[AT] Nvidia releases GK210

[Enigmoid]

http://www.anandtech.com/show/8729/nvidia-launches-tesla-k80-gk210-gpu

Looks like Kepler is still going.

Reworked cache system. Doubled L1 from 64 KB to 128 KB, register file size to 512 KB from 256 KB.

Perhaps Big Maxwell is delayed.

 

[96Firebird]

300W passive cooler? Quite impressive...

Edit - Ah, it relies on external fans in the setup.

 

[NTMBK]

Continuation of the bifurcation between HPC, double precision focused compute board, and graphics oriented single-precision focused board. Makes sense I guess as they don't need to compromise on graphics perf/W for the gaming market.

 

[Borealis7]

300W passive cooler?

Passive cooling in the full sense of the word: someone else has to cool it

 

[witeken]

Not really fair that they compare it against a CPU. Not apples to apples comparison. They should compare it against Xeon Phi, but then the comparison would be less flattering...

Edit: and don't forget power consumption...

 

[nvgpu]

http://blog.xcelerit.com/intel-xeon-phi-vs-nvidia-tesla-gpu/

For this application it can be clearly seen that NVIDIA’s Tesla GPU outperforms both other platforms significantly, being 5.1x faster than the multi-core dual Sandy-Bridge CPU and 2.2x faster than the Xeon Phi (512K paths).
Moreover, compared to the sequential implementation, the optimized Sandy-Bridge is 19x as fast, the Phi is 43.5x as fast, and the Kepler GPU is 96x as fast.

The Tesla GPU is about twice faster than the Xeon Phi

Xeon Phi gets trounced by a single Kepler GK110 already and thats embarassing. GK210 with doubled register file and doubled shared memory/L1 cache would embarass it even more.

 

[III-V]

The timing of this seems rather strange... I hope this does not mean GM200 will be delayed. We need it to bring improved perf/dollar to the market.

Xeon Phi gets trounced by a single Kepler GK110 already and thats embarassing. GK210 with doubled register file and doubled shared memory/L1 cache would embarass it even more.

I don't recall it being trounced, but Xeon Phi is certainly showing its age.

 

[Arachnotronic]

Well, now we know what GPU will power the next Titan!

 

[III-V]

Well, now we know what GPU will power the next Titan!

Not so sure about that. The upgrade would be pretty marginal.

 

[ShintaiDK]

Not really fair that they compare it against a CPU. Not apples to apples comparison. They should compare it against Xeon Phi, but then the comparison would be less flattering...

Also dual GPU vs single CPU

 

[NTMBK]

Also dual GPU vs single CPU

Yeah, that was pretty hilarious "Look, we doubled our FLOPs! By putting a second GPU on the board"

 

[alcoholbob]

Isnt this just a Tesla Titan Z?

 

[NTMBK]

Isnt this just a Tesla Titan Z?

Nope. New GPU with some tweaks. Titan Z is GK110, this is GK210.

 

[witeken]

http://blog.xcelerit.com/intel-xeon-phi-vs-nvidia-tesla-gpu/



Xeon Phi gets trounced by a single Kepler GK110 already and thats embarassing. GK210 with doubled register file and doubled shared memory/L1 cache would embarass it even more.

Intel's competitor will arrive in 2015 with Knight's Landing.

http://www.anandtech.com/show/8217/intels-knights-landing-coprocessor-detailed

 

[NostaSeronx]

I hope this does not mean GM200 will be delayed.

It doesn't, GK210 is for high-end Tesla. GM200 is taking the value mid-end Tesla.

 

[xpea]

Intel's competitor will arrive in 2015 with Knight's Landing.

http://www.anandtech.com/show/8217/intels-knights-landing-coprocessor-detailed

and KL will face Pascal, another lost battle for Intel...

 

[witeken]

and KL will face Pascal, another lost battle for Intel...

Let's not get too far ahead of ourselves, shall we? (It's rather easy to invent performance and specs of unannounced SKUs on the spot, blatantly ignoring all details (including price) and nuances that the reviews will uncover.)

 

[f1sherman]

The timing of this seems rather strange... I hope this does not mean GM200 will be delayed.

Well they didn't release K80 over night.
GK210 has been in pipeline at least since Feb. 2012 (driver traces)
And it was first time spotted in April 2014 (Zauba), along with some GM200 parts.

and KL will face Pascal, another lost battle for Intel...

It's not lost - the HPC/data center market is expanding, there is room for everyone.
And I base this on nothing

OK found it:

 

[Riek]

Rather curious how long the boost clock will be reached for intensive usage.

The default core clock is about 1/3th less. (or ~1TFlop between core clock and boost clock)

 

[3DVagabond]

I wonder why no comparison to Firepro "S" series cards? Surely that would make more sense than comparing it to CPU's?

 

[RaulF]

I wonder why no comparison to Firepro "S" series cards? Surely that would make more sense than comparing it to CPU's?

Marketing i guess.

But good job Nvidia, keep it up.

 

[f1sherman]

I wonder why no comparison to Firepro "S" series cards? Surely that would make more sense than comparing it to CPU's?

This might be the reason:

http://www.nvidia.com/object/tesla-servers.html

Selects the GPGPU platform to be used, currently the only supported value is CUDA (in future OpenCL support will be added)

http://www.gromacs.org/Documentation/Installation_Instructions_4.5/GROMACS-OpenMM


So beside inventing the market and promoting competitor with whom they don't even want/need to compete with
(has little foothold in scientific community and HPC, not a real threat that they need to address),

what is NVIDIA supposed to do:

Rewrite all existing software for their platform too?

 

[el etro]

I sense some work done on improving efficiency of GK110 chip.

One thing i don't understand is this low clocks: why instead of clock the card low like this don't clock the card high and lower the price for the end user pick two of these?
Since i not understand so much Tesla cards, i cannot give a more accurate opinion.

 

[3DVagabond]

This might be the reason:

http://www.nvidia.com/object/tesla-servers.html

http://www.gromacs.org/Documentation/Installation_Instructions_4.5/GROMACS-OpenMM


So beside inventing the market and promoting competitor with whom they don't even want/need to compete with
(has little foothold in scientific community and HPC, not a real threat that they need to address),

what is NVIDIA supposed to do:

Rewrite all existing software for their platform too?

All that was needed was to state that the comparison was limited to CUDA. I made no suggestion that nVidia should rewrite the software for their competition. We'll save those arguments for the Mantle threads.

 

[f1sherman]

I sense some work done on improving efficiency of GK110 chip.

One thing i don't understand is this low clocks: why instead of clock the card low like this don't clock the card high and lower the price for the end user pick two of these?
Since i not understand so much Tesla cards, i cannot give a more accurate opinion.

Because,
if die size is almost a non-issue - and it is because K80 has effectively twice the area of GK210,
you can achieve better perf/W by packing twice the count of cuda cores and clocking them very low

Also you are creating new product, new choices, selling two GPU's per 1 unit, bragging rights etc etc.