Nvidia CEO Vera Rubin confirms AI platform is 'fully operational'

AI Video & Visuals


00:00 jensen fan

So we decided that we had to move forward.

00:05 jensen fan

No year is left behind, and cutting-edge calculations are performed every year.

00:13 jensen fan

And now we've been shipping the GB 200 for a year and a half. We are currently in full-scale production of the GB300.

00:24 jensen fan

And if Vera Rubin were to make it in time this year, it would be in production by now.

00:30 jensen fan

And today we can tell you that Vera Rubin is fully committed.

00:43 jensen fan

Want to see Vera Ruben? Okay, come on. Please play.

00:54 narrator

Vera Rubin arrives in time for the next frontier of AI.

01:01 narrator

This is the story of how we built it.

01:04 narrator

This architecture is a six-chip system designed to work as one and is the result of extreme collaborative design.

01:13 narrator

It starts with the custom-designed CPU Vera, which has twice the performance of the previous generation.

01:21 narrator

And Ruben GPU. Vera and Ruben were co-engineered from the ground up to enable faster, lower-latency, two-way, and consistent data sharing.

01:31 narrator

And 17,000 components come together on a Vera Rubin computing board.

01:41 narrator

A high-speed robot places the components with micron precision before a Vera CPU and two Ruben GPUs complete the assembly.

01:50 narrator

Deliver 100 petaflops of AI.

01:55 narrator

5 times more than its predecessor.

01:59 narrator

AI requires fast data.

02:03 narrator

Connect X9 provides 1.6 terabits/second of scale-out bandwidth to each GPU.

02:11 narrator

Bluefield 4 DPU offloads storage and security, so compute remains fully focused on AI.

02:20 narrator

The Vera Rubin compute tray has been completely redesigned to eliminate cables, hoses, and fans.

02:27 narrator

It features a Bluefield 4 DPU, 8 Connect X9 NICs, 2 Vera CPUs, and 4 Ruben GPUs.

02:35 narrator

Vera Ruben Computational building blocks for AI supercomputers.

02:40 narrator

Next up is the 6th generation NV-link switch.

02:47 narrator

Move more data than the global internet, connect 18 compute nodes, and scale up to 72 Ruben GPUs working as one.

02:59 narrator

Next up is Spectrum X Ethernet Photonics.

03:07 narrator

The world's first Ethernet switch that simultaneously packages 512 lanes and 200 Gigabit-capable optics scales out thousands of racks into an AI factory.

03:20 narrator

15,000 engineer years since design began. The first Vera Rubin NVL 72 rack is now online.

03:28 narrator

6 breakthrough chips, 18 compute trays, 9 NV Link switch trays, 220 trillion transistors,

03:37 narrator

It weighs almost 2 tons.

03:44 narrator

A giant leap to the next frontier of AI.

03:51 narrator

Ruben is here.

04:00 jensen fan

What do you all think?



Source link