The Nvidia Hopper thread

[DarrenM343]

Pre-orders should start now, paid in full, for next gen. Might have a hope in hell of landing one around release time then and save F5 keys all over the world.

 

[TNA]

Pre-orders should start now, paid in full, for next gen. Might have a hope in hell of landing one around release time then and save F5 keys all over the world.

Haha

 

[Wrinkly]

If RTX4000 is Hopper and on 7nm then yeah I'd expect the 4070 would smash the 3090

Not sure Nvidia will stick with 7nm that long.

 

[Quartz]

I do hope that this isn't an April Fool's joke.

https://wccftech.com/nvidia-h100-ho...-dies-43008-cuda-cores-and-48-gb-hbm4-memory/

Perhaps a launch late this year?

The idea of using Steam and Epic accounts will go a long way to combatting scalpers and miners.

 

[Random Guy Esq.]

I couldn't have my yearly cup of coffee with Jensen this year due to the coronavirus, but we had a meeting on Zoom and he agreed to share some high-resolution slides of NVIDIA's upcoming plans.

Yeah, it's obviously legit.

 

[Quartz]

Yeah, it's obviously legit.

Shhh!!

 

[humbug]

2 year steam account as prof before purchase to get at MSRP would be a good idea.

 

[varkanoid]

I do hope that this isn't an April Fool's joke.

If you read the whole article especially the last two paragraphs you will get your truth

 

[Quartz]

If you read the whole article especially the last two paragraphs you will get your truth

Shhh!!

 

[4K8KW10]

Perhaps a launch late this year?

100.1B transistors when GA100 on its own has 54.2B and is a single die on N7 process?!

It is most likely a Fool's day article because the two dies on the photo don't look like 500ish sq. mm dies.

It should be possible on a N5 process, should cost north of $10000 or even $20000, but if it is on N7, why would they glue together 800+ sq. mm chips?!!

 

[varkanoid]

It should be possible on a N5 process, should cost north of $10000 or even $20000, but if it is on N7, why would they glue together 800+ sq. mm chips?!!

So they can more tomato sauce on their chips.

 

[4K8KW10]

So they can more tomato sauce on their chips.

Mmm

 

[LoadsaMoney]

 

[queamin]

Loads you did a video again at last.

 

[K1LaW8tt5]

Rumoured the flagship card will have 9gb vram

 

[HRL]

Thank god smellyvision isn’t a thing.

So Hopper’s not next, Lovelace then Hopper.

 

[willhub]

All that YouTube stuff is pure hype rubbish made up for clikz

 

[TNA]

Loads you did a video again at last.

 

[Grim5]

So RTX Hopper specs according to Kope and bare in mind leaks are flaky at the best of time, but Kope trying to leak specs for a GPU that is 3 years away is super flaky.

RTX5000 (Hopper) = MCM architecture, dual chiplet design with 144SM per chiplet

The layout of the cores within the SM isn't known, however if we assume the SM's carried the same configuration as Ampere then this would be the specs of the RTX5090: 144SM per chiplet, dual chiplet = 288SM per RTX5090, combined core counts as follows:

Cuda Cores: 36864 (RTX3090 = 10496)
ROPs: 392 units (RTX3090 = 112)
TMUs: 1152 units (RTX3090 = 328)
Tensor Cores: 1152 (RTX3090 = 328)
Ray Tracing Cores: 288 (RTX3090 = 82)
Transistors: 100 Billion (RTX3090 = 28 billion)

FP32 Performance: 124 Teraflops (RTX3090 = 36 Teraflops). This makes the assumption that the RTX5090 and RTX3090 operate at the same 1695mhz stock clock speed.

Also worth remembering that the latest leaks from the AMD camp shows that MCM architectures don't have a 1:1 gaming performance improvements, RX7000's MCM architecture appears to have a 0.5:1 scaling.

If that same scaling applied to Hopper, then the RTX5090 would not be 3.4 times times faster than the RTX3090 as the FP32 numbers would indicate but instead it would be 1.7 times faster than the RTX3090.

 

[LoadsaMoney]

Loads you did a video again at last.