Simulated Nvidia Kepler GK110 Performance Preview

[Don Karnage]

Source :

http://videocardz.com/33208/nvidia-kepler-gk110-performance-preview

OCAHOLIC.ch came up with some interesting review. Review of the card that doesn’t exist, well at least it’s not released yet. NVIDIA’s GK110 is the newest 7.1 billion transistor GPU, which will power upcoming Tesla K20 graphics card for High Performance Computing (HPC).

First rumors said that GK110 would be actually GeForce GPU, but NVIDIA decided to make some shifting to the naming and therefore move this GPU to other date. During GTC 2012 they presented this new chip as upcoming HPC unit for Tesla Series. But many of us still hope that this chip will eventually get into GeForce family

 

[njdevilsfan87]

That looks way too good to be true, but if it's anywhere near that performance, I expect a minimum $800 price tag...

 

[ShintaiDK]

I have a feeling its fake.

Logically even from the specs they write it gotta be fake. 2880Sps vs 1536sps, 48 vs 32 ROPS, 384bit vs 256bit, 240 TMUs vs 128TMUs. Everything at same clockspeeds. Those charts aint gonna happen with 100%+ speedboosts.

 

[RussianSensation]

Don, I think your title is misleading. That's not a real performance preview. It's a "simulated" performance preview: i.e., the authors have used the specifications and extrapolated / made a prediction on how they think GK110 would perform.

This can be very difficult to do since we don't know if 32 ROPs or the memory bandwidth are holding back GK104 the most. Also, in games such as Batman AC, Crysis 2, tessellation performance is a huge bottleneck. While in games such as Witcher 2 and Shogun 2, memory bandwidth and strong AA performance seem to be more important.

Some of their simulated graphs show performance nearly doubling over the GTX680. That sounds impossible to believe unless GK110 has 64 ROPs, 3072 SPs, 256 TMus, etc.

I think we can expect at least 40% faster but after that it becomes a game of darts.

 

[Don Karnage]

Don, I think your title is misleading. That's not a real performance preview. It's a "simulated" performance preview: i.e., the authors have used the specifications and extrapolated / made a prediction about how they think GK110 would perform.

This can be very difficult to do since we don't know if 32 ROPs or the memory bandwidth are holding back GK104 the most. Also, in games such as Batman AC, Crysis 2, tessellation performance is a huge bottleneck. While in games such as Witcher 2 and Shogun 2, memory bandwidth and strong AA performance seem to be more important.

Some of their simulated graphs show performance nearly doubling over the GTX680. That sounds impossible to believe unless GK110 has 64 ROPs, 3072 SPs, etc.

I think we can expect at least 40% faster but after that it becomes a game of darts.

Agreed RS. I changed the title

 

[tviceman]

I think the gap between the best GF110 geforce product and the gtx680 will be larger than the gap between the gtx580 and gtx560ti. TDP will likely be the biggest factor holding GK110 back.

 

[Don Karnage]

I think the gap between the best GF110 geforce product and the gtx680 will be larger than the gap between the gtx580 and gtx560ti. TDP will likely be the biggest factor holding GK110 back.

Easily solved. They'll all come watercooled

 

[blastingcap]

Performance rarely scales linearly. (E.g., the 5870 was a literal doubling of the HD4890 except adding DX11 performance, and it was what, 40% faster or something like that? DX11 slowed things down a little so it's not apples-to-apples but still, even if it were 50% faster that's nowhere near double.) Even if GK110 wound up being twice as fast as GTX680, expect to pay 3-4 times as much for 2x the performance, if it's available at all. K20 is going to soak up most early GK110 GPUs.

 

[Genx87]

It wont cost 3-4x a GTX 680.

 

[Smartazz]

348bit bus width eh? Regardless, seems like a reasonable hypothetical performance analysis. Time will tell though.

 

[RussianSensation]

I think the gap between the best GF110 geforce product and the gtx680 will be larger than the gap between the gtx580 and gtx560ti. TDP will likely be the biggest factor holding GK110 back.

Possibly. Then you are predicting GK110 to be >50% faster than GTX680? That would be more nearly 60% faster than an HD7970. That would give HD8000 a real headache. 2880 SP, 240 TMU GK110 coming out within 12 months from now sounds almost too good to be true. I could see something like a 2304 SP version.

 

[blastingcap]

It wont cost 3-4x a GTX 680.

How do you know what the street price of a GTX 680 is next year (because that's how long it'll take to filter down to GeForce)? If it's $250 then I can see GK110 priced at $750-1000, admittedly closer to the 750 than the 1000.

 

[RussianSensation]

Performance rarely scales linearly. (E.g., the 5870 was a literal doubling of the HD4890 except adding DX11 performance, and it was what, 40% faster or something like that? DX11 slowed things down a little so it's not apples-to-apples but still, even if it were 50% faster that's nowhere near double.)

I think HD5870 is about 70% faster on average than an HD4890 if looking at newer + older games and more like 80-90% faster in newer games. A lot of people probably remember HD5870 testing from TPU where it was 40% faster but that's because so many games in that review were so old (dating back to 2006-2008) and CPU bottlenecked. HD5870 launched in 2009 and since then a lot of newer games have come out in which HD4800 series struggled. Most people probably just never bothered revisiting newer reviews and that 40% faster number was stuck in their head.

As an estimate take off 15% off 6970 to get to 5870 in say Diablo 3:

If you were to redo HD5870 vs. HD4890 tests in games in the last 2-3 years, it would be more like 75-90% faster (of course except in games where tessellation hammers HD5870 such as Batman AC).

