Monday, June 17th 2013
NVIDIA GeForce GTX 760 Specifications Redux
There are many theories doing rounds about the specifications of NVIDIA's upcoming performance-segment GPU, the GeForce GTX 760. One states that it's largely similar to the GeForce GTX 660 Ti from previous-generation, with higher clock speeds, possibly 7.00 GHz memory, and GPU Boost; while another suggests a completely new core-configuration. According to a GPU-Z screenshot leaked by a ChipHell community member, NVIDIA is attempting to give the GeForce GTX 660 a successor, rather than merely retrofitting the GTX 660 Ti.
According to leaks that surfaced on ChipHell, NVIDIA is configuring a GK104 GPU with just three out of four GPC (graphics processing clusters) enabled, while keeping the memory and raster operations untouched. This approach would give the chip 1,152 CUDA cores, 96 TMUs, 32 ROPs, and a 256-bit wide GDDR5 memory interface. The card in the GPU-Z screenshot features 1072 MHz core, 1111 MHz GPU Boost, and 7.00 GHz memory.
According to leaks that surfaced on ChipHell, NVIDIA is configuring a GK104 GPU with just three out of four GPC (graphics processing clusters) enabled, while keeping the memory and raster operations untouched. This approach would give the chip 1,152 CUDA cores, 96 TMUs, 32 ROPs, and a 256-bit wide GDDR5 memory interface. The card in the GPU-Z screenshot features 1072 MHz core, 1111 MHz GPU Boost, and 7.00 GHz memory.
24 Comments on NVIDIA GeForce GTX 760 Specifications Redux
With that config it would probably not be faster than 660ti clock for clock at res 1080 and below , unless you cranked AA way up . 1080+ res would benefit a lot more of 256bus and 7k ram IMO .
I was figuring they just release 670 spec with maybe ram clocked higher for 760ti , but maybe that would perform to close to the 770 for price point they want .
No word on TDP?
somehow i doubt it..
But if You don't get it, "Looking at those numbers" strongly implied "faster on paper". Unless You don't know what that means.
If you take card like 670PE that can go to 1.212+ voltages which most reference 6xx can only go 1.175v w/o bios mod . that only benefit of boost 2 as it can do 1.2 with just software .
Yes, it was released at $400.00.
What more interesting is they could've been binning these chips like this all along. Those with one bad SMX, but still intact memory and raster part section where being shelved for like the past year. This whole model is another GK104 derivative, from the fact of Nvidia has to purchase a complete wafer production. If they’ve been amassing chips of this spec form wafers that basically already paid for while holding them back for this 7XX release that would be a new tactic and if true as the price for them dilutes the price for every GK104 over a longer period. I’d say it pretty smart/ballsy of them… sure geldings are always part of the mix, but I can’t think of an example where they’re used for a new product series release. (even the GTX9800/GTS250 wasn’t so physically different)
These thing could be priced cheap if Nvidia has large war-chest to off-load. If such an offering can stay well out in front of the GTX660Ti and cost $230 what's not to like. Then can they still have the 1344 Cuda part and 256-Bit but say 24ROP's show up later as a GTX760Ti?
The average shader count used is around 230-240sp per 4 ROPs on average, as that is what performance comparisons of products has shown over the course of many generations.
That basically means that 1344 (units including sfu) of 760 is less than the average of roughly 1408 on GTX660ti (limited by ROPs) but obviously more than (normalized) 1120 on 660 or 1280 on 7870. Obviously parts like 670/680 are limited by units, not ROPs, hence are a better example of pure scaling.
By allowing a higher tdp/clock, the stock core performance should increase by roughly 8% over a stock 660ti. The faster RAM/bus (that it doesn't even need) should boost it another 8%.
If you want a pure performance comparison, this at stock should be similar to a lot of things. The ones that immediately come to mind are a stock 670, a 7870 at a little over 1200mhz, or a 1536sp AMD part at 1000/6000 (slightly faster than a stock 7870XT ie TahitiLE).
Basically, they reconfigured the same amount of disabled logic as 660ti, gave it a higher TDP, and are using it to replace 670 (so as to beat 7870 and compete with parts such as TahitiLE and the upcoming 1536sp Hainan sku). They need to do this because no other configuration below 670 would be cost/performance competative, as 660 (vs 7870) and 660ti (versus 7870XT) currently lose those battles.
FWIW, 760ti should be a 670 with 7ghz ram and a similar clock to 760. They are likely waiting to see what happens with Hainan and will maximize margins in the meantime. Since 7 SMX essentially equals 1568 amd shaders, it will likely fall between the Hainan skus (which may be something like 1536 1000/6000 and 1792 1100/7000).
I think we see something like this happening in q4:
Hainan LE (7870) - 200
760 - 220-230
Hainan pro (Tahiti LE) - 250
760ti - 300
HainanXT - 350
770 - 400
Stagnant.
They did increase TDP on 680/>780 but that has more to do with going from GK104 to GK110 , they did raise TDP on 670>770 but not sure they will do with lower models .
When the cards go to 3rd party makers they generally crank up clocks according to model types .
From what I have seen reviewed so far with 7xx , I haven't seen much more headroom than 6xx , so I don't see where added clock headroom comes from.
Of course Nvidia could raise the reference defaults up a bit but all that would do is make the 3rd party makers not be able to increase there clocks for marketing .
I guess we should see soon how they perform