Nvidia BigK GK110 Kepler Speculation Thread

[Ailuros]

Well, they did claim they're on schedule for decemeber release on the GK110 Teslas, which is just over month to 2 months away now, so fairly close

tviceman said GeForce and not Tesla. As it seems there won't be any desktop variants with GK110 within this year. IF they'll release someting relevant in Q1 2013, then the first leaks should start around December or even January.

 

[UniversalTruth]

Yup, that's correct. As far as I remember, there was information about possible launch in March 2013.

 

[A1xLLcqAgt0qc2RyMz0y]

Titan Sets High Water Mark for GPU Supercomputing

In the conversion from Jaguar, a Cray XT5, ORNL essentially gutted the existing 200 cabinets and retrofitted them with nearly ten thousand XK7 blades. Each blade houses two nodes and each one of them holds a 16-core Opteron 6274 CPU and a Tesla K20 GPU module. The x86 Opteron chips run at a respectable 2.2 GHz, while the K20 hums along at a more leisurely 732 MHz. But because to the highly parallel nature of the GPU architecture, the K20 delivers around 10 times the FLOPS as its CPU companion (Using the 27 peak PF value for Titan, a back-of-the-envelope calculation puts the new K20 at about 1.2-1.3 double precision teraflops.

http://www.hpcwire.com/hpcwire/2012...mark_for_gpu_supercomputing.html?featured=top

 

[Gipsel]

Or to quote myself from the other Kepler thread:

According to HPCWire, the Teslas run at 732 MHz. Taking the claimed 27 PFlop/s theoretical peak and factoring in the CPU flops (2,63PFLOP/s for base clock, would be 2,99 PFlop/s for all core turbo clock), it is a 14 SMx GK110 version.

14 SMx
732 MHz
6 GB GDDR5 @ 384Bit memory interface (clock still unknown)
1.31 TFlop/s theoretical peak

It's probably the top version of the K20 line if one compares it with the one that was on sale in a shop (which had only 13 SMx and a 320Bit memory interface and only 1.17 TFlop/s peak).

What I find interesting, is that the XK6 nodes obviously gets upgraded to XK7 ones in the process. Originally it was planned that the Tesla cards are just drop in extensions for the new XK6 nodes (it should have worked as Jaguar just got upgraded from XT5 to XK6). What is strange, is that HPCWire claims a max power consumption of 12.7 MW (up from prior 10.8 MW afaik), while the specs from Cray say only 54.1 kW per rack, same as with XK6. No idea what to make out of that.

 

[tviceman]

14 SMX's with the all the memory controllers sounds like yields are decent and bodes well for the GK110 Geforce card!

 

[iMacmatician]

Anyone have estimates for TDP?

 

[tviceman]

Anyone have estimates for TDP?

I can't find the link at the moment, but I remember 225 watts being cited.

 

[iMacmatician]

For the (presumably) 14 SMX variant?

 

[Blazkowicz]

I would presume 13 SMX get it easily in 225W, while 14 SMX can be in higher power custom designs. The intent may be for you to get rack units made by some companies, while the card is more affordable. Two or more card variants are possible and probable, too.

 

[Ailuros]

It comes more down to frequencies then enabled/disabled cluster amounts on a chip.

 

[iMacmatician]

From Bright Side of News: "Titan Comes to Life: 46 Million Nvidia CUDA Cores, 300,000 AMD x86 Cores."

Their numbers work out to 2496 CCs and 5 GB GDDR5 per GPU.

 

[Gipsel]

From Bright Side of News: "Titan Comes to Life: 46 Million Nvidia CUDA Cores, 300,000 AMD x86 Cores."

Their numbers work out to 2496 CCs and 5 GB GDDR5 per GPU.

Just multiplying numbers with the specs of the lesser K20 version coming from that shop doesn't count as source of information in my book. For starters, Titan has definitely the 6 GB version with the full memory interface and HPCWire explicitly claimed a clock frequency of 732 MHz (instead of 705 MHz).

 

[tviceman]

From Bright Side of News: "Titan Comes to Life: 46 Million Nvidia CUDA Cores, 300,000 AMD x86 Cores."

Their numbers work out to 2496 CCs and 5 GB GDDR5 per GPU.

Just multiplying numbers with the specs of the lesser K20 version coming from that shop doesn't count as source of information in my book. For starters, Titan has definitely the 6 GB version with the full memory interface and HPCWire explicitly claimed a clock frequency of 732 MHz (instead of 705 MHz).

