NVIDIA Unveils Its First Data Center CPU

[Tsing]

Image: NVIDIA



During today’s GTC 2021 keynote, NVIDIA CEO Jensen Huang revealed his company’s first data center CPU: the NVIDIA Grace. NVIDIA Grace is an Arm-based processor that is said to deliver 10x the performance of today’s fastest servers for complex AI and other high-performance computing workloads when tightly coupled with NVIDIA GPUs. The CPU, which was named after computer scientist Grace Hopper, also boasts great efficiency thanks to its energy-efficient Arm-based cores and innovative low-power memory subsystem.



Meet NVIDIA Grace – the basic building block of the modern data center. This powerful @Arm-based #CPU allows us to accelerate the largest HPC and #AI...

Continue reading...

 

[Stoly]

This has been Jensen's dream for as long as I can remember, had a couple of bad starts, but this seems to be the real deal.

 

[Grimlakin]

I'm curious about how it is actually used... How does it accelerate AI type operations... I understand to some degree how a traditional CPU works... but this is something wholly new.

Unless of course it's just a new name for the same old thing. In that case.. cool glad it's here and advances the tech but also... meh... wish it was like quantum or something to that effect of processing.

 

[Brian_B]

I'm curious about how it is actually used... How does it accelerate AI type operations... I understand to some degree how a traditional CPU works... but this is something wholly new.

Unless of course it's just a new name for the same old thing. In that case.. cool glad it's here and advances the tech but also... meh... wish it was like quantum or something to that effect of processing.

My bet - ARM cpu with tensor cores

 

[Grimlakin]

My bet - ARM cpu with tensor cores

Yea you're probably right.

 

[Stoly]

My bet - ARM cpu with tensor cores

I think its more like the CPU to GPU channels which would allow the cpu to access the gpus via a high bandwidth bus. Rather than integrating tensor cores to a cpu. I mean why would you want that when you are tied to an AmpereNext GPU anyway.

 

[Brian_B]

I think its more like the CPU to GPU channels which would allow the cpu to access the gpus via a high bandwidth bus. Rather than integrating tensor cores to a cpu. I mean why would you want that when you are tied to an AmpereNext GPU anyway.

Well. so you can DLSS your data center, obviously. What other possible use could there be for Tensor cores anyway?