NVIDIA Kepler speculation thread

Tesla K10:
http://www.eetimes.com/electronics-news/4372936/Nvidia-upgrades-Tesla--eyes-cloud-gaming

GTX 690 with 2x 4GiB ECC and MPI(message passing interface).
.

We already discussed that.


I am not sure if i really doubt it, but Gamer-Kepler is not only weak at DP, but also Raytracing and some more OpenCL-thingies seem not to be it's greatest strength.

DirectCompute too, loses to GTX580 (and 7970 obviously) in Civ5 compute (leader) benchmark which uses directcompute to unpack textures on the fly

xEx said:
And not just that...If Nvidia cant make enough GK104 to meet the demand i have no idea how they would do it with Pro. cards that have to have harder regulations and even more if we think they are talking about dual GPUs cards...
 
A TechReport article (article mentioned from another thread) seems to discuss frame metering well before the 600 series cards were released. From what I understand they are adding lag to improve the jitter. Does that have anying to do with the polling issue mentioned earlier?
Discussion at the end of the page
 
AMD should take the opportunity and release a FirePro version with 2x 7970 cores as is the 2x GK104 K10 Tesla and follow the trend for K20

dual 7970 will end at 7.58Tflops SP and ~ 1.92 Tflops DP ...

A firepro version with lower clock speed will bring something maybe like...

~ 7.0 Tflops on SP vs 4.58 Tflops
~ 1.8 Tflops on DP vs 0,2 Tflops
( I have adjust a bit the DP SP count due to a possible decreased clock speed on the Pro version totally arbitary without calculation, just for the idea )

Imagine what AMD can do for take part in this market, they have an incredible latitude for it.

Even the K20, will have hard time to be a lot faster of an actual 7970 in peak DP/SP performance.

The K20 is aimed at 1Tflops DP ... not a lot of faster of a simple 7970 ..


When i see what Nvidia bring this time... If i was in AMD; i will immediately put my engineer on work for see what they can do .. specially if they want to push OpenCL computing, they need hardware. With GCN it seems they have the key right now for do it. Hardware wise they can obliterate the K10 without any problem ( what can happend if AMD bring a card with more SP performance but specially with an interessant DP rate at same time in the same category ? )... and looking at what is aimed the K20 in DP, a GCN 2 ( with more ALU ) will bring an interessant fight there too.
 
Last edited by a moderator:
Is it known that the chip itself is capable of 1/4 DP? Some sites IIRC claimed that it's 1/2, but limited on Radeon flawor to 1/4

Thoses are official numbers of the 7970 and confirmed in Sandrasys ..This is just for draw an idea of the capacity, radeon and FirePro should differ ofc.

Should be good to look the GCN presentation and compare both actual infos on the 7900 and what was bring during the presentation, the right infos should be there, GCN was presented there as the whole architecture capability, and ofc the 7970 dont have all of thoses capability. If it is 1/2 rate, the infos should be there ... ( i believe this is where this 1/2rate is coming )

For Nvidia K10, the infos are now official ofc.
 
Last edited by a moderator:
Imagine what AMD can do for take part in this market, they have an incredible latitude for it.

Even the K20, will have hard time to be a lot faster of an actual 7970 in peak DP/SP performance.

The K20 is aimed at 1Tflops DP ... not a lot of faster of a simple 7970

That would work if it weren't for the fact that peak paper flops are pretty irrelevant right now.

Don't you see how much nVidia is investing in relationships, education, software and infrastructure to make their HPC offerings attractive? AMD needs to focus on those things not flops. Alternatively they can let nVidia do all the work and jump on board after GPU computing matures and goes mainstream.
 
That would work if it weren't for the fact that peak paper flops are pretty irrelevant right now.

Don't you see how much nVidia is investing in relationships, education, software and infrastructure to make their HPC offerings attractive? AMD needs to focus on those things not flops. Alternatively they can let nVidia do all the work and jump on board after GPU computing matures and goes mainstream.


I completely agree ... thoses are just idea on hardware capability. the other part, the relation and the use of it is another story. Even on software side ... What if AMD cant bring the same capability of virtualistion described yesterdy for the K20? There's many thing outside " pure peak power ":
 
Last edited by a moderator:
Don't you see how much nVidia is investing in relationships, education, software and infrastructure to make their HPC offerings attractive? AMD needs to focus on those things not flops.
That is definitely true. The hardware appears to be quite capable.
AMD's ACEs can do basically an equivalent of HyperQ (they also offer independent hardware task queues complete with a priority system and all, but I have no idea how many, maybe just two, maybe more). And nV's "dynamic parallelism" reminds me an awful lot of AMD's "Compute Generated Task Graph Processing" which they said to be possible with GCN on last years AFDS (nV actually shows a task graph to visualize their dynamic parallelism ;)). Alone, one can't use it (the capabilities of the ACEs may be put to some use when you queue some kernels for different contexts, but AMD's device fission extension for OpenCL [or something more elaborate as you can expect from nV] is not yet supported on GPUs).
 
AMD needs to focus on those things not flops. Alternatively they can let nVidia do all the work and jump on board after GPU computing matures and goes mainstream.
That "wait for GPU computing matures then jump on board" is what AMD has been quoting for some time.

The problem is that time is fast approaching or may even have arrived and because AMD is severely lacking on all the for-mentioned "relationships, education, software and infrastructure" the only thing their jumping will invoke is them drowning.
 
From what I've gleaned GK110 and Tahiti allegedly have similar DGEMM/peak ratios in the 80-85% range, with Tahiti just shy of 800 GF and N20 somewhere north of 1 TF.

Tahiti as we know it has a TDP of 250W, and some Tesla cards are in the same area, so I'll just say they're equal in a hand-wavey way for now.
Giving N20 the benefit of the doubt and saying it gets around 1100 GF as opposed to 1001, that gives N20 more than a third more DP FLOPS/W.

The perf/watt argument can override price if there are restrictions in density or operating budget due to power and cooling.

On a one to one chip basis, that puts N20 at an advantage in DP in terms of peak and power efficiency. The equation does flip if SP is involved. (hence N10, possibly)

The software and features argument could override both price and power. I'm curious what AMD can really offer over N20 in this time frame. There is at least the perception, and anecdotally some very negative ones at that, that AMD will need to overcome an "it's AMD" disadvantage as far as the GPU compute side goes.
 
Even the K20, will have hard time to be a lot faster of an actual 7970 in peak DP/SP performance.

The K20 is aimed at 1Tflops DP ... not a lot of faster of a simple 7970 ..[/SIZE][/FONT]
Do note that 1Tflops double precision mentioned in the PDF is not peak performance for K20...
 
Is there a 670 recall by EVGA or something else going on?
Link

Here is what a New Zealand PC component retailer said a few days ago:

"First shipment of EVGA GTX670 SC has arrived in the country. Supplier has informed us that EVGA have contacted them and placed a recall on those cards - Some sort of fault. It does not look like EVGA will get cards back here until next week. (And it will be the non SC version - so we have been told)"
 
Back
Top