News Intels neue Grafikkarten: Xe²-HPG ("Battlemage") mit 3.072 Shader-Einheiten und 12 GiByte Grafikspeicher gesichtet

[DarkWing13]

Ich schätze bei Intel stellt sich inzwischen die Frage, ob sie zumindest die Consumer-Grafikkarten ganz unter den Tisch fallen lassen (klar BM wird noch kommen, wurde ja auch genug Geld und Zeit reingesteckt), und mehr den Weg Richtung KI-CPUs in Verbindung mit KI-GPUs gehen sollen, ala Nvidia.
Dort sind die großen Gewinne zu holen, wie Nvidia zeigt.

Passend dazu:

Für Super-KI-Supercomputer: Cerebras' Riesen-CPU mit 4 Billionen Transistoren

Dank Fertigung mit 5- statt 7-Nanometer-Technik wächst die Rechenleistung der Wafer Scale Engine WSE-3 auf bis zu 126 PFlops für KI-Datentypen.

www.heise.de

 

[gaav87]

Regarding the article, where do the 3072 shaders come from?
SiSoftware gives:
192CU's to 1728SM's with 8MB L2 and 12GB VRAM, or
160CU's to 1440SM's with 8MB L2 and also 12GB VRAM.
Or am I confusing something here?

Anhang anzeigen 1456330

Otherwise it could fit well with the G21.
The 8MB L2 cache in particular may suggest this.
Otherwise I read about a PCIe 5 connection with 8 lanes for this chip.

Now, according to the latest rumors from around four weeks ago, the G10, a previously considered but rejected flagship, is to be succeeded by the G31.
Where do you think he is positioned?
Shortened from an originally planned G10 with 512CU's to 384CU's? And then with the rumored 24-32MB L2 cache?
Somehow this all still seems very opaque to me

Edit: This doesn't make any sense to me, so I'll do some research.

XE cores: 24
4 cores per slice
8 EU per core
Vector engines(EU): 192
Total thread count: 1728
Cache 32MB

Metrics via EU
Thread count: 9
INT32 Ops per clock (ADD) 16
FP32 FLOPs per clock (MAD) 32
Register file 64KB

The "1728S is the thread count that got increased from 7 iris > 8 alchemist > 9 battlemage
"192C" is the EU count. they decreased EU count from 16 per core to 8 per core. They also increased register size by 2x compared to alchemist. To make it simple cache by 2x.

leak from slides "14,366.28 Mpix/s.
32 (MAD)*192(EU)*X(clock speed)=14366280 FP32 FLOP = Boost clock speed of 2338mhz
Boost clock speed 2338mhz Sisofsandra is showing the base clock of 1800mhz
You can take my calculations and fit it to the 160CU version and it matches .

At best this SKU is 4060 performance
Biggest they can go is 40.6TFLOPs ~4070ti performance
All the "leaks" were saying it is almost 2x faster then alchemist. Because 1 thread can do almost 2ops per clock now instead of 1.

This is not the G10. G10 was the 56 Xe core variant that was actually canceled. At the same time, work began on a version with 64 Xe cores. I don't know if it's called G31, but it could be.

The performance of Sisoft can only be seen to a limited extent. But it's interesting that Battlemage with only 192 vector units is only 10-15% below Alchemist with 512 vector units. Yes, Battlemage has twice as many ALUs per EU, but that alone doesn't explain it, especially since the clock speed is also lower.

Details for Result ID Intel(R) Arc(TM) A770 Graphics (4608S 512C SM3.0 2.4GHz, 16MB L2, 12.8GB) (OpenCL)

ranker.sisoftware.co.uk

Details for Result ID Intel(R) Graphics d gfx-driver-ci-master-16102 DCH-D (1728S 192C SM3.0 1.8GHz, 8MB L2, 11.6GB) (OpenCL)

ranker.sisoftware.co.uk

The overall score and single-float GP Compute. Half, double or quadruple precision doesn't matter to players.



192VE x 16 ALUs= 3072
160VE x 16 ALUs= 2560

Battlemage has twice as many units per EU as Alchemist. Sisoft is just reading this wrong, it was already like that with Alchemist. But the CUs are right at Sisoft.

Sisofsandra is right but its reporting the "thread count" and thread count increased to 9 compared to 8 on alchemist. And now each thread can do almost 2 IOPs / clock instead of 1.