Big Kepler news

[3DVagabond]

http://www.legitreviews.com/news/12803/

TBH I'm disappointed if it's true. I was expecting more than 20-25% faster than the GTX 680.

Look at Pitcairn and Tahiti. GK104 to GK110 is a similar design change. The big cards aren't going to be a "gaming efficient".

What happens if Nvidia drops the price of the 680 to 399.99 and sells the 685 for 499.99?

That looks more like wishful thinking.

Actually it depends on quantities. They will fulfill the Tesla market first, and then what's left will get sold for whatever the market will bear. We need for the rest of the Keplers to roll out. I don't think one card, the 680, is enough to bring true market competition.

 

[tviceman]

yeah that makes no sense at all. that would have to be the poorest scaling architecture in history if true.

really I do not see the point in going to 512 bit bus with the 384 bit bus would probably get the job done.

384 probably would get the job done, if the die was meant primarily for gaming. But that is not the case.

 

[tviceman]

No, more like $50 extra to own a "next gen" card since December instead of waiting to "try" and get one in April.

I am not AMD fan (never owned one), but have to give AMD credit for beating Nvidia out of the gate twice now (fermi anf kepler).

AMD absolutely destroyed Nvidia with the 40nm debut of GPU's, but this time around a less than 3 month head start with an overpriced lineup, in what will be a 2 year fight with 28nm products, is really not that impressive.

Kudos for being first, but when your best die is generally slower, less efficient in power and performance per mm^2, and priced higher on top of all that... well... you get the picture.

 

[soccerballtux]

AMD fans have been waiting for years to make that claim

very true lol

 

[3DVagabond]

Page after page of 7970 vs 680? People should just go out and buy whatever card they like better and stop this endless regurgitation.

Back to the topic, if possible. Keep in mind when expecting higher performance there is a TDP wall. While many/most gamers don't care how much power their card uses, it is important in the professional field. Also considering there is apparent issues with fabbing(?) this chip it might end up clock limited.

 

[lavaheadache]

7970 has 60% more cores and ~70% more mem bandwidth than 7870, but is only ~30-35% faster on average.
nV cranked the clocks on 680 high to beat 7970...

This

 

[skipsneeky2]

Figured as much that a gtx685 would be coming,might interest me if it retails for about $499,then again i love my 7970 but wow the drivers are terrible enough to make me sell this thing and switch over to the gtx685 when it debutes.

Just installed 12.3 today,didn't come with the issues 12.2 had so hope they are stable enough to keep me from making a impulse purchase come q3 2012,thus far so good.

 

[Don Karnage]

Figured as much that a gtx685 would be coming,might interest me if it retails for about $499,then again i love my 7970 but wow the drivers are terrible enough to make me sell this thing and switch over to the gtx685 when it debutes.

Just installed 12.3 today,didn't come with the issues 12.2 had so hope they are stable enough to keep me from making a impulse purchase come q3 2012,thus far so good.

Always a problem with amd. Their driver team is atrocious

 

[Cookie Monster]

Tbh, we've yet to know why compute performance on the GK104 is lacking. For all we know, it could be the lack of software/compiler optimizations.

 

[Vesku]

https://bitcointalk.org/index.php?topic=73681.20

If there is unlocked performance than Nvidia really needs to release a Kepler OpenCL/Cuda guide.

 

[Ajay]

Bigger leap of cores?
384 -> 512
1536 -> 2048
Sure, the number is bigger, but it's still the exact same percentage increase, 33%.

2304 for bigK, so a 50% increase in cores - according to the OP.

 

[BallaTheFeared]

If it's 2304 then nvidia dumped GF100/110.

I guess it's possible, but GF104 was never compute orientated, GK104 thus far is even less compute based than it's predecessor.

You guys really think Nvidia is going to simply tack on two more SMs to GK104 and add in a ton of compute? I mean this is their bread and butter market... And as of yet, it is still Fermi driven as Kepler in it's current form is incapable of exceeding what GF100 is.

http://www.nvidia.com/object/fermi_architecture.html

 

[toyota]

If it's 2304 then nvidia dumped GF100/110.

I guess it's possible, but GF104 was never compute orientated, GK104 thus far is even less compute based than it's predecessor.

