[SemiAccurate] Tesla K20 specs: 13 SMX, GeForce probably 12-13 SMX

[blastingcap]

http://semiaccurate.com/2012/11/02/n...t-28nm-yields/

[3DVagabond]

Charlie's scenario is only one possibility. Another is they fused them off purely for power usage, nothing at all to do with yields. The clocks being so low would also point to that.

[Jaydip]

So he is at it again

[3DVagabond]

Charlie is "usually" right when he is quoting his own sources. This is him just reading an article and drawing conclusions based off of it.

[blastingcap]

If they relieve the memory bandwidth constraints and allow GeForce to be somewhat less power efficient = higher voltages and clocks than Tesla, then we're looking at ~35% more performance than a GTX 680. Considering that HD8970 is supposed to be less than 20% faster than HD7970, it seems that NV is about to reclaim the single-GPU crown.

[Plimogz]

so, if I'm getting this straight, NV would've delivered 18K+ 14/15 SMX to ORNL, but yields would bad enough to only market a 13/15 SMX part as the professional market compute part and slot in below that a 12 or 13 (uncertainty per the article) SMX part for the consumer graphics market?

While I expect that even such a fraction of their monster 7.1 billion xtor part would be formidable in itself, I can't remember when gamers ever received a part so crippled in lieu of a high end chip.
Weird.

[boxleitnerb]

Originally Posted by blastingcap

http://semiaccurate.com/2012/11/02/n...t-28nm-yields/

Charlie purposefully left out that K20 at ORNL is 14 SMX, his other stuff is pure speculation to make Nvidia look bad. Nothing new in the west.

[Silverforce11]

He is incorrect one a major point, the reason for the fused part and low clocks = to fit the 225W TDP. Because these K20s are going to be used as a upgrade, replacing Fermi, which was also a 225W TDP HPC part. Massive sever upgrades would not risk replacing the infrastructure, especially the water cooling units and aircond for the facility. TDP being the same is critical.

I expect yields to be low, but it seems shipping 18,000 units in a few months points to TSMC either a) starting the ramp a long long time ago (possible, but unlikely due to 28nm constraints on every other part, ie. mobile), or b) yields are not so bad! (more likely)

[VulgarDisplay]

Originally Posted by blastingcap

If they relieve the memory bandwidth constraints and allow GeForce to be somewhat less power efficient = higher voltages and clocks than Tesla, then we're looking at ~35% more performance than a GTX 680. Considering that HD8970 is supposed to be less than 20% faster than HD7970, it seems that NV is about to reclaim the single-GPU crown.

You need to remember that the 7970 itself can be 20% faster than the 7970 through fairly reasonable overclocking. I don't think this refresh is going to be that negative for AMD.

[boxleitnerb]

The reason for selling 14SMX parts in K20 for ORNL is not only TDP. Obviously, yields on the full parts are always lower, so that means you need more wafers and more time to produce a certain amount. AMD and Intel are beginning to invade the HPC market with their own solution, so it is only logical to fulfill the contracts as quickly as possible. 5% more or less performance is irrelevant when you can supply 3 times as many cards.

[sontin]

Good old Charlie.

With the w9000 AMD is selling a 28 SIMD GCN chip in the server market. SO i guess there is no full GCN on the market right now?

[HurleyBird]

Originally Posted by Silverforce11

He is incorrect one a major point, the reason for the fused part and low clocks = to fit the 225W TDP. Because these K20s are going to be used as a upgrade, replacing Fermi, which was also a 225W TDP HPC part. Massive sever upgrades would not risk replacing the infrastructure, especially the water cooling units and aircond for the facility. TDP being the same is critical.

Originally Posted by 3DVagabond

Charlie's scenario is only one possibility. Another is they fused them off purely for power usage, nothing at all to do with yields. The clocks being so low would also point to that.

Wishful thinking. I thought the two of you would know better.

