It might be even higher than HD5870, considering even current GTX460 has already higher powerdraw than HD5850
How come there is no next generation Nvidia speculation?
- Thread starter Mendel
- Start date
Silent_Buddha
Legend
Well, considering 5870 already has the same or lower power draw as 5850.
Regards,
SB
Erm what are you smoking?? 5850's TDP is 151 W and 5870 is 188W
Silent_Buddha
Legend
Whoops, nevermind, you're right. I had a brain fart and was thinking of Idle Power. Ooopsie.
Regards,
SB
Regards,
SB
That seems to depend on the model and/or sample variance. I've seen some numbers (mostly from ht4u which measure real power draw) which suggests power draw is quite close between these two, and sometimes (at least for nonOC versions) not higher at all. The rest was extrapolated from OC cards, which typically don't increase voltage, hence suggesting a GTX475 card might not need more voltage neither, in contrast to HD5870 which has higher voltage than HD5850. It might have been a tad optimistic and power draw could end up quite similar to HD5870, but in any case I think it still would be slightly below that of GTX470.
I think you must have been thinking of the 5830…! (vs 5850)
I just checked the TPU review from last week. Only win for the 460 is idle power, for the rest 1 GB 460's consume more. http://www.techpowerup.com/reviews/MSI/GTX_460_HAWK/26.html
That Hawk is clocked 18% higher and it's peak power usage is also around 20% higher.
Actually the only test that makes the 460 look efficient are movie playback and Furmark, for typical gaming, you're getting less FPS/W (as evidenced by the Perf/W scores:http://www.techpowerup.com/reviews/MSI/GTX_460_HAWK/29.html where it's trailing the 5850 and 5870 by 30 and 10%.
But I think we're going off topic here aren't we?
Yes it's pretty interesting - at idle GTX460 use less power, I guess because they are lowering mem voltage too. "Normal" 3d load GTX460 use more power and Furmark is pretty much a draw again (could easily be sample variance) (I was going by Furmark numbers previously).
Oh yes there's no doubt - slightly higher power draw (with typical load) while somewhat slower.
Well to keep it on speculation topic, if you'd add 10-15% on top of the HAWX power consumption for that additional SM you'd definitely have very comparable max power draw to HD5870 (though higher for gaming), but in any case still less than GTX470.
Megadrive1988
Veteran
I think there should be an Nvidia Kepler & Maxwell thread.
Kepler is Nvidia's next-gen GPU architecture due in 2011 on 28nm.
Maxwell is their next-next-gen GPU architecture due in 2013 on 22nm.
Kepler is Nvidia's next-gen GPU architecture due in 2011 on 28nm.
Maxwell is their next-next-gen GPU architecture due in 2013 on 22nm.
We've all seen the slide, but is that an announcement date or an availability date?
Isn't Kepler Fermis 28nm upgrade (similar to what gtx280 was to 8xxx series) and Maxwell is their Next-gen GPU? (8xxx series to fermi)
Thus Kepler is closer to standard Fermi talk and Maxwell the pie from the sky speculation?
Last edited by a moderator:
There already is a Kepler thread, has been for quite some time already
http://forum.beyond3d.com/showthread.php?t=58668
Megadrive1988
Veteran
DarthShader
Regular
10 TFLOPs from nVidia by 2018
http://www.eetimes.com/electronics-news/4210815/Nvidia-describes-10-teraflops-processor
Somehow the numbers don't add up. Will the cores be running at close to 5Ghz, or will there be new instructions doing more FLOP's per clock.
http://www.eetimes.com/electronics-news/4210815/Nvidia-describes-10-teraflops-processor
Somehow the numbers don't add up. Will the cores be running at close to 5Ghz, or will there be new instructions doing more FLOP's per clock.
10 TFLOPs from nVidia by 2018
http://www.eetimes.com/electronics-news/4210815/Nvidia-describes-10-teraflops-processor
Somehow the numbers don't add up. Will the cores be running at close to 5Ghz, or will there be new instructions doing more FLOP's per clock.
The article stated explicitly that each core will be doing 4 DPFLOPS/clk. I can think of two ways to do this.
1. Adding lots more schedulers & dispatch ports to each core. Although I'm not sure how the inter-scheduler arbitration scales, this might be expensive.
2. Making each CUDA thread operate in VLIW (or SIMD) style, similarly to AMD's current architecture.
I'm sure there are other ways of doing this as well...
The way I understand the article is:
4 DP FLOPS per "core" => 8 SP FLOPS => 4 SP FMAs => he's actually talking about 4 SIMD lanes when he says "core".
He said 8 per SM, so that's 32SPs/SM, just like today but apparently with a different organization.
He also said 128SMs total, so that's 4096SPs total. You'd need 2.5GHz to get 10TFLOPs. That won't be easy, but it might be doable.
At 40nm, you get 512SPs. So you should get about 1024 at 28nm, 2048 at 20nm, and finally 4096 at 14nm.
28nm will be available in latish 2011, 20nm should follow in late 2013, and finally 14nm in 2015~2016. Let's call it 2017 for delays, and 2018 because things don't always scale perfectly, especially when you add more functionality, plus NVIDIA needs pretty high clocks. So I think 10TFLOPS from NVIDIA in 2018 is realistic. That is if NVIDIA still exists in 2018.
Obviously, AMD should be there a lot sooner, since Cayman will probably break 3TFLOPS next month.
4 DP FLOPS per "core" => 8 SP FLOPS => 4 SP FMAs => he's actually talking about 4 SIMD lanes when he says "core".
He said 8 per SM, so that's 32SPs/SM, just like today but apparently with a different organization.
He also said 128SMs total, so that's 4096SPs total. You'd need 2.5GHz to get 10TFLOPs. That won't be easy, but it might be doable.
At 40nm, you get 512SPs. So you should get about 1024 at 28nm, 2048 at 20nm, and finally 4096 at 14nm.
28nm will be available in latish 2011, 20nm should follow in late 2013, and finally 14nm in 2015~2016. Let's call it 2017 for delays, and 2018 because things don't always scale perfectly, especially when you add more functionality, plus NVIDIA needs pretty high clocks. So I think 10TFLOPS from NVIDIA in 2018 is realistic. That is if NVIDIA still exists in 2018.
Obviously, AMD should be there a lot sooner, since Cayman will probably break 3TFLOPS next month.
I think AMD will slow down a bit on the theoretical flops increases and invest elsewhere. Bumping up the raw flops has diminishing returns as the data and algorithms get more complex and data management / bandwidth become more important.
Agreed. But they're so much ahead that they stand a good chance of getting to 10TFLOPS significantly earlier, though perhaps not in DP.
If you count dual-GPU cards, they're even further ahead, with Hemlock almost at 5TFLOPS. I guess Antilles should end up a bit over 5TFLOPS.