[VC]AMD Fiji XT spotted at Zauba

[USER8000]

But NV's research paper shows that GM200 will have more than double the GFLOPs/Watt of GK110.
http://videocardz.com/51195/nvidia-maxwell-gm200-pascal-gp100-confirmed-research-paper

Nothing precludes NV from releasing an even faster professsional card and raise prices even more. After all, we didn't exactly have $1000 gaming Titans and $3000 dual-GPU Titan Z cards 10 years ago. We have seen NV and AMD both raise prices on tech, with Titan Z and 690 priced well above cards like GTX590/5970/6990 and so on. NV can easily create another class of professional cards priced even higher than GK210.

Based on the research paper, the timelines he provides are about 6-months off from wide launch availability. It says Summer 2014 for GK210 but it just launched recently. Since the paper specifies Dec 2014 for GM200, it should launch within 6 months of that.

Please reread the toms hardware test everyone seems to be taking this from.

This simply isn't true.

Yet it appears to be. Under constant compute loads and SP ones at that its nowhere near the improvement with gaming loads. You need to consider the GM204 cards have more refined throttling under such loads than the previous generation.

I am not saying there is no improvement but its much less than we saw with peaky gaming loads.

This is also another thing which people are ignoring - the GK210 is a GK110 with improved compute and improved performance/watt.

It also came out of the blue too.

There would be no reason to release it if the GM200/GM210 was such a massive improvement as Nvidia says it is for compute and GPGPU loads.

Not only does it take sales away from commercial users potentially wanting to upgrade from GK110 based cards,but more importantly it looks much less of an improvement going from a GK210 to a GM200/GM210 than a GK110 to a GM200/GM210.

Plus if the GM200/GM210 is coming so soon and is such a big improvement then why bothered wasting resources on another 28NM GPU,especially one which will PROBABLY never be in a gaming card,and hence not have reduction in overall dev and production costs(by taking the GPUs which don't pass the standard for commercial use) ??

Instead they are pitching the GK210 to customers.

Re-spinning the GK110 indicates something is not quite right ATM.

I would garner its delayed in some way for a reasonable length of time.

Edit to post.

That article says summer 2014 for the GK210 and yet its nearly the end of 2014,which is like a six month delay.

Second Edit to post.

If the GM210 is co-released with the GK210 at a higher pricing tier than the already expensive GK210 based cards,how much would a consumer card based on it cost then??

Third Edit to post.

The GK110 is already 565MM2,and the GK210 has more cache,which means it is probably bigger than that.

Unless Nvidia is making the GM200/GM210 on 16NM/20NM,then its likely its going to be even bigger then??

If it isn't much bigger than a GK110/GK210 then it does make any sense to be producing two similarly sized dies on the same node.

The only reason we would be seeing two the GK210 and the GM200/GM210 literally launching within months of each other is if the GM200/GM210 is a massive 28NM chip or made in limited quantities on the 16NM/20NM nodes hence driving cost up massively,which would bode poorly for pricing of gaming cards,and any delay could be down to yields in both cases.

Even a recent leak(from India) of the R9 390X indicated quite a high pricing too,which would fit it being a very large chip too.

 

[Enigmoid]

Yet it appears to be. Under constant compute loads and SP ones at that its nowhere near the improvement with gaming loads. You need to consider the GM204 cards have more refined throttling under such loads than the previous generation.

I am not saying there is no improvement but its much less than we saw with peaky gaming loads.

Did you double check?

Toms has since fixed their charts, 177W for the 980 reference under GPGPU load.

GM204 reference definitely throttles under heavy load, no doubt about that. However, the performance even under throttling conditions is high enough, that even with throttling, performance is generally improved more than in gaming. Ryan Smith in the other thread dedicated to this topic stated that the reference 980 never exceeded TDP, instead it throttled.

Anandtech 980 review Compute http://www.anandtech.com/show/8526/nvidia-geforce-gtx-980-review/20

Vs. 780 Ti

Luxmark + 54%
Compubench 1.5 Face Detection + 67%
Compubench 1.5 Optical Flow + 0%
Compubench 1.5 Particle Simulation + 35%
Sony Vegas Pro +25%
F@H Explicit SP +25%
F@H Implicit SP +45%
C++ AMP + 18%

(Not looking at DP as maxwell devotes much less FP 64 shaders per SMX)

Solid gains almost all around, generally equal to or better than the 980 does in comparison in gaming tests. If instead you compare to the 770 (much better comparison both being 104 chips) the improvement is often more than 2x and greater than the gain over the 770 in gaming. If anything the 980 is more efficient in compute (easily possible given how bad Kepler was).

 

[ocre]

Underestimating AMD would only make the buzz a million times bigger if they launch a groundbreaking new GPU.

Setting the bar real high before launch is a mistake, i dont care how you spin it

 

[SlickR12345]

Expect the high end graphic card by AMD to be about 50% faster than GTX 980.

Nvidia will release a new high end, the 980 TI, but won't be able to compete in performance with AMD's flagship either. Expect the 980TI to be about 10-15% faster than the 980.

Hopefully they release a 960 though at $200 so we can have mid range graphic cards go down in pricing.

 

[Silverforce11]

Expect the high end graphic card by AMD to be about 50% faster than GTX 980.

Nvidia will release a new high end, the 980 TI, but won't be able to compete in performance with AMD's flagship either. Expect the 980TI to be about 10-15% faster than the 980.

Hopefully they release a 960 though at $200 so we can have mid range graphic cards go down in pricing.

That's about a realistic expectation as Elvis being kidnapped by Aliens..

 

[tviceman]

The GK110 is already 565MM2,and the GK210 has more cache,which means it is probably bigger than that.

Unless Nvidia is making the GM200/GM210 on 16NM/20NM,then its likely its going to be even bigger then??

If it isn't much bigger than a GK110/GK210 then it does make any sense to be producing two similarly sized dies on the same node.

Maxwell has about 25-30% more perf/mm2 than Kepler on 28nm. It also has 50-75% more efficiency in perf/watt. GK110 was limited by it's TDP. Nvidia clocked it as high as it would go to stay at or under the 250w power envelope. Thus, it's clocks were in the low 900mhz range for reference models. Notice how maxwell is clocking much, much higher than Kepler.

Two things we can easily deduce from all this information.
1) GM200 can't have as much of a functional increase in transistors dedicated to shaders as GM204 had over GK104 because there isn't enough room left.
2) GM200 can, however, be clocked much higher than GK110 because it's much more efficient than GK110 AND with respect to GKM204, it won't be as a large an increase in die size and transistor count that GK110 was over GK104. GTX980 was able to clock 13.5% higher than GTX770 while maintaining the same power draw but growing 33% in die size. Comparatively, GM200 should have a larger percentage clock bump over GK110.

Basically GM200 is going to be a trade off of being less of a leap in transistors and shaders but will make up for it in it's ability to be clocked higher due to a more efficient design and a die size closer to it's smaller sibling. Think of it this way: Had GK110 been 12 SMX's instead full functioning instead of 15 SMX's, it would have been able to be clocked noticeably higher before hitting the 250w envelope that GK110 reference cards are at. The big downside is that there likely won't be as much OC headroom as GK110 has, since it's going to already be a huge die. GK110 clocked as high (most of the time) as most GK104's though, so who knows.