Pascal will be 10X faster than Maxwell

P

PRIME1

2[H]4U
Joined
Feb 4, 2004
Messages
3,942
http://blogs.nvidia.com/blog/2015/03/16/live-gtc/

Pascal has three great features. One is mixed precision, which is provides at 3X the level of Maxwell. Another is 3D memory, which provides more bandwidth and capacity simultaneously. A third is NVLink, which is an ability for multiple GPUs to be connected at very high speeds. It also has 2.7X more capacity than Maxwell.

As a result, Pascal will be 10X faster than Maxwell, although JHH mentions that this is just a very rough, high level estimate.
Click to expand...

:eek:
 
Pascal will be 10x faster than Maxwell*

* 1.) For neural nets. 2.) Assumes 8 GPUs instead of 4 3.) Utilizing half precision. 4.) Calculated with what Jen-Hsun calls "CEO math". ;)
 
Can't wait to see what Pascal will offer. Hopefully it continues to improve in terms of efficiency as well.
 
Jen-Hsun has turned into a giant winded bag. Besides 3D Memory there is nothing on the consumer side that's anything big. Well, he mentioned the "natural speed increase" from being on a smaller node, but the other stuff is mostly related to non-consumer uses.

NVLink has ZERO use for 99% of us.
 
Good news for me... no need right now to swap my 970, which is comfortable for 1080p gaming. In a year, getting 1070, which will beat Titan X for fraction of money :)
 
Liger88 said:
Jen-Hsun has turned into a giant winded bag. Besides 3D Memory there is nothing on the consumer side that's anything big. Well, he mentioned the "natural speed increase" from being on a smaller node, but the other stuff is mostly related to non-consumer uses.

NVLink has ZERO use for 99% of us.
Click to expand...

Well, it is a developer's conference, the presentation wasn't aimed at the consumer crowd. Apart from Titan X, all the other stuff presented here are for the industry, not gaming community.
 
LOL, CEO maths.

That said, my next system is going to be Skylake-E with at least 3 Pascal-powered video cards :cool:.
 
Yeah I'm totally skipping the Titan x, my 980s are beasts enough. Waiting for Pascal
 
Pascal is being hyped like every other new GPU architecture over the past decade...I'm sure it'll be faster then anything currently on the market...and whatever comes after Pascal will be 10 times faster...and so on...and so on
 
Hornet said:
Well, it is a developer's conference, the presentation wasn't aimed at the consumer crowd. Apart from Titan X, all the other stuff presented here are for the industry, not gaming community.
Click to expand...


lol right. Even his gaming presentation speeches are long winded these days boring the audience. I'm more of a, "Here it is." drop the mic kinda person I guess.
 
Sprayingmango said:
Yeah I'm totally skipping the Titan x, my 980s are beasts enough. Waiting for Pascal
Click to expand...

In another thread, some fanboys would say "cool story bro" and add a childish image like its 1995.

But because I'm a realist, I too agree with what you said.
 
Sprayingmango said:
Yeah I'm totally skipping the Titan x, my 980s are beasts enough. Waiting for Pascal
Click to expand...

Upgrading to every single new card is a huge waste of money.

I'm keeping my 2x R9 280xs until at least the new ~16nm process chips like Pascal come out. Maybe they'll be something worth buying on the CPU side then and I'll just build a whole new machine.
 
3D memory... seriously, they could not find a better term. what next Cloud Interlinks for faster PCIe connections!
 
MrGuvernment said:
3D memory... seriously, they could not find a better term. what next Cloud Interlinks for faster PCIe connections!
Click to expand...


It's actually called Stacked RAM. Sounds a lot better when you say it that way. Calling it 2.5/3D definitely does not bod well marketing wise.
 
Pascal will be 10X faster than Maxwell

I don't think that much Hyperbole could fit through a doorway.
 
Vega said:
Pascal will be 10X faster than Maxwell

I don't think that much Hyperbole could fit through a doorway.
Click to expand...

You just left out some key words, that's all.

Pascal will be 10X faster at extracting money out of people's wallets than Maxwell
 
I wonder how NVLink will actually improve multi-GPU experience. Is the current SLI link a bottleneck at all?
 
I will spend 10x what I paid for my GTX970 to upgrade to a Pascal!

Not.

I'll wait for the [H] review.
 
If PRIME1/Jen-Hsun told me it was sunny outside I would have to go to a window and check.
 
