Google Cloud A5X instances: Powering next-gen AI with NVIDIA Vera Rubin platform
Read the Portuguese translation →
Author: Google Cloud · Watch on YouTube
Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.
Original language: English
Original video: Google Cloud A5X instances: Powering next-gen AI with NVIDIA Vera Rubin platform · Channel: Google Cloud · 00:33 · Watch on YouTube
Topics: AI infrastructure and hardware
Summary
Section titled “Summary”Google Cloud A5X combines NVIDIA and Google technologies in a co-engineered system. Its sixth-generation NVLink connects 72 Rubin GPUs so a rack can operate as a unified accelerator, supported by Google’s fourth-generation Deschutes liquid cooling. Google’s Virgo Network and NVIDIA ConnectX-9 SuperNICs extend the system from a rack to as many as 80,000 GPUs in one data center. The video presents A5X on Google Cloud as NVIDIA Vera Rubin NVL 72 at Google scale.
Key points
Section titled “Key points”- Google Cloud A5X is described as a cohesive system co-engineered by Google and NVIDIA.
- Sixth-generation NVLink connects 72 Rubin GPUs.
- A rack can work as a unified accelerator.
- The racks use Google’s fourth-generation Deschutes liquid cooling.
- The cooling system is also used in Google’s accelerated infrastructure.
- Google’s Virgo Network and NVIDIA ConnectX-9 SuperNICs connect racks at data-center scale.
- The video cites up to 80,000 GPUs in a single data center.
- A5X is presented as NVIDIA Vera Rubin NVL 72 at Google scale.
Chapters
Section titled “Chapters”| Start | Chapter | Summary |
|---|---|---|
| 00:00 | A5X and rack-scale acceleration | Google introduces A5X as a co-engineered system and describes how NVLink connects 72 Rubin GPUs into a unified accelerator. |
| 00:15 | Liquid cooling | The video highlights Google’s fourth-generation Deschutes liquid cooling, also used in its accelerated infrastructure. |
| 00:33 | Scaling across a data center | Google’s Virgo Network and NVIDIA ConnectX-9 SuperNICs are presented as enabling connectivity across as many as 80,000 GPUs in one data center. |
Quotes
Section titled “Quotes”The best of NVIDIA and Google co-engineered into a single cohesive system. — 00:00
enabling a rack to work as a unified accelerator. — 00:00
A5X on Google Cloud. — 00:33
Concepts and entities
Section titled “Concepts and entities”Google Cloud A5X, NVIDIA, Google Cloud, Rubin GPU, NVLink, Deschutes, Virgo Network, ConnectX-9 SuperNICs, NVIDIA Vera Rubin NVL 72
Full transcript (original language)
Section titled “Full transcript (original language)”00:00 [MUSIC PLAYING] SPEAKER: Google Cloud A5X. The best of NVIDIA and Google co-engineered into a single cohesive system. Sixth generation NVLink connects 72 Rubin GPUs together– [MUSIC PLAYING] –enabling a rack to work as a unified accelerator. Every rack runs on Google’s fourth generation Deschutes liquid cooling, the same system behind more than a gigawatt of Google accelerated infrastructure.
00:33 Google’s Virgo Network and NVIDIA ConnectX-9 SuperNICs take it from one rack to 80,000 GPUs in a single data center. [? Cross ?] [? sites? ?] Nearly a million. NVIDIA Vera Rubin NVL 72 at Google scale. A5X on Google Cloud.