• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

Nvidia news direct from Jen-Hsun Huang!

Associate
Joined
24 Sep 2008
Posts
390
Location
Telford, UK
Last edited:
yeah, I saw the thread but had missed the interesting stuff so found info on Fud and posted it for those who also might have missed it. The article has a bit more info than the slide in post #24.
I presume the demo will be posted on youtube or summit?
 
Last edited:
Believe it when I see it lol, 4x the performance per watt :/ hmmm do they mean of the actual architecture or just due to 28nm lol?
 
Believe it when I see it lol, 4x the performance per watt :/ hmmm do they mean of the actual architecture or just due to 28nm lol?

I'm sure that they'll be able to come out with some part they're able to reduce the voltage of enough to claim that as fact. :p
 
i watched it and tbh was very good.

they are heading in the right direction if you ask me, i see it being the future and the cuda is defo gonna become the standard over its competition.
 
Roadmap pic i put in the other thread
cudagpu.jpg


I was most impressed by the adobe stuff though, re focus looked amazing.
 
A lot of that is because Fermi is a disaster from a performance per watt point of view.

Hmm, well it is but its a disaster due to design, not the process, and I see where the performance is coming from, he's talking about DP performance I would guess, not normal performance, the difference is very easy.

If the architecture is setup so for instance you can only run DP instructions at half the speed of SP instructions, then by tweaking the architecture so it can run the same DP as SP you'd be doubling performance per watt, switching to a new process generally brings with it a doubling of performance per watt, I would guess this is where 4x the performance/watt comes from.

Which unfortunately suggests a similar architecture, which will unfortunately mean its likely to suffer from every problem Fermi had.

We'll see, I just don't see him making the necessary changes, if they go with GloFo, which seems to be the case, then theres potential for less issues with a huge core if its a better process/yields.

However theres a very good chance only Tegra will get done at GloFo(as they'll be pretty much the masters of ARM core production as they are partners), meaning probably a lot of sharing of performance/yield tweaks to ARM based chips to be had by working with GloFo.

Didn't the 480gtx supposedly have 8times the throughput of DP instructions as the 285gtx anyway, also on a lower process probably meant it had 16x the increase in performance/watt for DP instructions anyway..... that doesn't mean Fermi was more power efficient or a great design compared to the 285gtx though ;)

You can spin anything, you HAVE to spin everything if you work at Nvidia though :p
 
That graph is confusing. Does 'Tesla' refer to G80 or GT200? I'd guess GT200 since I'm pretty sure G80 doesn't support double precision. However, Nvidia's site seems to suggest the 'Tesla architecture' is G80 (based on number of shaders and memory configuration). Then again G80 came out in 2006, and GT200 came out in 2008. If it is G80, then why is GT200 omitted when its performance and feature set are improved? God damn it Nvidia. You and your graphs. Oh well, at least we know some code names and some dates that are almost certainly wrong and probably in the wrong order. :confused:
 
Also don't confuse the Fermi TESLA cards DP with the GTX400 cards DP, remember they castrated the DP in the consumer cards.
 
Hmm, well it is but its a disaster due to design, not the process, and I see where the performance is coming from, he's talking about DP performance I would guess, not normal performance, the difference is very easy.

If the architecture is setup so for instance you can only run DP instructions at half the speed of SP instructions, then by tweaking the architecture so it can run the same DP as SP you'd be doubling performance per watt, switching to a new process generally brings with it a doubling of performance per watt, I would guess this is where 4x the performance/watt comes from.

It would require much more than a 'tweak' to make DP as fast as SP- I would guess that it is impossible.
A large part of why Fermi is poor when it comes to performance per watt IS the process. The 40nm process has a lot of leakage which ATi anticipated and because of that they did a much better job executing their design than Nvidia did. That doesn't mean that Nvida didn't mess up, they did, but it is possible that they would see a huge jump in performance per watt increase JUST by moving to 28nm- way above what ATi will get as ATi protected against it by using a slightly larger die than they would have liked.
 
Last edited:
Back
Top Bottom