Oak Ridge National Laboratory gets Kepler K20 "Big K"

Toggle sidebar Toggle sidebar
Keysplayr

Keysplayr

Elite Member
Jan 16, 2003
21,211
50
91
Last edited:
blastingcap

blastingcap

Diamond Member
Sep 16, 2010
6,654
5
76
As I've stated repeatedly, Tesla/Quadro + TSMC 28nm constraints = good luck waiting for major GeForce GTX 680 price drops. You may see more price movement on GeForce cards that use chips that are not used for Tesla/Quadro, such as harvested GK104 parts, GK106, etc. As TSMC 28nm matures, yields go up, etc. prices for all 28nm GPUs may slide down, though. We are already seeing that.

Btw, the better Quadro/Tesla do without impacting GeForce too much, the better for even GeForce/gamers, because it spreads the cost of GPU development more. At least for NV. AMD's pro graphics and HPC divisions are a joke compared to NV's.
 
ViRGE

ViRGE

Elite Member, Moderator Emeritus
Oct 9, 1999
31,516
167
106
Holy cow, that's a lot of 28nm wafers going to just Oak Ridge. Even an overly optimistic estimate would put GK110 at 100 working dies per wafer*. Oak Ridge needs 100 of those wafers.:eek:
 
M

mazeroth

Golden Member
Jan 31, 2006
1,821
2
81
I wonder if this thing could render a movie in real-time? Think about it. 24 fps = 417 GPU's rendering each frame. Probably not a recent 3D movie but I bet it would ship out Toy Story 1 in no time.
 
Keysplayr

Keysplayr

Elite Member
Jan 16, 2003
21,211
50
91
ViRGE said:
Holy cow, that's a lot of 28nm wafers going to just Oak Ridge. Even an overly optimistic estimate would put GK110 at 100 working dies per wafer*. Oak Ridge needs 100 of those wafers.:eek:
Click to expand...

How do you figure? Not saying you're wrong, just wondering where you're getting your math from.
 
BD231

BD231

Lifer
Feb 26, 2001
10,568
138
106
As if progression isn't slow enough as it is, super computers will still be suckin up all the fresh gear.
 
AnandThenMan

AnandThenMan

Diamond Member
Nov 11, 2004
3,991
626
126
TSMC is currently churning out 100,000+ 28nm wafers per month, so 100 is a drop in the bucket.
 
ViRGE

ViRGE

Elite Member, Moderator Emeritus
Oct 9, 1999
31,516
167
106
Keysplayr said:
How do you figure? Not saying you're wrong, just wondering where you're getting your math from.
Click to expand...
Sure.

300mm wafer = 70,685mm2

Assume GK110 is 530mm2 (it's probably larger, but we'll give NV the benefit of a doubt here). That means 133 dies per wafer.

Now you need to account for partial dies (let's say 13 to make the math easy) which brings us to 120 dies.

Now account for yield, we'll be generous and say 83%. That brings us down to a flat 100 dies per wafer, which means Oak Ridge would be getting 100 wafers worth of chips.

This is all quick & dirty math and I'm sure the numbers are off (I'm probably ending up with too many good dies per wafer). But I just wanted to point out that with 28nm wafers still being at a premium, Oak Ridge is getting what amounts to a pretty hefty allocation for any single user.
 
Last edited:
Keysplayr

Keysplayr

Elite Member
Jan 16, 2003
21,211
50
91
ViRGE said:
Sure.

300mm wafer = 70,685mm2

Assume GK110 is 530mm2 (it's probably larger, but we'll give NV the benefit of a double here). That means 133 dies per wafer.

Now you need to account for partial dies (let's say 13 to make the math easy) which brings us to 120 dies.

Now account for yield, we'll be generous and say 83%. That brings us down to a flat 100 dies per wafer, which means Oak Ridge would be getting 100 wafers worth of chips.

This is all quick & dirty math and I'm sure the numbers are off (I'm probably ending up with too many good dies per wafer). But I just wanted to point out that with 28nm wafers still being at a premium, Oak Ridge is getting what amounts to a pretty hefty allocation for any single user.
Click to expand...

Ah, very nice, TY. Sounds like it's very possible.

Lessee. Rumor had TSMC charging 8000 per wafer (good or bad dies not mattering).
ORNL will ultimately have just shy of 15000 K20s. Assuming your 100 GK110 per die is close to accurate, 150 wafers would be needed to accomodate Oak Ridge's order. 150x8000 =1.2million + PCB, GDDR5 and components, R&D. ORNL is paying 1000 per K20 according to the article. @15million.
Maybe I should get into the semiconductor biz!!! :D Anyone have 10 billion I can borrow?
 
Last edited:
T

tviceman

Diamond Member
Mar 25, 2008
6,734
514
126
www.facebook.com
ViRGE said:
Sure.

300mm wafer = 70,685mm2

Assume GK110 is 530mm2 (it's probably larger, but we'll give NV the benefit of a double here). That means 133 dies per wafer.

Now you need to account for partial dies (let's say 13 to make the math easy) which brings us to 120 dies.

Now account for yield, we'll be generous and say 83%. That brings us down to a flat 100 dies per wafer, which means Oak Ridge would be getting 100 wafers worth of chips.

