News Nvidia Reveals Hopper H100 GPU With 80 Billion Transistors

[Admin]

Nvidia unveiled it's new Hopper H100 GPU for datacenters, built on a custom TSMC 4N process and packing 80 billion transistors with 80GB of HBM3 memory.

Nvidia Reveals Hopper H100 GPU With 80 Billion Transistors : Read more

 

[peachpuff]

Do the datacenter guys have the same issues with Nvidia gpu stock as the regular folk?

 

[hotaru.hino]

Do the datacenter guys have the same issues with Nvidia gpu stock as the regular folk?

No, because they order directly from NVIDIA, who makes the final product. AIBs don't get to touch the "big" GPU products.

 

[renz496]

Do the datacenter guys have the same issues with Nvidia gpu stock as the regular folk?

don't think so. some of the professional client already orders 10 of thousands of them in advance. sometimes even before nvidia make their official product announcement. i heard one stories about how one company end up pre-ordering 15k of nvidia GK110 in mid 2012 when the first batch of GK110 chip not coming out until Q4 2012.

 

[JarredWaltonGPU]

No, because they order directly from NVIDIA, who makes the final product. AIBs don't get to touch the "big" GPU products.

Not to mention, these are in a completely different category of cost. Not even miners are willing to pay the prices that these GPUs command, and scalpers don't factory into the picture either. Buying a $700 GPU and reselling it for $1500 is one thing; buying a $400,000 DGX system and trying to flip it on eBay for even $500,000 isn't going to happen. :-D

 

[hotaru.hino]

Not to mention, these are in a completely different category of cost. Not even miners are willing to pay the prices that these GPUs command, and scalpers don't factory into the picture either. Buying a $700 GPU and reselling it for $1500 is one thing; buying a $400,000 DGX system and trying to flip it on eBay for even $500,000 isn't going to happen. :-D

I'm pretty sure the government also wants to know why you want to move more than 5 figures worth of money.

 

[drivinfast247]

I'm pretty sure the government also wants to know why you want to move more than 5 figures worth of money.

They want to know if you move $600 bucks!

 

[jp7189]

This article makes Hopper sound monolithic with talk of fully enabled die having low yields, etc. Has that been confirmed? I was expecting a shift to MCM on the datacenter lineup this round.

 

[spongiemaster]

Do the datacenter guys have the same issues with Nvidia gpu stock as the regular folk?

A regular ol' PCIe A100 costs about $20k ($15k off! It's like Black Friday in March!). Miners aren't touching these, and they are built by TSMC so Nvidia isn't deciding how to allocate their wafers, everything is going to these enterprise GPU's.

https://www.dell.com/en-us/shop/nvi...omer-install/apd/490-bhbk/graphic-video-cards

 

[spongiemaster]

This article makes Hopper sound monolithic with talk of fully enabled die having low yields, etc. Has that been confirmed? I was expecting a shift to MCM on the datacenter lineup this round.

Yes. The rumors pointed to MCM, but that appears to be for their upcoming CPU. GH100 is monolithic.

 

[JarredWaltonGPU]

Yes. The rumors pointed to MCM, but that appears to be for their upcoming CPU. GH100 is monolithic.

It's almost like people make up rumors when news is slow! Not mentioning any names... 😠😞

 

[hotaru.hino]

This article makes Hopper sound monolithic with talk of fully enabled die having low yields, etc. Has that been confirmed? I was expecting a shift to MCM on the datacenter lineup this round.

It probably won't be fully enabled. The A100 card has 20 or so of the SMs disabled and that's considered the highest end NVIDIA card.

Besides, considering the target market for this, I don't think NVIDIA's concerned about having yield issues unless it's really significant. You also have to consider Apple is selling an even larger die on as many, if not more, units. So there's likely a lot of confidence that TSMC can deliver on their 4nm.

Do the datacenter guys have the same issues with Nvidia gpu stock as the regular folk?

Also random update to this, I did find out that PNY and Gigabyte sell Quadro cards (PNY even sells the A100). However, judging by the lack of banding on the cards, I think NVIDIA simply wants them to second/third sources. It'd be like if NVIDIA allowed them to make FE cards, but they're making exact copies of what NVIDIA sells.

 

[renz496]

This article makes Hopper sound monolithic with talk of fully enabled die having low yields, etc. Has that been confirmed? I was expecting a shift to MCM on the datacenter lineup this round.

GH100 itself is monolithic. the way i see it AMD and nvidia both are using MCM but they take advantage of different aspect of it.

 

[jp7189]

Besides, considering the target market for this, I don't think NVIDIA's concerned about having yield issues unless it's really significant. You also have to consider Apple is selling an even larger die on as many, if not more, units. So there's likely a lot of confidence that TSMC can deliver on their 4nm

Yield issues aside I was thinking they'd hit a reticle limit before too long. MCM seems like the obvious next step - one that AMD has already taken with their CDNA2 datacenter GPUs. I guess TSMC 4 is able to cram enough transistors in to a single die that they didn't feel the need to change yet.

 

[jp7189]

My brain is off on a tangent this morning, but I was contemplating how they can funnel 600+ amps through a single die (that's assuming this die runs at 1 volt or less). A quick look at the EE charts says you need a wire with a cross section of ~760 sq mm to carry that much current without melting. I guess that's why so much engineering goes in to the cooling of dies like this.