congestion control for maximum bandwidth and noise isolation. It enables advanced cloud multi-tenancy, GPU compute elasticity and zero-trust security.

NVIDIA will launch new Spectrum-X products every year, delivering increased bandwidth and ports and enhanced software feature sets and programmability.  NVIDIA claims Spectrum-X accelerates generative AI network performance by 1.6x over traditional Ethernet fabrics. Data center networks based Spectrum-X switches are currently designed connecting for tens of thousands. With 1.6 Tbps interfaces on the horizon this will soon increase to millions of GPUs.

The next generation NVLink Switch chip will feature:

  • 50B transistors and use TSMC's 4NP 
  • 72-ports of 400G SerDes
  • 4 NVLinks at 1.8 TB/s
  • 7.2 TB/s full-duplex bandwidth

Another key networking aim for NVIDIA is bringing the capabilities of Infiniband to Ethernet for hyperscale data centers, including

  • Network-level RDMA
  • Congestion control using the switch telemetry
  • Adaptive routing using the Bluefield NICs
  • Noise isolation between training models





NVIDIA cited rapid adoption of its Spectrum-X Ethernet networking platform. 

CoreWeave, GMO Internet Group, Lambda, Scaleway, STPX Global and Yotta are among the first AI cloud service providers embracing NVIDIA Spectrum-X. Additionally, several NVIDIA partners have announced Spectrum-based products, including ASRock Rack, ASUS, GIGABYTE, Ingrasys, Inventec, Pegatron, QCT, Wistron and Wiwynn, which join Dell Technologies, Hewlett Packard Enterprise, Lenovo and Supermicro in incorporating the platform into their offerings.


“Rapid advancements in groundbreaking technologies like generative AI underscore the necessity for every business to prioritize networking innovation to gain a competitive edge,” said Gilad Shainer, senior vice president of networking at NVIDIA. “NVIDIA Spectrum-X revolutionizes Ethernet networking to let businesses fully harness the power of their AI infrastructures to transform their operations and their industries.”

https://nvidianews.nvidia.com/news/nvidia-supercharges-ethernet-networking-for-generative-ai

https://www.youtube.com/watch?v=pKXDVsWZmUU&t=4772s