SemiAccurate Forums  

 
Go Back   SemiAccurate Forums > Main Category > GPUs

GPUs Talk about graphics, cards, chips and technologies


Reply
 
Thread Tools Display Modes
  #21  
Old 01-09-2017, 10:41 AM
pTmd pTmd is offline
2^10
 
Join Date: Nov 2010
Posts: 1,954
pTmd is on a distinguished road
Default

Quote:
Originally Posted by Stuckey View Post
Btw- If the HBM2 shown in the pictures is SK Hynix, then it's 7.75 x 11.87 mm. So the GPU package is ~25 x 20 mm. Or ~400mm^2. By rough estimate. http://wccftech.com/amd-vega-gpu-pic...hbm2-official/

In other words, the entire thing is only slightly larger than a Fiji die, ie ~620mm^2. Give or take.
Two stacks take 200 mm^2. The die itself would be something larger than GP102 but smaller than Fiji. So you might end up with a min 700 mm sq interposer. Well it is true that it could not be closed to 2000 mm sq. My fault. But it doesn't affect other points.

Doing for the sake of doing is not a way forward. If you do not have an apparent reasonable use case to take advantage of it, like HBM, existing mature technology isn't too bad. At least Naples seem to be MCM w/o TSV. Crystalwell too.
Reply With Quote
  #22  
Old 01-09-2017, 11:32 AM
Stuckey Stuckey is online now
>intel 4004
 
Join Date: May 2012
Location: 173000G
Posts: 4,237
Stuckey will become famous soon enough
Default

Quote:
Originally Posted by pTmd View Post
Two stacks take 200 mm^2. The die itself would be something larger than GP102 but smaller than Fiji. So you might end up with a min 700 mm sq interposer. Well it is true that it could not be closed to 2000 mm sq. My fault. But it doesn't affect other points.

Doing for the sake of doing is not a way forward. If you do not have an apparent reasonable use case to take advantage of it, like HBM, existing mature technology isn't too bad. At least Naples seem to be MCM w/o TSV. Crystalwell too.
Yeah, they could probably do that. Vega is already using an organic substrate underneath the interposer. So I doubt there would be much to stop them simply making a bigger one and putting two Vega interposers on it ?

If they did that I'm not suggesting it would be doing for the sake of doing, I think it would be for the sake of testing out packaging for MCM. Which I'm pretty certain is the future for large GPUs.
__________________
Do you know what you call people who think everybody *else* is stupid ?

Stupid People.
Reply With Quote
  #23  
Old 01-09-2017, 07:28 PM
sirroman sirroman is offline
2^10
 
Join Date: Jul 2011
Posts: 1,915
sirroman is on a distinguished road
Default

Quote:
Originally Posted by Stuckey View Post
Btw- If the HBM2 shown in the pictures is SK Hynix, then it's 7.75 x 11.87 mm. So the GPU package is ~25 x 20 mm. Or ~400mm^2. By rough estimate. http://wccftech.com/amd-vega-gpu-pic...hbm2-official/

In other words, the entire thing is only slightly larger than a Fiji die, ie ~620mm^2. Give or take.
Sorry for the laziness but what HBM2 stack does GP100 uses?
Reply With Quote
  #24  
Old 01-10-2017, 05:18 AM
Stuckey Stuckey is online now
>intel 4004
 
Join Date: May 2012
Location: 173000G
Posts: 4,237
Stuckey will become famous soon enough
Default

Quote:
Originally Posted by sirroman View Post
Sorry for the laziness but what HBM2 stack does GP100 uses?
Could be SK or Samsung I'd guess ? Don't think they've said which. But I'm guessing both will follow the same package dimensions.

Btw- Sorry my bad, the Vega die is closer to 500mm^2. Not 400mm^2. So closer to a 700mm^2 interposer as pTmd said. Which is only ~17% larger than a bare Fiji die. Pretty remarkable as that should include 2-4 x as much memory.
__________________
Do you know what you call people who think everybody *else* is stupid ?