Enabling all of the SMXes while lowering the clock rate and voltage further to compensate would allow for better power consumption at the same performance level, that's basic mathematics. Yields are the only possible reason for disabling SMXes. Given the size of GK110 this isn't at all surprising, and may have been expected by Nvidia from the onset. Expect a fully functional part six months to a year after big Kepler officially launches.

[Silverforce11]

Originally Posted by HurleyBird

Wishful thinking. I thought the two of you would know better.

Enabling all of the SMXes while lowering the clock rate and voltage further to compensate would allow for better power consumption at the same performance level, that's basic mathematics. Yields are the only possible reason for disabling SMXes. Given the size of GK110 this isn't at all surprising, and may have been expected by Nvidia from the onset. Expect a fully functional part six months to a year after big Kepler officially launches.

Depends on the fusing techniques, they can effectively cut TDP by fusing if they need to fit a lower TDP mark.

You can bet a lot of their K20 sales will be for customers who are upgrading from Fermi. Slotting in an identical TDP card is easier.

[BenSkywalker]

Charlie is "usually" right when he is quoting his own sources.

If Charlie said the sun was going to come up tomorrow, I would be getting my wife and kids together to say goodbye.

[tviceman]

Ryan Smith said that at 1.3 TF, the K20's in ORNL has to be 14 SMX's at 732mhz. About half way to two thirds down in the comments: http://www.anandtech.com/show/6421/i...cores?all=true I'll believe anandtech and hpcwire any day over Charlie anything. Especially when using real, mathematical figures and facts.

EDIT: The article discussion thread so far on semiaccurate pretty much has everyone telling him he is wrong too. http://semiaccurate.com/forums/showthread.php?t=6720

[Keysplayr]

K20 could be 15 SMX with 1 disabled to meet the 225W power envelope they were looking for. With the last SMX enabled that might bring it up to 250W and multiply that by 18000+. Lot's O Watts.

[ShintaiDK]

Why do people read that sensationalist site again?

[tviceman]

Originally Posted by ShintaiDK

Why do people read that sensationalist site again?

It's good entertainment!

[Keysplayr]

I blocked it at my router just in case I lose my head and click the link. No hits from me for Charlie. Ever. Not even by accident or redirected url.

[sontin]

Originally Posted by tviceman

It's good entertainment!

It was in the past. Now he is writing nothing new. It always the same with the same sentences and the same message.

[SirPauly]

Imho,

I don't mind Charlie and try to understand the subjective context: sensationalism, conjecture, speculation and the name of the site is semi-accurate --- not accurate. Seems to have an odd negative fetish for nVidia and sometimes his journalism suffers from this.

He was semi-accurate on the GK-104 - win in many important metrics and efficiency but off on some things as well.

Pretty tough to ignore HPC and AnandTech and what they offer -- Time will tell to see if nVidia offers a GeForce derivative and if they do - what the specs may be.

[Silverforce11]

I'm pretty sure reading into Charlie's previous sources, he has friends at MSI and ASUS.. other AIBs im not certain.

It's pretty clear he has no source inside NV itself. But then again, NV is really leak-free.

I really do hope GK110 for consumers will be a high TDP high perf part, none of this crippled rubbish to fit a certain power envelope.. $600 GPU who cares?!

[iMacmatician]

Charlie implies here that the 14 SMX K20s are "special" chips for ORNL (and maybe other supercomputers), and the 13 SMX K20 is the "real" version that will actually be available on the market.

[tviceman]

Originally Posted by Silverforce11

I'm pretty sure reading into Charlie's previous sources, he has friends at MSI and ASUS.. other AIBs im not certain.

It's pretty clear he has no source inside NV itself. But then again, NV is really leak-free.

I really do hope GK110 for consumers will be a high TDP high perf part, none of this crippled rubbish to fit a certain power envelope.. $600 GPU who cares?!

I don't think we'll get a full on 15 SMX part at first. I think that won't come until the end of summer. Just my guess.

[sontin]

Until GF100 nVidia always released the full chip. There is no reason to believe that they will change their behavior.