GPUs Set to Increase Performance by 570 Times by 2015 – Chief Executive of Nvidia.

[B.E.E.F.]

than the mad ramblings of Nvidia's crazy man.

That's true. But maybe he's got a big can of whoop ass to open.

[Heinz68]

Is the Nvidia Chief Scientist talking about the same TFLOPs performance?
Looks like he has some different ideas or he is not yet fully trained with NVIDIA public speaking. I know few people on this forum would do much better than him.

He is predicting only 20 TFLOPS performance by the year 2015. Maybe a conservative estimate since he is also predicting 11nm technology in 2015.
Source EETimes

By the way the NVIDIA 11nm process technology is one year sooner than the Intel Technology Outlook. I think Drwho? is not going to be happy about that. j/k

[saaya]

I don't think we had that ten years ago.

Graphics chip

GeForce 6800 Ultra 54 GFLOPs
ATI Radeon X850 XT 66 GFLOPs
NVIDIA GeForce 7800 GTX 165 GFLOPs
ATI Radeon X1900 426 GFLOPs
NVIDIA GeForce 8800 GTX 518 GFLOPs
NVIDIA GeForce 8800 Ultra 576 GFLOPs
ATI Radeon HD 2900 475 GFLOPs
NVIDIA GeForce 9800 GTX 648 GFLOPs
ATI Radeon HD 3870 496 GFLOPs
NVIDIA GeForce GTX 280 933GFLOPs
ATI Radeon HD 4870 1.2 TFLOPs

http://forum.beyond3d.com/showthread.php?t=51677

Since the 5870x2 if we go by the specs rumour specs of 4.2 t flops. thats close to a 100x increase in about 5 years.

If we go by the past, performance used to triple across each generation. If we get 3x performance every year for 6 years, we could get a 729x increase in performance.

If we go by theo's unreliable rumours of NV next chip being 3+ t flops, we are getting another tripling in performance in about a year and a half.

I hope this is true, because what it means is NV has come to the table compete again rather than live off its laurels, since they means a rapid development cycle.

those older numbers are rather questionable, like the guy who started that thread pointed out himself...

the only big boost in those numbers is nv7x to nv8x and from r420 to r520, basically when gpus became so programmable that gpgpu started to become interesting. comparing pre n8x and r520 flops is pointless imo...

2006-09 ATI Radeon X1900 426 GFLOPs
2007-03 ATI Radeon HD 2900 475 GFLOPs
2007-11 ATI Radeon HD 3870 496 GFLOPs
2008-07 ATI Radeon HD 4870 1.20 TFLOPs
-----------------------------------------------
2006-09 to 2008-07 = 22 months
426 to 1200 gflops = 2.8x improvement in around 2 years

2006-11 NVIDIA GeForce 8800 GTX 518 GFLOPs
2008-04 NVIDIA GeForce 9800 GTX 648 GFLOPs
2009-01 NVIDIA GeForce GTX 285 1.06 TFLOPs
----------------------------------------------
2006-11 to 2009-01 = 26 months
518 to 1060 gflops = 2.0x improvement in around 2 years

2 years is a good enough time to get a reliable impression of how the industry evolves... and we see a 2-2.8x boost... ever 2 years... not every single year...

lets say the overall big picture boost is 3x every 2 years, which is rather optimistic, 3x every 2 years is quite a boost... then lets say this is possible to maintain until 2015, then we would end up with:

2011-01 3180 gflops
2013-01 9540 gflops
2015-01 28620 gflops

quite an impressive number... but 28620 gflops is not 570 times as fast as a gtx285 in flops performance, its only 27x faster...

and your theory about nvidia somehow making their shader processors more complex and efficient... well sure, but that means more transistors which means less shader units per gpu, and lower clockspeeds... nvidia and ati have both been improving their shader processors and gpus in the past years, and they havent even managed to reach a 3x perf boost over 2 years. dont get me wrong, thats quite an impressive perf boost!

but thats nowhere near what they would have to do to get a 570x boost in 6 years... performance does not magically grow exponentially...

[XSAlliN]

Predictions are easy to do, you just open your mouth and words come out. Back in the 80's some well known scientists, predicted "people using flying cars by year 2000". What can I say, it's almost 2010 and people still use planes to fly. Let's not forget those silly predictions were some of us or suppose to be on some colonized planets by now. Yes, with passing of time, people advance "eventually" - unless they get extinct or something like that. Like the ice melting - which is beyond the prediction part.

