Intels upcoming high end GPU Arctic Sound with 16384 Cores and 42 TFlops of FP32?

moebius · 2021-09-19, 16:53

article from wccftech.com

kriesel · 2021-09-19, 18:19

And https://www.tomshardware.com/news/in...xe-hp-pictured and not a word of DP performance on either article.

moebius · 2021-09-19, 21:21

HBM2E memory with a peak bandwidth of up to 716 GB/s is accetable, better than Titan V.

kriesel · 2021-09-20, 00:41

But a step down from 4kbit wide HBM2 1TB/sec on Radeon VII

M344587487 · 2021-09-20, 07:43

If I'm understanding correctly the memory bandwidth may scale with tile count, meaning the 2T version could exceed an R7's bandwidth.

henryzz · 2021-09-20, 12:19

Quote:

Originally Posted by kriesel

And https://www.tomshardware.com/news/in...xe-hp-pictured and not a word of DP performance on either article.

The one you linked mentioned FP64 support at least in the xe-hp variant. I would imagine that we will see 1/32 speed or similar dp support in the consumer variant as well for code compatibility.

kriesel · 2021-09-20, 12:41

Quote:

Originally Posted by henryzz

The one you linked mentioned FP64 support

"Supported". No performance number, as I said. "Support" means > 0 DP FLOPS is all. Windows 10 Calculator qualifies for that. It could be DP/SP 1/64 and "supported". By emulation in the driver might still be called supported.

kruoli · 2021-09-20, 13:47

Quote:

Originally Posted by kriesel

Windows 10 Calculator qualifies for that.

No, because it is not hardware. calc.exe needs hardware to execute DP instructions. You cannot compare software to hardware here. Only if you extended the meaning of "support" to "make use of", it might qualify. But this meaning would make no sense for hardware the other way around.

Edit: Additionally, calc.exe does not provide any kind of interface for executing DP instructions. The GUI is definitely not an interface for IEEE 754 numbers, not even in the programmer view.

kriesel · 2021-09-20, 14:51

Quote:

Originally Posted by kruoli

calc.exe does not provide any kind of interface for executing DP instructions

I only used calc as an extreme example of DP support/implementation with very low nonzero throughput. But there are folks who try interfacing to such things.
https://stackoverflow.com/questions/...win32-api-in-c

2021-09-19, 16:53	#1
moebius "CharlesgubOB" Jul 2009 Germany 3³·23 Posts	Intels upcoming high end GPU Arctic Sound with 16384 Cores and 42 TFlops of FP32? article from wccftech.com Last fiddled with by moebius on 2021-09-19 at 16:55

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Is there any sensible auxiliary task for HT logical cores when physical cores already used for PRP?	hansl	Information & Answers	5	2019-06-17 14:07
FP64 or FP32	marigonzes	Software	5	2017-02-11 01:06
Arctic sea ice thread	Oddball	Science & Technology	4	2010-09-09 18:45
sieving speeds for Intels	jasong	Sierpinski/Riesel Base 5	11	2007-08-09 00:15
Intels Intresting Road	moo	Hardware	7	2005-12-13 02:20

2021-09-19, 18:19	#2
kriesel "TF79LL86GIMPS96gpu17" Mar 2017 US midwest 3×2,161 Posts	And https://www.tomshardware.com/news/in...xe-hp-pictured and not a word of DP performance on either article.

2021-09-19, 21:21	#3
moebius "CharlesgubOB" Jul 2009 Germany 3³×23 Posts	HBM2E memory with a peak bandwidth of up to 716 GB/s is accetable, better than Titan V.

2021-09-20, 00:41	#4
kriesel "TF79LL86GIMPS96gpu17" Mar 2017 US midwest 3·2,161 Posts	But a step down from 4kbit wide HBM2 1TB/sec on Radeon VII

2021-09-20, 07:43	#5
M344587487 "Composite as Heck" Oct 2017 19×47 Posts	If I'm understanding correctly the memory bandwidth may scale with tile count, meaning the 2T version could exceed an R7's bandwidth.