Nvidia BigK GK110 Kepler Speculation Thread

If that GPU-Z is legit then it is a downgrade from the 780 for anyone who overclocks.
This should have been the 770Ti(clocked lower though).
 
Speculation based on K6000 power consumption. Also JHH kept going on about how low power and quiet it is.

But maybe the GPUs for the K6000 are specially binned? The K6000 also doesn't have boost. And we all know that boost is an efficiency killer because it increases voltage quite a bit from 1.0V (base clock) to 1.17V (max boost clock). Like 20-30% more power for 10-15% more performance.
 
Also JHH kept going on about how low power and quiet it is.

Yep, it sure sounds like the GK180 was called upon from the Quadro/Tesla-division to use for the "rescue" against R9 290X. nVidia want to keep their $650 price point, so this was the best solution they could come up with. Anyway, the GK180 might have better yields than a fully enabled GK110.
 
Last edited by a moderator:
Here's a comparison of some Fermi and Kepler GeForce and Quadro parts.

Fermi
GTX 470: 448 CCs, 1215 MHz shader clock, 215 W TDP, 5.06 GFLOPS/W.
Quadro 6000: 448 CCs, 1150 MHz shader clock, 204 W TDP, 5.05 GFLOPS/W.

Kepler
GTX 680: 1536 CCs, 1006 MHz clock, 1058 MHz boost, 195 W TDP, 15.85 GFLOPS/W.
Quadro K5000: 1536 CCs, 700 MHz clock, no boost, 122 W TDP, 17.62 GFLOPS/W.
Quadro K6000: 2880 CCs, 900 MHz clock, no boost, 225 W TDP, 23.04 GFLOPS/W.

Unlike the situation with the 470 and 6000, the 680 has 90% of the GFLOPS/W of the K5000, at least partially due to the presence of the boost on the 680. I'll assume the 780 Ti is a GK180 part like the K6000. Taking this 90% number gives 20.72 GFLOPS/W for 4660 GFLOPS at 225 W and 5180 GFLOPS at 250 W (both higher than Titan's 4500 GFLOPS).

Pictures from TweakTown indicate that it is 8-pin + 6-pin, so a maximum of 300 W.
 
Last edited by a moderator:
Could anyone explain the difference between GK110 and GK180?
I thought the Quadro K6000 was just a GK110 with more ALUs activated.
 
PCIe 3.0 compability on S.2011-systems seems to be fixed on K40 and K6000.
But since no real change in device ID it is probably more a GK110-A3.
 
Could anyone explain the difference between GK110 and GK180?
I thought the Quadro K6000 was just a GK110 with more ALUs activated.

Not a whole lot of info out here. Probably along the lines of GF110 vs GF100. Bug fixed and optimized for yields. Hopefully yields on the full 2880 core chip will be high enough to see a consumer version, etc.
 
Not a whole lot of info out here. Probably along the lines of GF110 vs GF100. Bug fixed and optimized for yields. Hopefully yields on the full 2880 core chip will be high enough to see a consumer version, etc.

Isn't that what the GTX 780 Ti is (the consumer version of the GK180 with DP neutered)?
 
Not a whole lot of info out here. Probably along the lines of GF110 vs GF100. Bug fixed and optimized for yields. Hopefully yields on the full 2880 core chip will be high enough to see a consumer version, etc.
gf110 definitely had some differences in the actual logic compared to gf100 (supported full-speed fp16 filtering for instance). So far I haven't seen anything that would indicate gk180 is really a new chip (at the very least requiring new masks) and not just something along the lines of pitcairn->curacao. But could be of course.
 
Not really, there are differences for example K6000 is the only one to support 64-bit atomics

[citation needed]
64bit atomics are already listed in the gk110 whitepaper. Doesn't mean there couldn't be more of them in future chips but that's the first I heard of this...
 
[citation needed]
64bit atomics are already listed in the gk110 whitepaper. Doesn't mean there couldn't be more of them in future chips but that's the first I heard of this...

This is straight from a conversation with an nvidia engineer, he said 64-bit atomics were exclusive to Big Kepler but was even more specific since he said "K6000 and Maxwell" so I thought it was an exclusive feature of the Quadro, maybe I misunderstand him though because you're right it's mentioned in the GK110 whitepaper.
 
A claimed GPU-Z of the GTX 780 Ti from Chiphell:

http://www.chiphell.com/data/attachment/forum/201310/21/200004cldc8fc9b2syz8sb.jpg
EDIT: The picture seems to be down for me, in case that's true for anyone else here are the specs (I saved the picture): GK110 A1, 2496 CCs, 902 MHz clock, 954 MHz boost, 1502 MHz (6008 Mbps) memory, 384-bit bus, 3 GB VRAM, Device ID "10DE - 1006," 48 ROPs, 208 TMUs, PCI-e 2.0.

The GFLOPS (4503) are very close to Titan's (4500). Also, MyDrivers.com has a rumor of a "new GTX 780" that is not the 780 Ti.

Kun Peng (MyDrivers) (Google Translate) said:
Performance, the new GTX 780 is said to be able to easily surpass Titan, although looks very tricky, but this is not a difficult task, to know GTX Titan in the previous test, the magnitude of the leading GTX 780 on average only about 12% , while in the GTX 780 overclocked to 1GHz or so later, GTX Titan in performance are not any obvious advantage.
 
Last edited by a moderator:
Back
Top