News Intel details Guadi 3 at Vision 2024 — new AI accelerator sampling to partners now, volume production in Q3

[Admin]

Intel announced its new Gaudi 3 AI processors at its Vision 2024 event, claiming they offer up to 1.7X the training performance, 50% better inference, and 40% better efficiency than Nvidia’s market leading H100 processors, but for significantly less money.

Intel details Guadi 3 at Vision 2024 — new AI accelerator sampling to partners now, volume production in Q3 : Read more

 

[thisisaname]

So energy efficient may if the world used these instead of Nvidia the energy problems with AI would end over night! /s

 

[CmdrShepard]

They can say it's faster and cheaper than NVIDIA, but in the end TCO depends on power usage / efficiency and also on the amount of money you need to spend writing and maintaining custom software for their card.

That's why people will continue buying NVIDIA.

 

[gg83]

Intel using 200GBps network interface vs everyone using PCIe? Another bad decision? I am not seeing any real world stats. And it is unclear how the software platforms in use will drive Intel forward or reverse. Intel has been in reverse for a long time.

The software and ecosystem are key to AI Accelerator adoption right?

 

[JayNor]

"Intel also announced that it is developing an AI NIC ASIC for Ultra Ethernet Consortium-compliant networking and an AI NIC chiplet that will be used in its future XPU and Gaudi 3 processors..."

Gaudi 3 processors are already sampling. So are their NIC chiplets UEC-compliant?

 

[JayNor]

But Intel has a habit of muddying the waters. They took there CPU and designed cores that have different power properties which IMHO was insane and added unnecessary complexity. So now lets makes fast ethernet connections a 'thing' when in fact we can use the PCIe bus which is MUCH faster.

Intel also put PCIE5 X16 on each of the 8 Gaudi3 modules. In the Supermicro Gaudi3 server, there are 8 Gaudi3 modules, and the host CPU is reportedly a Granite Rapids-AP, with 136 PCIE5 lanes.

So, there is PCIE5 if you prefer that.

I think the Gaudi focus is on using their per NIC RoCE v2 DMA controllers, and controlling data transfers in a fully distributed manner. That's a lot of DMA going on without having to rely on a CPU.

 

[JayNor]

not clear why you consider pcie5 (32GT/sec) "PCIe bus which is MUCH faster"

The Gaudi 3 ethernet signals are being routed around as
48 x 112Gb/s PAM4 SerDes Links

let's see the math

https://www.intel.com/content/www/u...intel-gaudi-3-ai-accelerator-white-paper.html

 

[sjkpublic]

not clear why you consider pcie5 (32GT/sec) "PCIe bus which is MUCH faster"

The Gaudi 3 ethernet signals are being routed around as
48 x 112Gb/s PAM4 SerDes Links

let's see the math

https://www.intel.com/content/www/u...intel-gaudi-3-ai-accelerator-white-paper.html

Thank you for correcting me. My fault. I have deleted my posts as they are misinformation.