This is all quick & dirty math and I'm sure the numbers are off (I'm probably ending up with too many good dies per wafer). But I just wanted to point out that with 28nm wafers still being at a premium, Oak Ridge is getting what amounts to a pretty hefty allocation for any single user.
Click to expand...

I'm betting the release version of GK110 will be less problematic than GF100 was for Nvidia. Also, I'm guessing K20 will compromise of not-fully-functional GK110's. That would be the smart move, anyways. As yields improve, Nvidia can make a K25 with fully functional GK110's that they can deliver in volume.
 
T

tviceman

Diamond Member
Mar 25, 2008
6,734
514
126
www.facebook.com
blastingcap said:
As I've stated repeatedly, Tesla/Quadro + TSMC 28nm constraints = good luck waiting for major GeForce GTX 680 price drops. You may see more price movement on GeForce cards that use chips that are not used for Tesla/Quadro, such as harvested GK104 parts, GK106, etc. As TSMC 28nm matures, yields go up, etc. prices for all 28nm GPUs may slide down, though. We are already seeing that.

Btw, the better Quadro/Tesla do without impacting GeForce too much, the better for even GeForce/gamers, because it spreads the cost of GPU development more. At least for NV. AMD's pro graphics and HPC divisions are a joke compared to NV's.
Click to expand...

100,000 wafers a month. I think the supply constraints are going to be over as of Oct. 1st. And if the rumors of slow gtx660ti sales are true, as well taking into account Intel adjusting it's quarterly revenue forecast downward, I think we have a recipe for noticeably lower prices (at least on Nvidia's current lineup) before Christmas time. Not that it matters for you and I; we already blew our load and bought new cards.
 
Last edited:
Keysplayr

Keysplayr

Elite Member
Jan 16, 2003
21,211
50
91
AnandThenMan said:
What else is significant in this link besides the wafer capacity that I posted, or just a confirmation/source?
Click to expand...

Just an additional source. Not disputing anything you said, although we've come to expect that from each other eh? :sneaky:
 
A

Arzachel

Senior member
Apr 7, 2011
903
76
91
ViRGE said:
Sure.

300mm wafer = 70,685mm2

Assume GK110 is 530mm2 (it's probably larger, but we'll give NV the benefit of a doubt here). That means 133 dies per wafer.

Now you need to account for partial dies (let's say 13 to make the math easy) which brings us to 120 dies.

Now account for yield, we'll be generous and say 83%. That brings us down to a flat 100 dies per wafer, which means Oak Ridge would be getting 100 wafers worth of chips.

This is all quick & dirty math and I'm sure the numbers are off (I'm probably ending up with too many good dies per wafer). But I just wanted to point out that with 28nm wafers still being at a premium, Oak Ridge is getting what amounts to a pretty hefty allocation for any single user.
Click to expand...

83% isn't generous, it's impossible. It's likely what Nvidia sees with GK104, they'd be happily cranking GK110 since spring with that kind of yields. Between 30% and 5% is probably the best bet.
 
ViRGE

ViRGE

Elite Member, Moderator Emeritus
Oct 9, 1999
31,516
167
106
Arzachel said:
83% isn't generous, it's impossible. It's likely what Nvidia sees with GK104, they'd be happily cranking GK110 since spring with that kind of yields. Between 30% and 5% is probably the best bet.
Click to expand...
Perhaps. The yield isn't particularly important. The only point I'm making is that Titan is using a lot of silicon.:eek:
 
Keysplayr

Keysplayr

Elite Member
Jan 16, 2003
21,211
50
91
ViRGE said:
Perhaps. The yield isn't particularly important. The only point I'm making is that Titan is using a lot of silicon.:eek:
Click to expand...

Even if it took 1000 wafers at 15% yield to fill ORNL's order, that is still only less than 1/3 of TSMC's wafer start ability for a single day which is somewhere around 3300 wspd.
 
S

SlowSpyder

Lifer
Jan 12, 2005
17,305
1,002
126
I wonder if they ever did end up getting that Fermi supercomputer? My guess is they did not, and Charlie was right about that. But much like your average video card buyer, waiting for the next gen always sucks, until the next gen is finally here and you feel that much better about buying 'now' instead. :)
 
B

BenSkywalker

Diamond Member
Oct 9, 1999
9,140
67
91
I wonder if they ever did end up getting that Fermi supercomputer? My guess is they did not, and Charlie was right about that. But much like your average video card buyer, waiting for the next gen always sucks, until the next gen is finally here and you feel that much better about buying 'now' instead.
Click to expand...

http://i.top500.org/system/176544

They did, Charlie was wrong(what a shocker, heh) and it is currently ranked sixth in the world. This machine they are building with BigK parts is actually an upgrade to their Fermi based machine.
 
Keysplayr

Keysplayr

Elite Member
Jan 16, 2003
21,211
50
91
What was that about Chuck, SlowSpyder? :)
Don't ever give that man credit for anything.
A sensationalist "journalist" that writes for a site designated as a "satire" site. Zero cred.
 
Last edited:
You must log in or register to reply here.

TRENDING THREADS

COMPANY
RESOURCES
FOLLOW
Top Bottom