I brought this up on anandtech and the mods there confidently said hpcwire is much more accurate (they take the time to verify their info) than websites like brightsideofnews and semiaccurate. I'll believe the mods and hpcwire until proven otherwise.

 

[A1xLLcqAgt0qc2RyMz0y]

Or to quote myself from the other Kepler thread:

14 SMx
732 MHz
6 GB GDDR5
1.31 TFlop/s theoretical peak

Ryan Smith of AnandTech agrees with you.

http://www.anandtech.com/show/6421/...-x86-cores-and-186k-nvidia-gpu-cores#comments

We're basing our numbers off of the figures published by HPCWire.

http://www.hpcwire.com/hpcwire/2012...mark_for_gpu_supercomputing.html?featured=top

For a given clockspeed of 732MHz and DP performance of 1.3TFLOPs, it has to be 14 SMXes. The math doesn't work for anything else.

 

[iMacmatician]

From SemiAccurate: "Nvidia Tesla K20 specs gives hints about 28nm yields."

Posted for discussion purposes only, since he completely ignores the 14 SMX GK110 at ORNL.

 

[homerdog]

Yay for posting an article about NVIDIA written by Charlie.

 

[iMacmatician]

And also from the thread at SA, which puts a spin on the "two variants of K20" theory:

Did you ever consider that supercomputers may get parts that are not the same as you can buy off the shelf? Think about what that might mean in the context of poor yields. Hint: If a company suddenly needs 20K units at one bin above where things are yielding, and is contractually obliged to do it after screwing the pooch on the last generation, what might the result look like?

Discuss.

 

[A1xLLcqAgt0qc2RyMz0y]

Charlie is a hack

And also from the thread at SA, which puts a spin on the "two variants of K20" theory:

Originally Posted by charlie
Did you ever consider that supercomputers may get parts that are not the same as you can buy off the shelf? Think about what that might mean in the context of poor yields. Hint: If a company suddenly needs 20K units at one bin above where things are yielding, and is contractually obliged to do it after screwing the pooch on the last generation, what might the result look like?

Discuss.

Charlie is a hack. His articles start with a conclusion he wants then he massages the data (or ignores it completely as in the 14 SMX 732 MHz K20) to make his per-conceived conclusions believable.

If Nvidia can provide 18.6K T20's that have 14 SMX's working at 732 MHz then Nividia's yields are fine.

 

[jimbo75]

If Nvidia can provide 18.6K T20's that have 14 SMX's working at 732 MHz then Nividia's yields are fine.

And you know how many wafers it took to get them, obviously.

 

[A1xLLcqAgt0qc2RyMz0y]

And you know how many wafers it took to get them, obviously.

Rebuttal to Charlie from tviceman

Yet the super computers got parts inferior to what us mundane desktop users got when Fermi first arrived. First gen Fermi Tesla was 440 cores, 384 bit bus while Geforce was 480 cores and 384 bit bus. I have no doubt we'll be seeing 12 & 13 SMX Geforce GK110's; GF110 had 4 total iterations (GTX580, GTX570, GTX560 448 ti, GTX560 OEM). But in addition to those cut down chips, I also believe we'll be getting 14 & 15 SMX Geforce chips as well. I have said myself many times the 15 SMX part will likely not come until sometime the end of summer. Wasn't ORNL's schedule to compete Titan March 2013? Wouldn't that mean getting the Teslas now put them ahead of schedule? Do you think Nvidia would pump out 3,500+ GK110 wafers with god awful yields just for the sake of ORNL being 4 months ahead of schedule?

You have consistently claimed doom and gloom for Nvidia when it's new chips are either in development or about to arrive. And to your credit you were right with some things about Fermi as well as with Kepler. But Nvidia's gross margins have steadily risen since a quarter or two before Fermi came out, so I'm inclined to believe not all is doom and gloom with each new Nvidia chip. Nvidia hasn't provided any updated guidance on it's upcoming quarterly earnings, but Intel warned ahead of it's earnings, as did AMD. And remember, Nvidia revised their earnings UPWARD during their last conference call. I think if GK110 was yielding god awful, and the rest of Kepler was also yielding god awful, they wouldn't have been looking up and up with the last two reports and would have warned by now.

http://semiaccurate.com/forums/showpost.php?p=170827&postcount=13

I agree. If yields were terrible on GK110 then there would be a hit on gross margins. No hit so no yield problems.