P.S. HD4000 series sucks in Diablo 3!

 

[blastingcap]

Actually HD5870 is about 80% faster than an HD4890. A lot of people probably remember HD5870 testing from TPU where it was 40% faster but that's because so many games at the time were so old (dating back to 2006-2008). HD5870 launched in 2009 and since then a lot of newer games have come out in which HD4800 series struggles.

As an estimate take off 15% off 6970 to get to 5870 in say Diablo 3:

If you were to redo HD5870 vs. HD4890 tests in games in the last 2-3 years, it would be more like 75-90% faster (of course except in games where tessellation hammers HD5870 such as Batman AC).

P.S. HD4000 series sucks in Diablo 3!

Fine say it's 80%. You can see SLI scaling in that area averaged over a broad range of games. The bottleneck may not be in the SMX's though as demonstrated by the GTX 680 vs 670.

 

[RussianSensation]

Fine say it's 80%. You can see SLI scaling in that area averaged over a broad range of games. The bottleneck may not be in the SMX's though as demonstrated by the GTX 680 vs 670.

I agree. The graphs are pure speculation. Since GTX680 is what like 30-35% faster than a GTX580, that preview expects GK110 to be 1.6-1.8x faster than a GTX680? That sounds insanely optimistic or GTX680 would have a price of $299 next year....

 

[LOL_Wut_Axel]

FFFAAAAAKKKKEEEE!

Not much else to add, because even going by their specs it wouldn't be anywhere near 50% faster. More like 25% faster, and even then these specs scream fake.

And the gap between GK104 and GK110 should be smaller than GF114 and GF110 because of the fact that GK104 is an all-out gaming GPU. On the other hand, GF114 didn't give up too much of its compute performance.

 

[Grooveriding]

So someone made up a prediction and drew a graph with their imagination ?

hehe :sneaky:

This is worse than the predictions around that the 294mm2 die GTX680 would be 60% faster than the GTX 580, oh, and would cost only $300.

I have a 8970 vs 780 performance preview ready though

8970 >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
780 >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>

Someone make a thread.

 

[Phynaz]

How do you review something that doesn't exist? And simulation software is very expensive, which is the only way to really do this.

Anything for page hits I guess.

 

[guskline]

Don: First Kudos for disclosing the source and accurately titling this thread. I actually read (didn't say I could understand it all) the white paper from Nvidia on the GK110. Suffice it to say that toward the end of the paper they say the following:

"Kepler GK110 GPUs will be used in numerous systems ranging from workstations to supercomputers to address the most daunting challenges in HPC."​

I don't think we will see this GPU in the consumer market anytime soon.

 

[princess12]

Hi I am just new to this thread. It is really helpful. Indeed a brilliant explanation at that. Thanks for helping.

 

[Olikan]

This is worse than the predictions around that the 294mm2 die GTX680 would be 60% faster than the GTX 580, oh, and would cost only $300.

still it better than the 600mm² 7970...

 

[tviceman]

Possibly. Then you are predicting GK110 to be >50% faster than GTX680? That would be more nearly 60% faster than an HD7970. That would give HD8000 a real headache. 2880 SP, 240 TMU GK110 coming out within 12 months from now sounds almost too good to be true. I could see something like a 2304 SP version.

A 2304 core part might come first, but I am referring to the "best" GK110-based geforce product that we will see with GK110. I think 60% is going to be the average over what GK104 currently does. Remember, GK104 can always be refreshed to run a little faster (needs faster vram!!!) just like Nvidia did with GF104.

But when comparing the differences with GF114 and GF110 with GK104 and GK110, GK110 has a higher percentage of cores AND also has a larger gap in die size. So yeah, I think GK110's best geforce card (whether it comes first or comes sometime down the road) will be ~50-60% faster than GK104, and will cost $699 (I think there will be 3 total GK110 geforce products, $649-699, $549-599, and $449-499).

 

[blackened23]

So someone made up a prediction and drew a graph with their imagination ?

hehe :sneaky:

This is worse than the predictions around that the 294mm2 die GTX680 would be 60% faster than the GTX 580, oh, and would cost only $300.

I have a 8970 vs 780 performance preview ready though

8970 >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
780 >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>

Someone make a thread.

You can never be too early to the internet with fake performance slides, GF100 is a testament to it. Doesn't matter if its a year out...crysis 3 is a year out and you can pre-order it.

Besides, anyone who actually reads the article they state plainly that the result isn't real , its hypothetical. Just goes to show that 99% of readers skim the title and/or conclusion and never read the body of articles.

 

[Genx87]

How do you know what the street price of a GTX 680 is next year (because that's how long it'll take to filter down to GeForce)? If it's $250 then I can see GK110 priced at $750-1000, admittedly closer to the 750 than the 1000.

Right, the GTX 680 is going to drop by half its price. Has the 580 dropped its price by half in 20 months? It will be discontinued before it is sold en mass at 250 dollars.

When this is released as the top card I expect it to run 500-600 bucks as it replaces the 680 GTX.

 

[ShadowOfMyself]

Was this chart made by Wreckage or something?

Over twice as fast? Bahaha! It will barely be 50% faster, especially because as someone else mentioned, the GTX680 is only as fast as it is because it is basically a "handicapped" version of Kepler focused purely on gaming performance

Either way, its such a long way off that its not even worth speculating about (and the same for the 8000 series)