Intel Ponte Vecchio Xe-HPC GPU holds 100 billion transistors

Published 2021-03-24 17:02 by Hilbert Hagedoorn

Click here to post a comment for Intel Ponte Vecchio Xe-HPC GPU holds 100 billion transistors on our message forum

#5898980

cucaulay malkin

2021-03-24 17:22

😱 ga100 has 54b

#5898993

Fox2232

2021-03-24 17:58

47 tiles, now, that's impressive.

#5899006

Teo_

2021-03-24 18:17

INB4 Intel HBM wit many tiles on consumer GPUs. (JK)

#5899027

labidas

2021-03-24 19:00

but can it do anything?

#5899030

fantaskarsef

2021-03-24 19:11

labidas:

but can it run Crysis Remastered or Cyberpunk with RT native?

Fixed that for you.

#5899032

Truder

2021-03-24 19:13

labidas:

but will it blend?

FTFY

#5899062

wavetrex

2021-03-24 20:10

labidas:

but can it do anything?

It can mine coin for sure.

#5899067

Gomez Addams

2021-03-24 20:16

"Yeah, that does put a GPU named like Big NAVI to shame really." No, it doesn't. Not in the slightest. It's a GPU, FFS. Where is the shame? I don't see it. The fact is AMD has extensive experience in multi-chip module designs. If they wanted to build a GPU with an MCM design don't you think they would?

#5899068

CPC_RedDawn

2021-03-24 20:17

1exaFLOP 😱 Can it run.... YES.

#5899077

CPC_RedDawn

2021-03-24 20:24

Gomez Addams:

The fact is AMD has extensive experience in multi-chip module designs. If they wanted to build a GPU with an MCM design don't you think they would?

RDNA3, or more than likely RDNA4, will be MCM design. Will be interesting if they copy what they did in the CPU space and have an I/O die for a huge amount of L3 cache (infinity cache) to feed all the other GPU chiplets. Or if they will specify each chiplet to do different things? Like have 4 GPU chiplets, 2 for graphics compute, 1 for ray traced effects, and 1 for some form of upscaling tech? Then a 5th chiplet as an I/O die with a few GB of cache? Now I think about it maybe each chiplet would have its own smaller amount of infinity cache (L3) then the I/O die could have a large pool of shared cache (L4?) which would pull data from an even larger pool of VRAM (GDDR6/7). Or will HBM2/3 finally be viable by then?

#5899123

Kaarme

2021-03-24 23:10

Luckily supercomputer builders make deals directly with Intel (or a company like HP that speaks for Intel in such a case), so miners can't snatch all of those super GPUs.

#5899124

cucaulay malkin

2021-03-24 23:14

Kaarme:

Luckily supercomputer builders make deals directly with Intel (or a company like HP that speaks for Intel in such a case), so miners can't snatch all of those super GPUs.

or intel made them to mine on their own

#5899127