LinkBack URL
About LinkBacksTake with the usual grains of salt:
http://wccftech.com/alleged-gtx-tita...sisoft-sandra/
http://wccftech.com/amd-fiji-r9-390x-specs-leak/
Take with the usual grains of salt:
http://wccftech.com/alleged-gtx-tita...sisoft-sandra/
http://wccftech.com/amd-fiji-r9-390x-specs-leak/
Bandwidth calc for the AMD card is mussed: the 128GB/s per stack figure is a 1GHz, not 1.2GHz. @ 1.25GHz, you're looking at 160GB/s per stack, or 640GB/s for the 4 combined. That's monstrous....
Holly hell that must be one crazy fast GPU to need 640Gb/s of memory bandwidth. If the Sisoft link in right this card will be a performance jump like nothing else.
One thing though, if the GPU is 4096SP that would make the chances of this being a 28nm card pritty slim.
Having and needing are two different thingsOriginally Posted by jigger
Roughly the same as previous generation or two ago in CF/SLI? That's not an unusual performance leap. Especially for a cryptography bench that could be sped up with some dedicated acceleration hardware.If the Sisoft link in right this card will be a performance jump like nothing else.
It could simply be a big die. Or simpler/smaller SPs.One thing though, if the GPU is 4096SP that would make the chances of this being a 28nm card pritty slim.
Maybe, but thats all pritty unlikely to be the case I would say.Having and needing are two different things
Roughly the same as previous generation or two ago in CF/SLI? That's not an unusual performance leap. Especially for a cryptography bench that could be sped up with some dedicated acceleration hardware.
It could simply be a big die. Or simpler/smaller SPs.
If we assume AMD have made at least some kind of pipeline performance improvements to this GPU and haven't gone overboard on the memory for nothing, this looks all set to be God like fast come release. Just scaling Hawaii up to around the 3200 SP mark and no other improvements would make an incredibly fast card.
The 290X already has 2816 SPs, so going up to 3200 isn't going to be life changing. The 295 shows you what an equivalent 5632SPs performs like and it's not what I'd call god-like.
With efficiency the loss that comes with mulit GPU setups.
Last edited by jigger; 16-11-2014 at 02:06 AM.
Sure, but that the overheads would still allow for a greater than 3200SP representation.
Funnily enough it seems we are going to see a 3200 pipe Hawaii card.
http://www.sisoftware.eu/rank2011d/s...98ffcc1f9&l=en
But with only 3 GB RAM? Still, it doesn't seem to have this amazing performance you were suggesting.
What hell are you talking about? The 3200SP card I mentioned first was in response to you saying a GPU that requires an insane amount of memory bandwidth will not be particularly fast.
I don't think that's what was said - just that a card that *has* that much bandwidth doesn't necessarily *need* that much bandwidth. As to a 3200sp Hawaii card being "incredibly fast" (your words) - it would at best be ~ 14% faster than a 290X (assuming that number of SPs is the only current limiting factor on Hawaii), which wouldn't even take it past an overclocked GTX 980.
As kaniel says, 4098sps would only match two 7970s in crossfire, plus maybe 15% generational improvements and maybe a bit of extra headroom thanks to the hugely excessive bandwidth that 4 HBM cubes would give. Compare a couple of 280Xes in crossfire to a stock 290X on Uber mode (i.e. custom designs *will* be faster) here: http://www.anandtech.com/bench/product/1058?vs=1059 and you'll see that the best result the CF setup gets is about 50% increase. That *used* to be the expected generational advance back in the day. And the GTX 980 is the second rung Maxwell card - there will be a bigger, significantly faster, NVidia card coming out. A 4098 shader AMD rival card might be struggling to keep up, tbh....
kalniel (17-11-2014)
The thing is though,the R9 285 with 256 less shaders and significantly lower bandwidth is matching an HD7970 at similar clockspeeds(around 900MHZ),and is getting up to R9 280X levels in certain games which are more tessellation limited. Its quite possible a full fat 2048 shader Tonga chip will actually get quite close to a R9 290,even if clocked at 1GHZ,so it will be interesting to see how a larger chip will perform.