[Mechromancer]

.....But can it run Crysis?

Seriously though, unless theyr'e using a new type of exotic circuit design technology like new materials, memristors, optical processors, etc., I see no way they can claim this type of speedup in a mere 6 years. Then again, tech is advancing very fast nowadays, so we'll see. I doubt we'll forget this statement in the next 6 years and we have something incredibly good, or incredibly disappointing, to look forward to in 2015. Let's revisit this in 6 years.

[Smartidiot89]

.....But can it run Crysis?

Seriously though, unless theyr'e using a new type of exotic circuit design technology like new materials, memristors, optical processors, etc., I see no way they can claim this type of speedup in a mere 6 years. Then again, tech is advancing very fast nowadays, so we'll see. I doubt we'll forget this statement in the next 6 years and we have something incredibly good, or incredibly disappointing, to look forward to in 2015. Let's revisit this in 6 years.

Agreed unless he meens 570 times the performance with new hardware and some handpicked well coded GPGPU app?

[randomizer]

I guess Dave House was wrong about performance doubling every 18 months. For NVIDIA it's actually 4.8x every 18 months!

[Manicdan]

Is the Nvidia Chief Scientist talking about the same TFLOPs performance?
Looks like he has some different ideas or he is not yet fully trained with NVIDIA public speaking. I know few people on this forum would do much better than him.

He is predicting only 20 TFLOPS performance by the year 2015. Maybe a conservative estimate since he is also predicting 11nm technology in 2015.
Source EETimes

By the way the NVIDIA 11nm process technology is one year sooner than the Intel Technology Outlook. I think Drwho? is not going to be happy about that. j/k

nice link, heres a snip

Throughput processors have hundreds of cores today and will have thousands of cores by 2015, Dally said. By then, Nvidia will have GPUs implemented on 11 nm process technology that feature roughly 5,000 cores and 20 teraflops of performance, Dally predicted.

[lookmomnobrains]

Moore's law states that the speed of chips dubles in about 2 years. So 64 times in 6 years is a more realistic target

[randomizer]

Moore's law states that the speed of chips dubles in about 2 years.

I don't think Moore's Law actually says anything about performance. That was Dave House, and it was 18 months.

[deathman20]

I don't think Moore's Law actually says anything about performance. That was Dave House, and it was 18 months.

Also thought that was dealing with CPU's not GPU's.

[Manicdan]

moores law is too simple. there are really only a few factors that go into chip advancements

the biggest is the process, if we go down a size every 2 years we can get ~100% from that. double check my math, example: (65nm*65nm)/(45nm*45nm) = 65nm is 2.08x larger than a 45nm chip

then we have IPC improvements, very minor jumps with each revision

and finally we have power consumption. back 15 years ago cpus were passively cooled, 10 years ago gpus only had a fan on the high end cards, and duel slot coolers were inexistent. now we have gpus that use 100+W per core, and people can throw 4 cores into a machine. while power consumption is simplified, the idea is that chips are getting bigger.

factor in all those things and its possible that we had 500x increase in the last few years. but to do the SAME THING (not to be confused with same increase) in the next 6 years means we would have to see gpus that use over 500W per core, and have 16 of them in our PC. if we dont get any bigger, then either we will need to have 80 gpus in a PC (basically 5x for not getting 5x more power hungry, and 4x for not getting 4x as many, and we already see computers with 4 gpus, for a total increase of 20x) or they better optimize those chips to be 20x more effective in how they work.

some one print this thread out and put it into a time capsule and take it out in 6 years

[randomizer]

Moore's Law itself applies to alot of things, even though it was originally talking specifically about ICs. According to Moore, this 18 month performance doubling that House observed was about transistor performance, rather than CPU performance (from memory anyway).

Moore's Law should really be called Moore's Observation. Nobody has to follow it, like you have to follow a law

[Dragy2k]

well that must mean that by that time they will have cracked dx11 for there card hence to performance jump.....

[B.E.E.F.]

I guess Dave House was wrong about performance doubling every 18 months. For NVIDIA it's actually 4.8x every 18 months!

... on paper. Applications don't run 4.8x faster.

[saaya]

Moore's law states that the speed of chips dubles in about 2 years. So 64 times in 6 years is a more realistic target

rofl.... you might want to read up on moores law...