Stupid People.
Reply With Quote
  #25  
Old 01-10-2017, 05:53 AM
JeeBee JeeBee is offline
2^11
 
Join Date: Jan 2010
Location: UK
Posts: 3,111
JeeBee will become famous soon enough
Default

Yeah, HBM2 is not a productized proof-of-concept! It's a shame the stacks are larger though, but you do need half as many.
Reply With Quote
  #26  
Old 01-10-2017, 06:06 AM
Zizy Zizy is offline
2^10
 
Join Date: Jun 2010
Posts: 1,286
Zizy is on a distinguished road
Default

I think HBM is not core part of Vega and that it is quite interchangeable between HBM and GDDR5. So, Scorpio could have GDDR5 and consumer Vega would have HBM.

The other interesting bit I saw on vega schematic was that cache thingy on the chip. Seems MS will not have any troubles with eSRAM this time - Vega includes it by default.
Reply With Quote
  #27  
Old 01-11-2017, 08:47 PM
integrated integrated is offline
2^11
 
Join Date: Dec 2009
Posts: 3,098
integrated will become famous soon enoughintegrated will become famous soon enough
Default

Quote:
Originally Posted by DCO View Post
My guess is 44CUs, 32ROPs, 408GB/s using 2 stacks of HBM2 if it reaches frequencies higher than 1400MHz. That would be at least 40% faster than Polaris 10.

MS released renders of the Scorpio mobo that are clearly not HBM-based




AMD will probably use the same compute section, but mem-controller and other uncore would be different.

Small Vega probably won't need the USB, SATA, etc controllers that Scorpio will include.
Reply With Quote
  #28  
Old 01-12-2017, 05:26 AM
JeeBee JeeBee is offline
2^11
 
Join Date: Jan 2010
Location: UK
Posts: 3,111
JeeBee will become famous soon enough
Default

I expect Scorpio to be half Vega 10 but with GDDR5X, plus CPU cores and NB functionality.
I wouldn't be surprised to see SB functions remain in a separate chip.

The lack of HBM might not be down to cost, but volume requirements.
Reply With Quote
  #29  
Old 01-12-2017, 11:00 AM
raghu78 raghu78 is offline
8-bit overflow
 
Join Date: Oct 2010
Posts: 439
raghu78 is on a distinguished road
Default

Quote:
Originally Posted by JeeBee View Post
I expect Scorpio to be half Vega 10 but with GDDR5X, plus CPU cores and NB functionality. I wouldn't be surprised to see SB functions remain in a separate chip.

The lack of HBM might not be down to cost, but volume requirements.
I think Scorpio is likely Vega 11 with 384 bit GDDR5 at 7 Ghz.

Vega 11 - 2304 x 2 x 1600 Mhz = 7.3 TFLOPS
Scorpio - 2304 x 2 x 1300 Mhz = 5.99 ~= 6 TFLOPS
Reply With Quote
  #30  
Old 01-12-2017, 11:32 AM
testbug00 testbug00 is offline
>intel 4004
 
Join Date: Dec 2012
Location: San Francisco Bay Area
Posts: 6,656
testbug00 will become famous soon enough
Default

Quote:
Originally Posted by raghu78 View Post
I think Scorpio is likely Vega 11 with 384 bit GDDR5 at 7 Ghz.

Vega 11 - 2304 x 2 x 1600 Mhz = 7.3 TFLOPS
Scorpio - 2304 x 2 x 1300 Mhz = 5.99 ~= 6 TFLOPS
about right in my estimates. Although maybe a little wider (and hence slower for scorpio).
__________________
-Q
Reply With Quote
Reply


Thread Tools
Display Modes

Posting Rules
You may not post new threads
You may not post replies
You may not post attachments
You may not edit your posts

BB code is On
Smilies are On
[IMG] code is On
HTML code is Off


Forum Jump


All times are GMT -5. The time now is 03:26 PM.


Powered by vBulletin® Version 3.8.7
Copyright ©2000 - 2017, vBulletin Solutions, Inc.
SemiAccurate is a division of Stone Arch Networking Services, Inc. Copyright 2009 Stone Arch Networking Services, Inc, all rights reserved.