OTH,the R9 390X and Titan II will be made for mixed compute/gaming workloads,which means power and die area will be devoted towards features which won't add anything to gaming performance.
Look at where the GTX780TI and R9 290X are,when compared to the R9 280X/HD7970GE and GTX770/GTX680. Despite the large increase in shaders and memory bandwith the GTX780TI and R9 290X did not show similar scaling in performance,despite the R9 290X having 40% shaders than a R9 280X/HD7970GE and the GTX780TI having 88% more shaders than a GTX770/GTX680.
If anything the GK110 seems to be incredibly unbalanced for gaming against the GK104 - more than even Hawaii when compared to Tahiti.
So when people were surprised by the performance of the GTX980,its not really that surprising - with a 33% increase in shaders over the GTX770/GTX680 and 10% to 20% extra clockspeeds,more ROPs,more RAM and better texture compression it was pretty easy to match or beat a GTX780TI,especially when the GTX980 does not need all the functionality for GPGPU which adds to power consumption,heat production and lower clockspeeds.
The GM204 is around 33% bigger than the GK104 for the performance boost it gains.
Unless both the Fiji and GM200/GM210 are more efficient gaming wise against the midrange chips,unlike the previous generation,I am not sure we are going to see a huge performance bump from either - and this is especially true for Nvidia.
Remember,AMD managed to get most of a performance of a GK110 for gaming in a 438MM2 chip - the GK110 was around 565MM2,and Hawaii was packing a larger memory controller and audio DSPs too.
This is the thing too - Tahiti was around 365MM2,so if anything Hawaii,further improved the gaming performance/mm2 over Tahiti and added audio DSPs too,and this is for GPUs which have the added burden of considerably higher overall DP performance and large memory controllers meant for certain GPGPU operations too.
The GK110 went the other way - just like all previous top end Nvidia chips(even worse than AMD).
AMD simply have more die area to improve performance with. OTH,one area Nvidia has an advantage is usable TDP,but again this assumes AMD makes very little gains in performance/watt with their next generation.
Last edited by CAT-THE-FIFTH; 17-11-2014 at 12:32 AM.
Agree with pretty much everything you say CAT, except this: while Hawaii had a wider memory controller, I believe in silicon terms is was actually a little smaller, due to the simplification of the controller and the lower target memory speeds.
Well I bought a R9 290 about a year ago and while it cost me a lot of money I was amazed by the performance. A 3200SP version of that card would be incredible.
We have to assume AMD have made some improvements to the GPU since the last design, and AMD have fitted the 4096SP card with 640Gb/s of memory bandwidth for a reason. I mean, it is a 4096 SP chip afterall...
I suspect that's because they don't want to go back a step in terms of memory offered on their flagship card, and HBM only comes in 1GB cubes at the minute. To make up a 4GB framebuffer they *have* to use four cubes, which means they *have* to offer a 4096bit controller.
The bottom line is modern cards aren't predominantly bandwidth limited. You can tell this fairly easily: overclocking just the GPU tends to yield better results than overclocking just the memory. HBM will provide a point at which memory bandwidth will be taken completely out of the equation: doubling the bandwidth is a huge advance in one generation. 2008's GTX 280 has 141GB/s bandwidth, the 4870 had 115GB/s. Compare those with 2012's GTX 680 and HD 7970 - both three generations on - and nvidia had only increased bandwidth to 192 GB/s, less than 50% more. AMD had bumped up a lot further: to 264GB/s, but that's still only double and a bit in three generations (and AMD will try to tell you that the 7970 is actually a different class of card, and you should be comparing the 7870). The increase from 7970 - 290X was only 21%. So to suddenly double the bandwidth available in one generation? Yeah, I don't buy that the card needs that much bandwidth.
There are currently 1 users browsing this thread. (0 members and 1 guests)
Posting Permissions
Reply With Quote
