[ExtremeTech] Building GPUs Out of Entire Wafers Could Turbocharge Performance, Efficiency - Overclock.net - An Overclocking Community
Forum Jump: 

[ExtremeTech] Building GPUs Out of Entire Wafers Could Turbocharge Performance, Efficiency

Reply
 
Thread Tools
post #1 of 23 (permalink) Old 02-21-2019, 08:21 AM - Thread Starter
Waiting for 7nm EUV
 
tpi2007's Avatar
 
Join Date: Nov 2010
Posts: 11,471
Rep: 898 (Unique: 504)
[ExtremeTech] Building GPUs Out of Entire Wafers Could Turbocharge Performance, Efficiency

Quote:
The overarching finding is that a wafer-scale GPU would scale vastly better than any alternative configuration. The researchers modeled both a 24 GPM (GPU Module) and 40 GPM configuration. Configurations above 40 GPUs weren’t feasible due to the difficulty of utilizing an entire 300mm wafer for a single GPU (while we refer to this as wafer-scale processing, the team suggests 40 GPUs rather than a theoretical 100 GPUs per wafer is currently a practical maximum). Performance across a range of tests was 2.97x faster than the equivalent MCM configuration on average for a 24-GPM configuration and 5.2x faster for a 40-GPM configuration.
Source.




Maybe this is what we need to make mGPU a reality again, if this can be good enough to treat multiple GPUs as a single one, that is. If they can come up with an interconnect that scales over ~half the wafer (since the article suggests the whole wafer is infeasible, but then they could put another GPU on the other half), you can cut out the pieces that don't work and salvage probably just as much as before. Not to mention that building smaller GPUs will then be a matter of just cutting it smaller.


tpi2007 is offline  
Sponsored Links
Advertisement
 
post #3 of 23 (permalink) Old 02-21-2019, 08:45 AM
What goes here?
 
speed_demon's Avatar
 
Join Date: Nov 2006
Location: Wisconsin
Posts: 1,919
Rep: 102 (Unique: 80)
Quote: Originally Posted by Kimir View Post
Turbocharge? what if I want to supercharge my GPU instead. [IMG class=inlineimg]/forum/images/smilies/biggrin.gif[/IMG]
It's a good idea until you need to cut a hole in the front of your case to fit it. The turbocharged GPU is more space friendly 😉

Quis custodiet ipsos custodes?


speed_demon is online now  
Sponsored Links
Advertisement
 
post #4 of 23 (permalink) Old 02-21-2019, 09:49 AM
New to Overclock.net
 
DNMock's Avatar
 
Join Date: Jul 2014
Location: Dallas
Posts: 3,647
Rep: 173 (Unique: 127)
Quote: Originally Posted by speed_demon View Post
It's a good idea until you need to cut a hole in the front of your case to fit it. The turbocharged GPU is more space friendly 😉
Obviously those beasts would be enterprise/research/server units and require everything be built around them, but once the tech is developed to get that to work well, trickling down to consumer it's a way to get chiplets into consumer GPU's without all the issues AMD is having using chiplets on GPU's.

Would be super cost effective for gpu makers too only having to design one chip and simply chop the wafer into sections of 1 chip - entry level, 2 chips -low end mainstream, 4 chips - high end mainstream 8 chips - flagship sections which would themselves all be cut down versions of the 24 or 40 chip wafers listed in the thread.

I think it's bloody brilliant.


DNMock is online now  
post #5 of 23 (permalink) Old 02-21-2019, 11:11 AM
New to Overclock.net
 
white owl's Avatar
 
Join Date: Apr 2015
Location: The land of Nod
Posts: 5,374
Rep: 136 (Unique: 103)
Quote: Originally Posted by DNMock View Post
Obviously those beasts would be enterprise/research/server units and require everything be built around them, but once the tech is developed to get that to work well, trickling down to consumer it's a way to get chiplets into consumer GPU's without all the issues AMD is having using chiplets on GPU's.

