NVIDIA "Hopper" Might Have Huge 1000 mm² Die, Monolithic Design

[Raevenlord]

Renowned hardware leaker kopike7kimi on Twitter revealed some purported details on NVIDIA's next-generation architecture for HPC (High Performance Computing), Hopper. According to the leaker, Hopper is still sporting a classic monolithic die design despite previous rumors, and it appears that NVIDIA's performance targets have led to the creation of a monstrous, ~1000 mm² die package for the GH100 chip, which usually maxes out the complexity and performance that can be achieved on a particular manufacturing process. This is despite the fact that Hopper is also rumored to be manufactured under TSMC's 5 nm technology, thus achieving higher transistor density and power efficiency compared to the 8 nm Samsung process that NVIDIA is currently contracting. At the very least, it means that the final die will be bigger than the already enormous 826 mm² of NVIDIA's GA100.

If this is indeed the case and NVIDIA isn't deploying a MCM (Multi-Chip Module) design on Hopper, which is designed for a market with increased profit margins, it likely means that less profitable consumer-oriented products from NVIDIA won't be featuring the technology either. MCM designs also make more sense in NVIDIA's HPC products, as they would enable higher theoretical performance when scaling - exactly what that market demands. Of course, NVIDIA could be looking to develop an MCM version of the GH100 still; but if that were to happen, the company could be looking to pair two of these chips together as another HPC product (rumored GH-102). ~2,000 mm² in a single GPU package, paired with increased density and architectural improvements might actually be what NVIDIA requires to achieve the 3x performance jump from the Ampere-based A100 the company is reportedly targeting.



View at TechPowerUp Main Site

 

[oobymach]

1000mm is 100cm which is 1meter or around 40" for you Americans, that is a big honking gpu die. I think we took a leap backward here,

* imagines trying to fit that motherboard sized block along with a pcb and rams and such required to run it in a cpu case.

 

[HTC]

1000mm is 100cm which is 1meter or around 40" for you Americans, that is a big honking gpu die. I think we took a leap backward here, * imagines trying to fit that motherboard sized block along with a pcb and rams and such required to run it in a cpu case.

It's actually just 31.63mm squared ...

I remember that video: that is HILARIOUS.

 

[Anymal]

31,6227

1.245 inch

 

[Space Lynx]

TIE FIRST ROUND OF SALES TO STEAM ACCOUNTS WITH AT LEAST 1K+ GAMES AND 10 YEARS OLD IN LENGTH+

then lower the requirements for each round of graphics card sales after that.


THIS IS what nvidia/amd would do if they cared about gamers, yes i know not everyone plays on steam, but at this is is a starting point for the first say 7 rounds before you open it up to everyone.

but w.e

 

[oobymach]

31,6227

1.245 inch

Yeah it's less impressive after realizing my 2080 has a 545mm square die and it's not that big.

 

[R-T-B]

TIE FIRST ROUND OF SALES TO STEAM ACCOUNTS WITH AT LEAST 1K+ GAMES AND 10 YEARS OLD IN LENGTH+

then lower the requirements for each round of graphics card sales after that.


THIS IS what nvidia/amd would do if they cared about gamers, yes i know not everyone plays on steam, but at this is is a starting point for the first say 7 rounds before you open it up to everyone.

but w.e

This part isn't even intended for gamers, and NVIDIA has cared about HPC more than gamers for longer than crypto has even existed... Why? Because they pay more, and more consistently.

 

[Space Lynx]

This part isn't even intended for gamers, and NVIDIA has cared about HPC more than gamers for longer than crypto has even existed... Why? Because they pay more, and more consistently.

AMD to the moon then

fuck those leather jackets anyway, i like my regular jacket.

 

[phanbuey]

sounds expensive.

 

[R0H1T]

And cost $100k

 

[Lycanwolfen]

When video cards cost more than the computer itself then might as well forget it. Still running 4k on everything with my twin 1070ti's in SLI working fine

 

[z1n0x]

1000mm² on 5nm. Yeah, this can't be good for yields and price. I hope it's fake news/rumors, and they have a working MCM. That's the only way to keep the ever increasing costs under control. Otherwise further price increases or give up on the 65% margins (and lose the Wall Street darling statue, which of course, it won't happen). The same applies for gaming gpus.

 

[Fluffmeister]

For a HPC only chip this isn't going to be too much of a surprise.

 

[Legacy-ZA]

Interested to see Lovelace

 

[Patriot]

1000mm² on 5nm. Yeah, this can't be good for yields and price. I hope it's fake news/rumors, and they have a working MCM. That's the only way to keep the ever increasing costs under control. Otherwise further price increases or give up on the 65% margins (and lose the Wall Street darling statue, which of course, it won't happen). The same applies for gaming gpus.

Hopper was supposed to be mcm from long ago

 

[TheoneandonlyMrK]

Not at all what I expected , didn't Huang show off a 9/12 chip MCM hopper prototype a while ago.

With much much smaller chip's.

 

[oobymach]

When video cards cost more than the computer itself then might as well forget it. Still running 4k on everything with my twin 1070ti's in SLI working fine

I had 1070's in sli until the sli went poop after about 4 years and 2 builds, the downside to sli is latency and games that actually support it are few and far between.

 

[Anymal]

Probably estimation for inferior sammys process.

 

[Wirko]

TIE FIRST ROUND OF SALES TO STEAM ACCOUNTS WITH AT LEAST 1K+ GAMES AND 10 YEARS OLD IN LENGTH+

If anything, sales will be tied to NASDAQ accounts with at least 1K+ NVDAs bought each year, 10 years in length+.

 

[Space Lynx]

meh, i didn't realize this wasn't for gamers, so it was a mistake post anyway

 

[mechtech]

build it on 90nm, that will be a nice chunk of silicon then

 

[WhoDecidedThat]

I see two ways this 1000 mm2 isn't bullshit:

1. They are using tech similar to Cerebras' Wafer Scale Engine to link multiple dies together. (See image below)
2. They are quoting the size of the silicon interposer they are using for connecting HBM to the processing die.

 

[bug]

There's also the possibility, just like Ampere, that this is a HPC-only architecture. I mean, even if they cut that in half, it's still going to make for a damn expensive GPU. Then again, "damn expensive" is where pricing for video cards starts these days

 

[windwhirl]

I see two ways this 1000 mm2 isn't bullshit:

I don't see why it would be bullshit. It's not unfeasible to produce such a large chip. Just ridiculously expensive if the yield rate isn't good enough.

And expensive isn't really a problem if this is indeed targeted at HPC markets, which can and will pay for it if the performance is good enough.

 

[RH92]

I see two ways this 1000 mm2 isn't bullshit:

1. They are using tech similar to Cerebras' Wafer Scale Engine to link multiple dies together. (See image below)
2. They are quoting the size of the silicon interposer they are using for connecting HBM to the processing die.

View attachment 234491

Actully neither of those is a necessity , they can print beyond reticle limit by making multiple passes it's just that yealds will be poor but this is HPC so money is not really an issue as long as perf is there .

There's also the possibility, just like Ampere, that this is a HPC-only architecture.

You mean Volta but yeah many peoples completely forget this side of the equation and this is a certitude not only a possibility . Nvidia has been segmenting their gaming and HPC architectures and we know Hopper is going to be HPC so we have no idea what their gaming architecture will look like .