You guys really think Nvidia is going to simply tack on two more SMs to GK104 and add in a ton of compute? I mean this is their bread and butter market... And as of yet, it is still Fermi driven as Kepler in it's current form is incapable of exceeding what GF100 is.

http://www.nvidia.com/object/fermi_architecture.html

so you are saying that it should be more than 2304 sp? I don't think you realize how much more gpu that is than the gk104. heck gf110 had only had 33% more sp, 50% larger bus and the same number of tmus as gf114. the gk110 is a massive leap in specs over the gk104 in comparison.

 

[BallaTheFeared]

No I'm saying if it's a GF100 derived part, than assuming it follows the same progression as GF104, than this part should be 512x4 = 2048

Starting with 512, double it because of a loss of hot clocks = 1024 = GTX 580 double again for the generational increase just like GK104 = 2048

2048 = GF100 doubled, 2304 = something else...


2304 is... doing math

GK104 = 192 CUDA cores per SM, so GK104 has 8 SM units. 2304 would need 12 SM units, which is actually 4 more SM units instead of what I said before which I think was 2.

I mean it's totally possible Nvidia would use the GK104 design, I don't know enough about the differences in GF104 vs GF100 to say unequivocally that they couldn't use it for compute.

Just spitballing.

 

[bunnyfubbles]

No I'm saying if it's a GF100 derived part, than assuming it follows the same progression as GF104, than this part should be 512x4 = 2048

Starting with 512, double it because of a loss of hot clocks = 1024 = GTX 580 double again for the generational increase just like GK104 = 2048

2048 = GF100 doubled, 2304 = something else...


2304 is... doing math

GK104 = 192 CUDA cores per SM, so GK104 has 8 SM units. 2304 would need 12 SM units, which is actually 4 more SM units instead of what I said before which I think was 2.

I mean it's totally possible Nvidia would use the GK104 design, I don't know enough about the differences in GF104 vs GF100 to say unequivocally that they couldn't use it for compute.

Just spitballing.

I was thinking the same thing

while GK104 is clearly based off of GF114, 2304 shaders suggests a complete departure from GF100/110

 

[Keysplayr]

No I'm saying if it's a GF100 derived part, than assuming it follows the same progression as GF104, than this part should be 512x4 = 2048

Starting with 512, double it because of a loss of hot clocks = 1024 = GTX 580 double again for the generational increase just like GK104 = 2048

2048 = GF100 doubled, 2304 = something else...


2304 is... doing math

GK104 = 192 CUDA cores per SM, so GK104 has 8 SM units. 2304 would need 12 SM units, which is actually 4 more SM units instead of what I said before which I think was 2.

I mean it's totally possible Nvidia would use the GK104 design, I don't know enough about the differences in GF104 vs GF100 to say unequivocally that they couldn't use it for compute.

Just spitballing.

GF104 fully unlocked was 384 cores, not 512. 2304 is divisible by 384 evenly.
384 x 6 = 2304.

The progression you described above in your first sentence would be GF100/GF110.
So in your second sentence, you should be starting with 384 instead of 512.

Then everything else falls into place.

 

[BallaTheFeared]

When I said that I meant GF100 would follow the same progression that GF104 did, becoming GK104 with the quadrupling of CUDA cores.

I hope it makes more sense now with what I wrote, I'm assuming BigK would be a derivative of GF100, not GF104 as GK104 is. However the same basic principles that governed how we arrived at GK104 from GF104 would be applied to GF100.

Ugh it's not coming out as smoothly as I had hoped... Does it make more sense to you now though as far as what I'm trying to convey here?

 

[Keysplayr]

When I said that I meant GF100 would follow the same progression that GF104 did, becoming GK104 with the quadrupling of CUDA cores.

I hope it makes more sense now with what I wrote, I'm assuming BigK would be a derivative of GF100, not GF104 as GK104 is. However the same basic principles that governed how we arrived at GK104 from GF104 would be applied to GF100.

Ugh it's not coming out as smoothly as I had hoped... Does it make more sense to you now though as far as what I'm trying to convey here?