... on paper. Applications don't run 4.8x faster.

i think he was being sarcastic...

570x boost in 6 years means almost a 3x boost per year in the next 6 years... why jensen mentioned this.. well it probably has to go with gt300 supposedly delivering 3x the theoretical max flops as a gtx285, and gt300 will come out around 1 year after gtx285... thats complete nonsense though as there is no way nvidia can keep up that pace for 6 years... the reason gt300 will launch around 1 year after gtx285 is because there hasnt been anything new for a long time, the 285 is a mere refresh, the cycle from one to the next architecture is at least 18 months if not more, not 1 year like jensen seems to assume in his pr math...

it easy to outdo yourself and beat your previous product by 300% in flops if your previous product is a refresh of a refresh of a much older part... its basically like sleeping in the office for one day and then doing your job the next and telling your boss you are doubling your productivity in only 1 day and predict you will be 6x as productive in 6 days as you have been today

but most analysts and investors are stupid enough to fall for this and will probably give jensen a thumbsup and shuffle more money into nvidia

[randomizer]

... on paper. Applications don't run 4.8x faster.

Paper is the only thing that matters. After all, that's what the marketing team shows us right?

i think he was being sarcastic...

Aye, this is true.

[Mungri]

Well, maybe if they use Ceramix Circupipe heatsinks, Ultra Durable PCBs, and MemOK VGA Ram they will actually manage to improve performance by 1000x by 2015.

[strange|ife]

now all we need are games.

being tied alongside the console market sucks. so few pc exclusives anymore

[strange|ife]

Looks like this CD will be back in style in 2015



Perkam

Quoted for Win

[Dude_Person]

There is no way he's talking about GPU performance. The slide compares "CPU-alone" against "CPU+GPU". The CPU GPU combination 6 years from now will be 570x more powerful than CPU's be themselves today (or so he claims). He's banking on CUDA/OpenCL/DirectX Compute adoption.

[eRacer]

Obviously x-bit got the story wrong. There is no projected 570x increase in GPU performance. It's more like 11.4x.

The photo in the article shows that the 570x comparison is between the combined power of a 2015 GPU and CPU versus today's CPU alone. The first equation is CPU performance growth alone, while the second is GPU+CPU.

According to the photo NVIDIA already considers today's GPU+CPU combo roughly 50x faster than a CPU alone, which is where the 50x multiplier comes from in the second equation. Dividing 570x by 50x gives 11.4. 1.5^6 = 11.4. So NVIDIA is claiming that 2015 CPU+GPU performance will be 11.4x that of now, increasing on average 50% each year, with almost all of that performance increase due to the GPU.

The GTX 295 currently does about 1.79TFlops with two cores. Perhaps the upcoming GT300 will do about the same with one core. Taking 1.79Tflops x 11.4x = 20.4 TFlops in 2015 which is almost identical to the 20TFlops target quoted by NVIDIA chief scientist William Dally.

[saaya]

the whole math starts by claiming cpu+gpu = 50% faster than cpu alone in 2009? no wonder the final result is nonsense...
even if its cpu+gpu for the 570x boost by 2015 its nonsense...

marketing people + maths = fail... always....

[Luka_Aveiro]

Obviously x-bit got the story wrong. There is no projected 570x increase in GPU performance. It's more like 11.4x.

The photo in the article shows that the 570x comparison is between the combined power of a 2015 GPU and CPU versus today's CPU alone. The first equation is CPU performance growth alone, while the second is GPU+CPU.

According to the photo NVIDIA already considers today's GPU+CPU combo roughly 50x faster than a CPU alone, which is where the 50x multiplier comes from in the second equation. Dividing 570x by 50x gives 11.4. 1.5^6 = 11.4. So NVIDIA is claiming that 2015 CPU+GPU performance will be 11.4x that of now, increasing on average 50% each year, with almost all of that performance increase due to the GPU.

The GTX 295 currently does about 1.79TFlops with two cores. Perhaps the upcoming GT300 will do about the same with one core. Taking 1.79Tflops x 11.4x = 20.4 TFlops in 2015 which is almost identical to the 20TFlops target quoted by NVIDIA chief scientist William Dally.

Now this makes sense.

[Chumbucket843]

What a move by Huang. way to be a party pooper at hot chips.

and to all of you guys preaching moores law.... http://en.wikipedia.org/wiki/Pollack's_Rule