Nvidia CEO Jensen Huang demonstrated the first version of Spectrum-X, the Spectrum-4 chip, with a hundred billion transistors in a 90 mm x 90 mm die. Nvidia
Nvidia CEO Jensen Huang, giving the opening keynote at the Computex computer technology conference, on Monday in Taipei, Taiwan, unveiled a host of new products, including an assortment of converters new ethernet dedicated to moving large volumes of data for artificial intelligence tasks.
“How do we introduce a new, backwards-compatible ethernet of everything, to turn every data center into a composite AI data center?” Huang set out in his keynote speech. “We brought the capabilities of high-performance computing to the ethernet market for the first time,” said Huang.
According to Nvidia, Spectrum-X, as the ethernet family is known, is “the world’s first high-performance ethernet for AI”. Gilad Shainer, senior vice president of networking, says a key feature of the technology is that it “doesn’t drop packets”.
Nvidia says the first version of Spectrum-X is Spectrum-4, which it calls “the world’s first 51Tb/s Ethernet switch built specifically for AI networking.” This switch works in conjunction with Nvidia’s BlueField data processor, or DPU, data fetching and queuing chips, and an Nvidia fiber optic transceiver. The company says the switch can route 128 400 gigabit ethernet ports or 64 800 gig ports end-to-end.
Huang held up the silver Spectrum-4 ethernet converter chip on stage, noting that it was “huge”, consisting of one hundred billion transistors on a 90 mm x 90 mm die built using “4N” process technology. ” by Taiwan Semiconductor Manufacturing Company. Huang said the unit runs at 500 watts.
“We brought the capabilities of high-performance computing to the ethernet market for the first time,” said Huang. Nvidia
Spectrum-4, the first in the Spectrum-X family of chips, is a purpose-built new ethernet to provide lossless packet transmission for AI workloads. Nvidia
Nvidia’s chip, the switch that houses it, has the potential to change the ethernet market. The majority of switch silicon is provided by chip manufacturer Broadcom. These switches are sold to network equipment manufacturers Cisco Systems, Arista Networks, Extreme Networks, Juniper Networks, and others. Those companies have been expanding their equipment to better handle AI traffic.
The Spectrum-X series is built to address data center fragmentation into two forms. One is what he calls “AI factories,” hundreds of millions of dollars worth of facilities for the most powerful GPUs, based on Nvidia’s NVLink and Infiniband, used for AI training, serving a small number of people. very large workload.
The other data center facility is the AI cloud, which is multi-tenant, ethernet-based, and handles hundreds of hundreds of workloads for customers concurrently and is focused on things like providing predictability to consumers. using AI, will be served by Spectrum-X.
VP Shainer says Spectrum-X can “disperse traffic across the network in the best way,” said Shainer, using “a new mechanism for congestion control,” which helps prevent the superposition of packets that can may occur in the caching of network routers.
“We use advanced telemetry to understand network-wide latency to identify hotspots before they cause anything, to keep the network free of congestion.”
Nvidia said in a prepared comment that “the world’s leading hyperscale companies are adopting NVIDIA Spectrum-X, including industry-leading cloud innovators.”
Nvidia is building a test computer in its Israeli offices, called Israel-1, a “next generation AI supercomputer”, using Dell PowerEdge XE9680 servers including H100 GPUs running data on Spectrum-4 switches.
All news at Computex available in the Nvidia newsroom.
In addition to the new ethernet technology, Huang’s keynote also introduced a new model in the company’s “DGX” series of AI computers, the DGX GH200, which the company touts as “a kind of supercomputer” new big memory AI computer for giant generation AI models.”
Creative AI refers to programs that generate more than just scores, sometimes text, sometimes images, other artifacts, such as OpenAI’s ChatGPT bot.
The GH200 is the first system to ship with what the company calls a “super chip”, the Grace Hopper board, containing on a single board the GPU Hopper and the Grace CPU, an ARM instruction set-based CPU meant to compete with Intel x86 CPUs and Advanced Micro Devices.
The company says Nvidia’s Grace Hopper “super chip,” a board containing the Grace CPU, on the left, and the Hopper GPU, is now in full production. Nvidia
Huang said the first version of the Grace Hopper, the GH200, is “in full production”. Nvidia said in a press release that “global supercomputing and supercomputing centers in Europe and the United States are among a number of customers that will have access to the systems offered by the GH200. ”
The DGX GH200 combines 256 of the superchips, Nvidia says, to achieve 1 aggregate exaflop — ten times the power of 18, or a billion, billion floating point operations per second — using 144 terabytes of memory. shared. According to Nvidia, the computer is 500 times faster than the original DGX A100, which was released in 2020.
The keynote also revealed MGX, a reference architecture for system manufacturers to quickly and cost-effectively build over 100 server variants. Nvidia says the first partners to use this specification are ASRock Rack, ASUS, GIGABYTE, Pegatron, QCT, and Supermicro, with QCT and Supermicro being the first to bring the system to market in August. know.
The whole speech can be considered as one playback from Nvidia website.
MGX is a reference architecture for computer system manufacturers to quickly and cost-effectively build over 100 server variants using Nvidia chips. Nvidia says QCT and Supermicro will be the first to market the systems in August. Nvidia
