NVIDIA Kepler speculation thread

[mczak]

Hmm, I can't "lock" on any reference measure in this one. Anyone have the dimensions of those DDR3 chips?

Some hynix 2gbit ddr3 chip (mentioned in the "graphic card memory" category) is 13.0x9.0mm. Can't tell you though if they have standardized physical dimensions.

edit: with that ddr3 chip size I get about 130-140mm² for gk107, so slightly bigger than CV.

 

[ninelven]

And what about the possibility that the big guy's more optimized for HPC workloads, owing his bigness to big 52 bit multipliers (or two 32 bit multipliers and corresponding carrier networks) and according datapaths and register files throughout the chip?

Doesn't really matter... people have always paid a premium for highest performance available regardless of the price/performance ratio.

 

[DarthShader]

it kinda works out, 50% more raw shading power at the same base clock as 580, so if its clocked a bit higher 45-50% better perf in 3dmark seems possible.

... at the same power draw as GTX580. And we know it's going to be much lower (actualy, do we really know that?).

So for example, if we guess the TDP of the GK104 card to be 180W, and agree the real TDP of a GTX580 is 270W, that is exactly 50% more. So plus/minues the inaccuracies of this very simplistic analysis, the performance should be around/slightly higher than a GTX580, as predicted earlier. Or did I miss something?

 

[CarstenS]

Doesn't really matter... people have always paid a premium for highest performance available regardless of the price/performance ratio.

That would not be the problem.

 

[ninelven]

Well the only problem I can see is if you are suggesting that GK110 would actually be lower performance than GK104 in gaming scenarios. Is that what you are saying? Or to put it another way, then what exactly is the problem?

 

[mczak]

Looks like 50% more performance over a 550M @ same 1.8Gbps DDR3 in the SM3.0 of 3DM06:
http://i51.tinypic.com/nq44li.jpg
http://www.cooltoyzph.com/viewitem.php?iid=632

Hmm 30% more for SM2.0 and 50% more for SM3.0 doesn't exactly sound stellar for a card which has 110% peak shader flops (as the default clock for 550M is 740Mhz). The score is nowhere close to a HD7750.
Granted though that's an unfair comparison, I'm willing to believe this card is probably (a lot more so than GF108 was) very bandwidth limited with ddr3 (no idea how Cape Verde with ddr3 would perform - probably not so good neither...).

 

[trinibwoy]

The score is nowhere close to a HD7750.

Shouldn't it be compared to the 7750M?

 

[AlphaWolf]

Shouldn't it be compared to the 7750M?

Dunno what the 7750 will be called in a laptop, but it runs low enough power that it could well find itself in laptops without any changes. The 6990m is already a faster part.

 

[mczak]

Shouldn't it be compared to the 7750M?

Sure. You've got some benchmark scores for it?

Dunno what the 7750 will be called in a laptop, but it runs low enough power that it could well find itself in laptops without any changes. The 6990m is already a faster part.

Yes, but the part performing very close to 7750 will probably be called 7870M or so.
A hypothetical 7750m based on CV would likely be downcut/downclocked a lot (or come with ddr3 memory too), making this a more fair comparison.
In any case gk107 with ddr3 just doesn't tell us much about how fast the chip really is.

 

[trinibwoy]

Sure. You've got some benchmark scores for it?

Nope just clarifying where we expect performance to fall

 

[AlphaWolf]

Yes, but the part performing very close to 7750 will probably be called 7870M or so.
A hypothetical 7750m based on CV would likely be downcut/downclocked a lot (or come with ddr3 memory too), making this a more fair comparison.
In any case gk107 with ddr3 just doesn't tell us much about how fast the chip really is.

I expect you'll see it used in a variety of ways (OEMs love to cripple stuff with cheap memory), however, as is it would be a good part for a performance laptop (with 256bit parts based on pitcairn in the enthusiast segment). But yes a 7750m should still be CV, unless they pull out the re-brand.

 

[Picao84]

I'm guessing this is GK100/110's PCB, with the 6+6+8, 7 phase

edit: for compare
Some early GF100 boards weren't exactly the same as their design diagrams

jaredpace, just to be sure: The diagram pic.. is it a new pic? from a possible GK100/GK110 board? Or is it an early diagram from a GF100/GF110 board?

Thanks

 

[jaredpace]

No those are old GF100

 

[Kaotik]

Which Tahiti exactly for one and directly after that since when does buswidth by itself define a GPU's performance? What a surprise Cayman had a 256bit bus yet was on average about 15% behind a GF110 with a 384bit bus.

***edit: and just for the "rumor's" sake I count 8 memory chips on that picture above.

The real card has 8 yes, the PCB image had 12 slots didn't it (yes, now i'm really drunk)

Cayman vs GF110 was different - Cayman had high clocked memories vs GF110's low clocked ones, GTX580 had only 9.3% higher mem bandwidth than 6970, but Tahitis memories are already quite high clocked so GK104 can't have similar clock advantage to make up for the narrower membus

 

[CarstenS]

Well the only problem I can see is if you are suggesting that GK110 would actually be lower performance than GK104 in gaming scenarios. Is that what you are saying? Or to put it another way, then what exactly is the problem?

That there is a - however remote - possibility, that the changes from GK104 to GK110 won't be benefitting gamers at all. For example, and I'm purely making this up mind you, Nvidia could have opted to stay with the design, clocks, # of functional units but only improve double precision throughput from, say one fourth (physically, Geforce products would be throttled down to 1/12th I expect) to one-half or - heavens forbid - even full rate and only attach a wider memory bus in order to enable larger data set buffers on professional products.

edit:
Just to make it even more clear: I'm not saying that this is what GK110 will turn out to be, but depending on business analysis and competitive analysis outcomes at Nvidia they could have become convinced that the focus on HPC markets with GF100/b (half-rate DP, fine grained work scheduling, large dataset buffers and less texturing power) was the right decision and that they should continue in that direction. They could have come to the conclusion that the professional market would be growing large enough to justify it's very own ASIC from which gamers would only profit in the way of wider memory busses.

 

[AnarchX]

Some hynix 2gbit ddr3 chip (mentioned in the "graphic card memory" category) is 13.0x9.0mm. Can't tell you though if they have standardized physical dimensions.

edit: with that ddr3 chip size I get about 130-140mm² for gk107, so slightly bigger than CV.

Samsung 2gbit ddr3 chips are ~13.3x7.5mm
.

Seems to be a comparable size to CV. Let's see what it can do with ~1GHz GPU clock and >4Gbps GDDR5 in >2010 benchmarks.

 

[fellix]

I got ~132mm² with perspective compensated measurement.

 

[chavvdarrr]

Then the chip is 110-120 ?
Are CV sizes from external measurments, or "official" AMD data?

 

[mczak]

Samsung 2gbit ddr3 chips are ~13.3x7.5mm
.

Oh so the ddr3 chip size isn't quite the same for everybody. So are those samsung or hynix chips (or something else)? The width/height ratio of the chips doesn't quite seem to match neither, I guess too much perspective distortion.

Seems to be a comparable size to CV. Let's see what it can do with ~1GHz GPU clock and >4Gbps GDDR5 in >2010 benchmarks.

Yes, if that memory chip is 13.5x7.5mm it should probably be slightly smaller than Cape Verde, otherwise slightly larger, in any case quite close. And performance could be quite close too.

 

[CarstenS]

Are CV sizes from external measurments, or "official" AMD data?

AMDs official CV size is 123 mm².