3D memory and a faster link between GPUs are actually very good performance enhancements that will make an impact on gaming. AMD already has the XDMA engine in the R9 290 series, which provides higher throughput for inter-GPU communication than the CrossFire bridge did, and it made a pretty big impact on frame times (coming from someone who has used both older generation AMD cards in CrossFire and NVIDIA cards as well my 290Xs.)

Sure, if you don't use more than a single card NVLink won't benefit you - but 3D memory will, and what else exactly do you expect? 90% of the gains in single GPU boards have always been from core count, clock speed, and process improvements.
 
chenw said:
I wonder how NVLink will actually improve multi-GPU experience. Is the current SLI link a bottleneck at all?
Click to expand...
It looks like it is yes. Read the 980 SLI review that HardOCP did where it compares against the 290X in 4K.
 
chenw said:
I wonder how NVLink will actually improve multi-GPU experience. Is the current SLI link a bottleneck at all?
Click to expand...

It appears NVLink is really oriented towards GPU compute applications, and is meant to overcome the current bandwidth limitations of PCIe 3.0.

NVLink is an energy-efficient, high-bandwidth path between the GPU and the CPU at data rates of at least 80 gigabytes per second, or at least 5 times that of the current PCIe Gen3 x16, delivering faster application performance.
Click to expand...

Source

For gaming it's going to make 1/4 of an iota of difference at best.
 
He clarified his statement at the beginning by saying 3x the FP performance on the same power envelope. Aka what might be applicable to games. Minus the gains from HBM and 16FF process and it's probably half that, or 1.5x more efficient per CUDA core. Which is still a huge jump considering Maxwell is 40% more efficient.
 
Rizen said:
3D memory and a faster link between GPUs are actually very good performance enhancements that will make an impact on gaming. AMD already has the XDMA engine in the R9 290 series, which provides higher throughput for inter-GPU communication than the CrossFire bridge did, and it made a pretty big impact on frame times (coming from someone who has used both older generation AMD cards in CrossFire and NVIDIA cards as well my 290Xs.)

Sure, if you don't use more than a single card NVLink won't benefit you - but 3D memory will, and what else exactly do you expect? 90% of the gains in single GPU boards have always been from core count, clock speed, and process improvements.
Click to expand...

NVLink won't be seen on consumer GeForce cards.
 
So I guess we should stop all purchases from Nvidia now and wait for the release of super duper Pascal?
 
Just to kill urban legends-in-the-making: PASCAL MIGHT BE 10x FASTER ONLY FOR VERY SPECIFIC NARROWMINDED TASKS RELATED TO SOME EGGHEADS ALGORITHMS. THIS HAS ABSOLUTELY NOTHING TO DO, REPEAT NOTHING TO DO, WITH THE GAMING PERFORMANCE.
 
I'm 95% certain anyone with a functional brain pretty much assumed that the 10x referred to specific compute scenarios, at least that was the first thing I assumed.

I mean really, gaming-wise never in the entire history of either ATi/AMD or nVidia have they put out a card that managed more than 3x improvement in one generation, nevermind 10x. And I think I'm being generous with the 3x figure already.
 
They obviously see FP16 as the next big HPC market and there is some validity to that if NN really take off in the next 5-10 years.

So the 10x of Pascal boils down to:
2x the performance by moving from 4GPUs to 8GPUs in an array.
2x the performance of Maxwell's FP32, easy peasy with HBM and 16FinFet.
2x the performance by adding the ability to execute paired FP16 instructions on a Pascal FP32 ALU, ala Tegra X1.

All you have to do to meet the 10x now after that 8x improvement is get 25% more performance from making the SMMs more efficient, scheduler more efficient, the drivers more efficient, increasing the clockspeed or a mix and match of them.
 
Last edited:
n=1 said:
I'm 95% certain anyone with a functional brain pretty much assumed that the 10x referred to specific compute scenarios, at least that was the first thing I assumed.
Click to expand...

Seen those "Yeah I'm totally skipping the Titan x, my 980s are beasts enough. Waiting for Pascal" ? Before this will shape into new urban legend it is worth trying to kill it, otherwise there will be endless waste of time on this. Let's face it, if Pascal is 2xTitan X in gaming it will be spectacular, realistically 50-60% might be expected.
 
What a let down. I already told myself I wouldn't invest in anything less than 11x faster than Maxwell. Oh well, guess I'll skip Pascal and wait for Turing.
 
You must log in or register to reply here.
Back
Top