An amazing card in a niche is worth releasing, especially for AMD because they're not the default choice in the market. So it's better for them to have something that really stands out in a niche than an also-ran generalist that can be safely ignored.
Agreed. If you really need this, you'll be willing to pay the higher cost, which is great for AMD and for the consumer, since they get a killer card for the job.
not gruffi, on servers memory comes first than performance, since you can put as many cards as you can, but the memory, limit the size of the job you can do on the entire cluster. But as the article says, have a niche where 4GVram is enough, so the performance comes very handy.
Honestly, this seems like a really smart move by AMD. They had a card sitting in the stack, found a niche for it and put it out there. The margins are MUCH higher here than as a Radeon Pro Duo and it didn't take much R&D to make it happen. It all comes back to having the right tool for the job. Sure this is a more specialized tool than we typically see, but I'm sure there will be people that will be quite happy to have it. It's not like it is the only card on the market either. If you have massive datasets or primarily need FP64 grunt, then this card won't be all that compelling compared to the rest of the market.
Well Google just invested in these for their Cloud platform and Deep Learning, which apparently rely heavily on FP16/32 performance. This should really boost confidence in AMD. Their stock jumped from $6.5 to about $9.00 as a result. The fact that Vega will be succeeding the chips that are in this S9300x2, are half the size via the 14nm FinFET process (not 1/4 as one might think as 14nm FinFET is 1/2 the size of 28nm Bulk), are more powerful and more efficient, with HBM 2.0, means that AMD could literally put 4x Vega GPU's on a card the same size as the S9300x2.
The implications of that should speak for themselves.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
19 Comments
Back to Article
Pork@III - Thursday, March 31, 2016 - link
Too small VRAM and tooooo small FP64bill.rookard - Thursday, March 31, 2016 - link
Depends on the job at hand. Yes, FP64 is not good, but FP32/FP16 should be phenomenally good.Drumsticks - Thursday, March 31, 2016 - link
I think you missed the point of the entire article...ImSpartacus - Thursday, March 31, 2016 - link
He's just trolling. Anyone that reads the article gets it immediately.xthetenth - Thursday, March 31, 2016 - link
An amazing card in a niche is worth releasing, especially for AMD because they're not the default choice in the market. So it's better for them to have something that really stands out in a niche than an also-ran generalist that can be safely ignored.ChefJeff789 - Thursday, March 31, 2016 - link
Agreed. If you really need this, you'll be willing to pay the higher cost, which is great for AMD and for the consumer, since they get a killer card for the job.Samus - Thursday, March 31, 2016 - link
Nobody is buying this for Solidworks, Pork. RTFA.gruffi - Thursday, March 31, 2016 - link
If the performance is good it's irrelevant if the VRAM is small or large. And the card targets FP16 and FP32 areas. Have you even read the article?Marcelo Viana - Thursday, March 31, 2016 - link
not gruffi, on servers memory comes first than performance, since you can put as many cards as you can, but the memory, limit the size of the job you can do on the entire cluster.But as the article says, have a niche where 4GVram is enough, so the performance comes very handy.
prtskg - Friday, April 1, 2016 - link
And yet it already has secured a contract -http://www.amd.com/en-us/press-releases/Pages/fire...
nismotigerwvu - Thursday, March 31, 2016 - link
Honestly, this seems like a really smart move by AMD. They had a card sitting in the stack, found a niche for it and put it out there. The margins are MUCH higher here than as a Radeon Pro Duo and it didn't take much R&D to make it happen. It all comes back to having the right tool for the job. Sure this is a more specialized tool than we typically see, but I'm sure there will be people that will be quite happy to have it. It's not like it is the only card on the market either. If you have massive datasets or primarily need FP64 grunt, then this card won't be all that compelling compared to the rest of the market.zangheiv - Thursday, March 31, 2016 - link
Perfect GPU for HPC, cloud HW Virtualization as well as large commercial 3D content displaysPork@III - Thursday, March 31, 2016 - link
Yes for 1/3 of 3D content 4GB is enough.StereoPixel - Thursday, March 31, 2016 - link
Is AMD HPC Software has FP16 support?MrSpadge - Thursday, March 31, 2016 - link
OpenCL surely has, that's enough.BillyONeal - Thursday, March 31, 2016 - link
> Popular GPU neutral networkDid you mean neural network?
Ryan Smith - Thursday, March 31, 2016 - link
Well wouldn't a *neutral* neural network be a good thing?;-)But yes, you are correct. Thanks!
SeanJ76 - Monday, April 11, 2016 - link
AMD is GARBAGE!!MLSCrow - Tuesday, November 22, 2016 - link
Well Google just invested in these for their Cloud platform and Deep Learning, which apparently rely heavily on FP16/32 performance. This should really boost confidence in AMD. Their stock jumped from $6.5 to about $9.00 as a result. The fact that Vega will be succeeding the chips that are in this S9300x2, are half the size via the 14nm FinFET process (not 1/4 as one might think as 14nm FinFET is 1/2 the size of 28nm Bulk), are more powerful and more efficient, with HBM 2.0, means that AMD could literally put 4x Vega GPU's on a card the same size as the S9300x2.The implications of that should speak for themselves.