I understand. But you're looking at this as just a simple die shrink if you think GF104/GF114 or GF100/110 has anything in common with GK104/GK110. It's a radically changed architecture and you are trying to compare the progression based on shader count per SM. Nvidia changes this often. The new current "SMX" has 192 shaders and the last gens had what? 32? 48?.
It wouldn't surprise me in the least either way if the Kepler design is a modular one where they can more easily add or remove more SMXs to or from a chip design.
I expect GK107 to have at most 768 shaders or 4 SMXs.
What you suggest is always a possibility though. I never write anything off completely.

 

[BallaTheFeared]

I'm still looking at Kepler as a revision of Fermi, which is still "loosely" based on the G80 uArch.

 

[Keysplayr]

I'm still looking at Kepler as a revision of Fermi, which is still "loosely" based on the G80 uArch.

Almost every new gen has ingredients of the older gens. Nothing is set in stone, however.

 

[BallaTheFeared]

Yep yep, I'm just armchair engineering over here I figured it would be good for the discussion.

From the Anand review:

Big Kepler may not end up resembling GK104, but if it does then it may be an extremely potent FP64 processor if it’s built out of CUDA FP64 blocks..

They're talking about compute here (I believe), not graphical performance for those that aren't sure what I'm trying to use this quote for.

Speculation, the next best thing to actually owning it!

 

[Don Karnage]

Yep yep, I'm just armchair engineering over here I figured it would be good for the discussion.

From the Anand review:



They're talking about compute here (I believe), not graphical performance for those that aren't sure what I'm trying to use this quote for.

Speculation, the next best thing to actually owning it!

Big Kepler will be better in compute? But I loved hearing people whine about how the 7979 was so much better in it.

 

[aaksheytalwar]

Big Kepler will be better in compute? But I loved hearing people whine about how the 7979 was so much better in it.

Are you half brain dead?

If nvidia can make a card 9 months after the 7970 which is like 20-40% better in compute and will replaced within 3-4 months by a card from AMD which will thrash the big kepler, then I consider AMD the winner here, not nvidia. Can't be simpler than that.

If you didn't get this, count the number of months for which AMD won and the number of months for which nvidia may win and then again the number of months for which AMD will begin again. Even if both tie in the end, the one who started winning first is the winner in that special case of tie

 

[BallaTheFeared]

AMD isn't even a factor in compute, Nvidia doesn't care about bitcoins.. lol they're making movies and do scientific HPC.

I was specifically talking about what we have now (GK104 GTX 680) vs what could possibly happen with BigK.

Here is the entire quote from the Anand review.

The other change coming from GF114 is the mysterious block #15, the CUDA FP64 block. In order to conserve die space while still offering FP64 capabilities on GF114, NVIDIA only made one of the three CUDA core blocks FP64 capable. In turn that block of CUDA cores could execute FP64 instructions at a rate of ¼ FP32 performance, which gave the SM a total FP64 throughput rate of 1/12th FP32. In GK104 none of the regular CUDA core blocks are FP64 capable; in its place we have what we’re calling the CUDA FP64 block. The CUDA FP64 block contains 8 special CUDA cores that are not part of the general CUDA core count and are not in any of NVIDIA’s diagrams. These CUDA cores can only do and are only used for FP64 math. What's more, the CUDA FP64 block has a very special execution rate: 1/1 FP32. With only 8 CUDA cores in this block it takes NVIDIA 4 cycles to execute a whole warp, but each quarter of the warp is done at full speed as opposed to ½, ¼, or any other fractional speed that previous architectures have operated at. Altogether GK104’s FP64 performance is very low at only 1/24 FP32 (1/6 * ¼), but the mere existence of the CUDA FP64 block is quite interesting because it’s the very first time we’ve seen 1/1 FP32 execution speed. Big Kepler may not end up resembling GK104, but if it does then it may be an extremely potent FP64 processor if it’s built out of CUDA FP64 blocks.

 

[Vesku]

I don't believe the kind of installations that use GPGPU are going to want to put their installations on hold until the Fall. The people in charge of them will be trying very hard to get an inkling of when the next Nvidia compute product will show up. If it's a ways out then AMD's GCN arch will be mighty tempting for at least some of them.

I bet behind closed doors Nvidia is promising the 28nm compute card will be glorious and worth waiting for.