Would be super cost effective for gpu makers too only having to design one chip and simply chop the wafer into sections of 1 chip - entry level, 2 chips -low end mainstream, 4 chips - high end mainstream 8 chips - flagship sections which would themselves all be cut down versions of the 24 or 40 chip wafers listed in the thread.

I think it's bloody brilliant.
What issues is AMD having with applying chiplets to GPUs?

Quote: Originally Posted by SpeedyVT
If you're not doing extreme things to parts for the sake of extreme things regardless of the part you're not a real overclocker.
Quote: Originally Posted by doyll View Post
The key is generally not which brands are good but which specific products are. Motherboards and GPUs are perfect examples of companies having everything from golden to garbage function/quality.
Hot n Bothered
(12 items)
CPU
4790k 4.7Ghz
Motherboard
Asus Sabertooth Z97 MkII 2
GPU
EVGA GTX 1080 SC
RAM
16gb G.Skill Sniper 2400Mhz
Hard Drive
2x Kingston v300 120gb RAID 0
Hard Drive
WD Blue
Power Supply
Seasonic 620w M12 II EVO
Cooling
Cooler Master 212 Evo
Case
Corsair 450D
Operating System
Windows 10
Monitor
Nixeus EDG27
Other
I have pretty lights.
▲ hide details ▲
white owl is offline  
post #6 of 23 (permalink) Old 02-21-2019, 11:33 AM
New to Overclock.net
 
DNMock's Avatar
 
Join Date: Jul 2014
Location: Dallas
Posts: 3,647
Rep: 173 (Unique: 127)
Quote: Originally Posted by white owl View Post
What issues is AMD having with applying chiplets to GPUs?
Software side. The software (games) still recognizes each chiplet as a different gpu. Using the same Ryzen approach they are unable to homogenize it so that it is recognized by software as a single unit. That's why Navi is monolithic instead of using chiplets.


Gotta backtrack now a bit. 8 200mm chiplets would be way too huge.

Here is scaled up image of about what 6 chiplets at 200mm^2 would look compared to a Volta Tesla GV100 and lol, it's stupid huge.
Attached Thumbnails
Click image for larger version

Name:	bigdie.jpg
Views:	36
Size:	1.39 MB
ID:	254786  



DNMock is online now  
post #7 of 23 (permalink) Old 02-21-2019, 12:07 PM - Thread Starter
Waiting for 7nm EUV
 
tpi2007's Avatar
 
Join Date: Nov 2010
Posts: 11,471
Rep: 898 (Unique: 504)
Quote: Originally Posted by DNMock View Post
Software side. The software (games) still recognizes each chiplet as a different gpu. Using the same Ryzen approach they are unable to homogenize it so that it is recognized by software as a single unit. That's why Navi is monolithic instead of using chiplets.


Gotta backtrack now a bit. 8 200mm chiplets would be way too huge.

Here is scaled up image of about what 6 chiplets at 200mm^2 would look compared to a Volta Tesla GV100 and lol, it's stupid huge.

You don't need to backtrack, after all your baseline GPU block was entry level, not a ~200mm² block, which is usually in the mainstream/mid-range (GTX 1060 / RX 480).

You take the 1050 Ti, which is 132 mm² or go even lower, the 1030 is 74 mm². Ideally, something in between, at ~100mm². If you put 8 together you get a chip around the size of the Titan V, but since you have freedom to scale and cut from wherever you want on the wafer, it would be much cheaper to get to market.




Edit: This approach needs a chiplet design though, but it's easy: you make wafers filled with memory controllers + video decode/encode block + video output block + PCIe interface.


Edit 2: And as a bonus, GPU makers can do like AMD is doing with Zen 2, make the chips that contain the memory controllers + video decode/encode block + video output block + PCIe interface on an older node, for even lower costs and better yields.



Last edited by tpi2007; 02-21-2019 at 01:52 PM.
tpi2007 is offline  
post #8 of 23 (permalink) Old 02-21-2019, 01:02 PM
What goes here?
 
speed_demon's Avatar
 
Join Date: Nov 2006
Location: Wisconsin
Posts: 1,919
Rep: 102 (Unique: 80)
Agreed the chiplet design makes a lot of sense for this type of application. Any ideas on when we'll see a GPU using this design for sale?

