NVIDIA Grace

Simbelmynë · April 14, 2021, 05:13

This may or may not be a golden nugget for CFD.

What do you think?

https://www.anandtech.com/show/16610...-in-ai-systems

Seeing that EPYC with DDR4 3200 MT/s memory has a theoretical bandwidth of about 205 GB/s. If Genoa will have 12 memory channels and DDR5 memory, then it may have even greater bandwidth than the ">500 GB/s" of NVIDIA Grace. So perhaps this is not a golden nugget, but if NVIDIA continues to push increasing bandwidth as a selling point (also on CPU) then perhaps we will see more rapid improvement in this area from AMD and INTEL as well. This is really good news for CFD folks.

LaminarFlow · April 16, 2021, 05:32

Well DDR5 for AMD and Intel platforms is just around the corner. Also ARM has a lack of command sets compared to x86 architecture. So it depends on the problem which one is better. Let's see which platform will be most profitable (peformance, investment cost, power consumption).

April 14, 2021, 05:13	NVIDIA Grace	#1
Simbelmynë Senior Member Join Date: May 2012 Posts: 546 Rep Power: 15	This may or may not be a golden nugget for CFD. What do you think? https://www.anandtech.com/show/16610...-in-ai-systems Seeing that EPYC with DDR4 3200 MT/s memory has a theoretical bandwidth of about 205 GB/s. If Genoa will have 12 memory channels and DDR5 memory, then it may have even greater bandwidth than the ">500 GB/s" of NVIDIA Grace. So perhaps this is not a golden nugget, but if NVIDIA continues to push increasing bandwidth as a selling point (also on CPU) then perhaps we will see more rapid improvement in this area from AMD and INTEL as well. This is really good news for CFD folks.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Ansys Installation on Docker Container	mohsen.shiea	ANSYS	11	March 14, 2024 11:18
OpenFOAM and nvidia maximus	firat	OpenFOAM Programming & Development	1	December 10, 2012 12:17
NVIDIA Tesla GPU	ztdep	Hardware	0	December 7, 2010 21:00
[OpenFOAM] Error API mismatch the NVIDIA kernel	Erik	ParaView	0	February 3, 2009 06:01
New Nvidia gpu aimed at gpgpu	bmeagle	OpenFOAM	0	November 9, 2006 09:41

April 16, 2021, 05:32		#2
LaminarFlow New Member Join Date: Dec 2020 Location: Germany Posts: 12 Rep Power: 5	Well DDR5 for AMD and Intel platforms is just around the corner. Also ARM has a lack of command sets compared to x86 architecture. So it depends on the problem which one is better. Let's see which platform will be most profitable (peformance, investment cost, power consumption).