Nvidia CEO predicts GPU performance to increase 570x by 2015

[Megadrive1988]

http://www.tomshardware.com/news/Nvi...570x,8544.html
http://www.tgdaily.com/content/view/43745/135/
http://blogs.nvidia.com/nTersect/

Reminds me of something from a few years ago,
Intel's 'Platform 2015' projections

 

[CNCAddict]

How much faster have they become over the past 6 years?? With a failing economy and physical limits hitting harder than ever...it's difficult to see the next 6 years outpacing the past 6.

I think that puts us back to the NV30

http://www.tomshardware.com/picturestory/464-12-nvidia-graphics-cards.html
http://forum.beyond3d.com/showpost.php?p=1256412&postcount=14

Which looks like about 31 GFLOPS which is not even close to 570x slower then what we have now. More like 33x no??

 

[ShaidarHaran]

He's talking about GPU compute. Totally different than rendering capability.

 

[Davros]

whats this 1.2 to the power of 6 thing
does he mean cores will be 1.2 times faster and there will be 6 of them ?

ps: remember chameleon mark (still available on nzone) it was a benchmark for the shaders in the first card to have them the gf3 iirc that card scored around 50fps on a 260 i get about 1500

 

[prunedtree]

whats this 1.2 to the power of 6 thing
does he mean cores will be 1.2 times faster and there will be 6 of them ?

My guess would be that the slide is:

CPU-Alone  1.2^6        3X
CPU+GPU    50 * 1.5^6   570X

My interpretation is that Huang compares future speedups vs a current CPU, single core. Which is why GPU gets a 50x advantage (that's today, one core vs GTX295 in the best ideal case). Then he predicts 20% of single-core performance scaling per year for CPUs, for the next six years (hence 1.2^6) while GPUs would get 50% faster per year.

Which would mean that he predicts `GPU compute' will be 570x faster than using a single core on a CPU in 6 years, which is a sound argument...

... If you believe that no mortal can make a CUDA kernel scale on multiple CPU cores.

 

[neliz]

I'll post this as I've posted somewhere else..


nV engineers have a LOT of work to do:

G80 -> G92 -> GT200
8800Ultra-9800GTX+-GTX285
in transistors:
681M - 754M - 1400M
in GFlops:
576 - 705 - 1063

In Comparison:

R580=426GFlops
R600=614GFlops
RV670=416Gflops
RV770=1200GFlops

 

[aaronspink]

Which looks like about 31 GFLOPS which is not even close to 570x slower then what we have now. More like 33x no??

you forgot to put the magical 50x in front. With that it is 1650x slower! See it's all about the magical 50x. Everything looks better with that!

On a more realistic note, my prediction: GPUs in 2015 will be 8-16x faster than today. Most of the 33x vs the NV30 is because of increases in thermal and area budgets of the GPUs which are now effectively at their limit.

In all honesty, 2x for gpus every major process node will likely be extremely hard to achieve due to the effective reduction in power scaling in process technology.

 

[rpg.314]

Which would mean that he predicts `GPU compute' will be 570x faster than using a single core on a CPU in 6 years, which is a sound argument...

... If you believe that no mortal can make a CUDA kernel scale on multiple CPU cores.

Seems about the only way I can agree with jhh's statement.

 

[Davros]

while GPUs would get 50% faster per year.

but the 50 refers to cpu no? + 1.6 to gpu
as it says CPU+GPU not GPU+CPU

 

[no-X]

In Comparison:

R580=426GFlops
R600=614GFlops
RV670=416Gflops
RV770=1200GFlops

Are you sure about those numbers?

R580: PS: 650MHz, 48 vec3+1 MADD + 48 vec3+1 ADD = 650*(48*4*2+48*4) = 374.4 GFLOPS
VS: 650MHz, 8 5D MADD = 52 GFLOPS
R600: US: 740MHz, 320 MADD = 740*320*2 = 473.6 GFLOPS
RV670: US: 775MHz, 320 MADD = 775*320*2 = 496 GFLOPS
RV770: US: 750MHz, 800 MADD = 750*800*2 = 1200 GFLOPS

 

[CarstenS]

nV engineers have a LOT of work to do:
G80 -> G92 -> GT200
8800Ultra-9800GTX+-GTX285
in transistors:
681M - 754M - 1400M
in GFlops:
576 - 705 - 1063

In Comparison:

R580=426GFlops
R600=614GFlops
RV670=416Gflops
RV770=1200GFlops

Is this pipeline-counting 2009-style?

 

[rpg.314]

but the 50 refers to cpu no? + 1.6 to gpu
as it says CPU+GPU not GPU+CPU

I think 50 refers to today's speedup of gpu vs a single cpu core.

 

[neliz]

Are you sure about those numbers?

R580: PS: 650MHz, 48 vec3+1 MADD + 48 vec3+1 ADD = 650*(48*4*2+48*4) = 374.4 GFLOPS
VS: 650MHz, 8 5D MADD = 52 GFLOPS
R600: US: 740MHz, 320 MADD = 740*320*2 = 473.6 GFLOPS
RV670: US: 775MHz, 320 MADD = 775*320*2 = 496 GFLOPS
RV770: US: 750MHz, 800 MADD = 750*800*2 = 1200 GFLOPS

I took them from a post here that does R580/R600/RV670 comparisons, maybe overlooked something like clockspeeds.

 

[ChrisRay]

Is this pipeline-counting 2009-style?

Hehe to be fair. GFLOPS have still expanded at an alarming rate over the past couple of years. This "570x" number is really hard to quantify right now.

 

[CNCAddict]

I'm confused about something though. CPUs should now be scaling as well as GPUs in performance since they are basically doing the same thing, namely tacking on more and more cores. I believe AMD is aiming for 16 cores in 2 years.

 

[rpg.314]

I'm confused about something though. CPUs should now be scaling as well as GPUs in performance since they are basically doing the same thing, namely tacking on more and more cores. I believe AMD is aiming for 16 cores in 2 years.

Yeah, CPUs are scaling in core count too. But if you want to make your point to tech-averse journalists, there's nothing better than comparing performance of a single cpu core with the entire gpu's performance. :smile:

 

[elsence]

Maybe Huang is taking for CPU scaling (1,2X per year), the scaling we had (general performance) from Q4 2006 with a high end CPU like Intel Core 2 Extreme QX6700 (quad 2,66GHz) to a future Q4 2009 Intel CPU (like a high end quad Nehalem)

Just kidding!

the 1,5X scaling per year for GPU is a consistent number with what NV said in the past (2015=20 TFLOPs) (GPU=General Computing on GPU=TFLOPs indicative measurement)

the 50X part maybe it refers to the difference a future Q4 2009 GPU like GT300 (512SP?, with around 2,5 TFLOPs?) with a high end CPU (QX9770 was around 50GFLOPs, i don't know the nehalems figure)

Huang essentially was giving a marketing speach about how useful GPU is going to be for the future (570X instead of 3X) (if we move the code from CPU to GPU)

I wonder how many suicides we will have among programmers that way!

Just kidding!

 

[Malo]

Welcome to B3D

Just kidding!