• Light/Dark
  • Ever wanted an RSS feed of all your favorite gaming news sites? Go check out our new Gaming Headlines feed! Read more about it here.

Nvidia GTC - 14 May 6am PST

JesseDeya

JesseDeya

Member
Oct 27, 2017
164
CypherSignal said:
I guess someone missed the memo that that press release was meant for 6am Pacific, not 6am Eastern?
Click to expand...
Click to shrink...

OOOPS!

Also for those hoping for 3000 series snippets...

"Nvidia did not release any information Wednesday about consumer GPUs using Ampere, but when asked by a reporter in the briefing about the difference between enterprise and consumer approaches to Ampere, Huang said "there's great overlap in the architecture, but not in the configuration."
 
ILikeFeet

ILikeFeet

DF Deet Master
Banned
Oct 25, 2017
61,987
Wonder if the changes in the tensor cores will allow for acceleration of other functions now
 
ILikeFeet

ILikeFeet

DF Deet Master
Banned
Oct 25, 2017
61,987
c0de said:
You mean something like a trend to "general purpose" tensor cores? ;)
Click to expand...
Click to shrink...
Cuda cores with another name!

tuxfool said:
Like what.

They're pretty much stuck to processing NN. Their utility outside of that is nil. Tensor cores are hyper specific, more generic, you're talking about CUDA cores.
Click to expand...
Click to shrink...
Who knows. That's why I ask. I know it's limited to NN but I'm questioning if the change would expand its potential. Faster simulations, more accurate sums, stuff like that.
 
Vash63

Vash63

Member
Oct 28, 2017
1,681
Hopefully we get some more real details today. That transistor count is absolutely staggering but it's also somewhat expected for their enterprise class part. I would be curious if (G)A102 and 104 have similar gains over Turing.
 
Trago

Trago

Member
Oct 25, 2017
3,600
giphy.gif
 
dgrdsv

dgrdsv

Member
Oct 25, 2017
11,846
videocardz.com

NVIDIA announces A100 Tensor Core GPU specifications - VideoCardz.com

NVIDIA Tesla A100 features 6912 CUDA Cores The card features 7nm Ampere GA100 GPU with 6912 CUDA cores and 432 Tensor cores. This GPU has a die size of 826mm2 and 54-billion transistors. The GPU is divided into 108 Streaming Multiprocessors. The GPU in Tesla A100 is clearly not the full chip...
videocardz.com videocardz.com
So GA100 in DGX-A100 is cut down likely, with only 108 SMs being active out of 128 present in the chip.
 

Deleted member 65994

User requested account closure
Banned
Apr 14, 2020
627
damn, on 1,6ghz that's 22 tflops, on 1,8ghz that would be 24 tflops.

Also it looks like one out of the 6 hbm2e stack is put off, that's why it have 40gb vram instead of 48gb.
 
ILikeFeet

ILikeFeet

DF Deet Master
Banned
Oct 25, 2017
61,987
dgrdsv said:
videocardz.com

NVIDIA announces A100 Tensor Core GPU specifications - VideoCardz.com

NVIDIA Tesla A100 features 6912 CUDA Cores The card features 7nm Ampere GA100 GPU with 6912 CUDA cores and 432 Tensor cores. This GPU has a die size of 826mm2 and 54-billion transistors. The GPU is divided into 108 Streaming Multiprocessors. The GPU in Tesla A100 is clearly not the full chip...
videocardz.com videocardz.com
So GA100 in DGX-A100 is cut down likely, with only 108 SMs being active out of 128 present in the chip.
Click to expand...
Click to shrink...
Then what's gonna use the full chip?
 

Deleted member 65994

User requested account closure
Banned
Apr 14, 2020
627
Seems like a huge focus on transistors, like how navi used way more transistors compared to GNC. And Mark Cerny mentioned why one tflop performance better than another.
 
pswii60

pswii60

Member
Oct 27, 2017
26,662
The Milky Way
thefollower said:
And Mark Cerny mentioned why one tflop performance better than another.
Click to expand...
Click to shrink...
We already knew about that. Nvidia GPUs have had significantly more performance per flop than AMD GPUs for years. RDNA2 seeks to close that gap, albeit Ampere might widen it again. We'll see.
dgrdsv said:
videocardz.com

NVIDIA announces A100 Tensor Core GPU specifications - VideoCardz.com

NVIDIA Tesla A100 features 6912 CUDA Cores The card features 7nm Ampere GA100 GPU with 6912 CUDA cores and 432 Tensor cores. This GPU has a die size of 826mm2 and 54-billion transistors. The GPU is divided into 108 Streaming Multiprocessors. The GPU in Tesla A100 is clearly not the full chip...
videocardz.com videocardz.com
So GA100 in DGX-A100 is cut down likely, with only 108 SMs being active out of 128 present in the chip.
Click to expand...
Click to shrink...
What would a GA100 be in the consumer line-up? The new Titan RTX? Assuming GA102 is 3080Ti.
 
DSP

DSP

Member
Oct 25, 2017
5,120
It has fewer tensor cores, huh. Probably something has changed with them.
 

Deleted member 10428

User requested account closure
Banned
Oct 27, 2017
1,083
I hope we'll get a timeframe for mobile Ampere GPUs. That would make the wait just a tiny bit easier.

edit: eh
 
FHIZ

FHIZ

Member
Oct 28, 2017
1,942
So last time around, how long was it between announcement of the architecture and announcement of consumer cards?
 

Genio88

Banned
Jun 4, 2018
964
As expected no RTX 3000 information and it's not even a live conference, just a few videos about different professional topics and AI, i hope you weren't that amped
 
sirap

sirap

Member
Oct 25, 2017
8,210
South East Asia
I can't believe I have a better kitchen than the co-founder and CEO of Nvidia.

TBF a man of his wealth probably does very little cooking lol.
 
dunkzilla

dunkzilla

alt account
Banned
Dec 13, 2018
4,762
this is probably the wrong thread but is that image of the 3XXX release dates and capabilities correct? I need to upgrade my 1060.
 

Deleted member 13560

User requested account closure
Banned
Oct 27, 2017
3,087
I feel bad for the people who thought there was going to be actual consumer news from this... Even though for the last few years these keynotes have never announced anything of the sort.

That Mellanox Switch is hype.
 
You must log in or register to reply here.