Quis custodiet ipsos custodes?


speed_demon is online now  
post #9 of 23 (permalink) Old 02-21-2019, 04:40 PM
OG AMD
 
Redwoodz's Avatar
 
Join Date: Jan 2010
Location: Silicon Valley
Posts: 8,605
Rep: 597 (Unique: 465)
Quote: Originally Posted by DNMock View Post
Software side. The software (games) still recognizes each chiplet as a different gpu. Using the same Ryzen approach they are unable to homogenize it so that it is recognized by software as a single unit. That's why Navi is monolithic instead of using chiplets.


Gotta backtrack now a bit. 8 200mm chiplets would be way too huge.

Here is scaled up image of about what 6 chiplets at 200mm^2 would look compared to a Volta Tesla GV100 and lol, it's stupid huge.

So in other words this as a concept has the same problem. Nice story.

https://valid.x86.fr/cache/banner/4d8m5e-6.png
https://valid.x86.fr/cache/banner/bs72bx-6.png
AMD Athlon II 450 @4.74GHz http://valid.canardpc.com/show_oc.php?id=2204588
AMD MaxxMemm Rankings
960T @ 4.6GHz http://valid.canardpc.com/show_oc.php?id=2171450
RedwoodCustomPC
AM4
(6 items)
Redwoodz
(10 items)
Professional
(13 items)
CPU
Ryzen 5 2600
Motherboard
ASRock X470 Taichi
GPU
MSI RX570 Armor MKII 8GB
RAM
GSKill Trident Z D-die
RAM
TeamGroup Excalibur
Power Supply
GSKill Ripjaws PS850G
CPU
AMD FX Vishera 8350
Motherboard
ASRock Extreme 9 990FX
GPU
HIS X 7970 IceQ
RAM
Gskill Trident
Hard Drive
Samsung 840 Pro
Power Supply
Seasonic SS-760XP
Cooling
TT Water 2.0
Monitor
Samsung Series 9 27"
Keyboard
Mionix Zibal
Mouse
Mionix 5000
CPU
Phenom II X6 1100t
Motherboard
MSI 890FX GD65
GPU
MSI Radeon HD5670
RAM
GSkill RipjawsX DDR3 PC3 12800 2x4GB CL8
Hard Drive
WD Black 1TB SATA III
Optical Drive
Samsung BD
Power Supply
Kingwin Lazer Platinum 500w
Cooling
Zalman 9900MAX
Case
Fractal Design R3
Operating System
Windows 7 64 Professional
Monitor
AOC 22" LED
Keyboard
Logitech
Other
Samsung 470 SSD 128GB
▲ hide details ▲


Redwoodz is offline  
post #10 of 23 (permalink) Old 02-21-2019, 05:52 PM
New to Overclock.net
 
DNMock's Avatar
 
Join Date: Jul 2014
Location: Dallas
Posts: 3,647
Rep: 173 (Unique: 127)
Quote: Originally Posted by Redwoodz View Post
So in other words this as a concept has the same problem. Nice story.
The problem is there isn't enough incentive to force the issue yet with software developers. If the article is to be believed, a 300% to 500% performance increase is big enough to prompt a big push from AMD/Nvidia/Intel to make it happen.


DNMock is online now  
Reply

Quick Reply
Message:
Options

Register Now

In order to be able to post messages on the Overclock.net - An Overclocking Community forums, you must first register.
Please enter your desired user name, your email address and other required details in the form below.
User Name:
If you do not want to register, fill this field only and the name will be used as user name for your post.
Password
Please enter a password for your user account. Note that passwords are case-sensitive.
Password:
Confirm Password:
Email Address
Please enter a valid email address for yourself.
Email Address:

Log-in



Currently Active Users Viewing This Thread: 1 (0 members and 1 guests)
 
Thread Tools
Show Printable Version Show Printable Version
Email this Page Email this Page


Forum Jump: 

Posting Rules  
You may post new threads
You may 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
Trackbacks are Off
Pingbacks are Off
Refbacks are Off