Geforce GTX Titan 6GB

[ET_]

Originally Posted by sdbardwick

And they de-Nerfed the FP64 abilities.
Titan: 1/3 FP32
680: 1/24 FP32
580: 1/8 FP32

Now, that's interesting...

Luigi

[Prime95]

Originally Posted by ET_

Now, that's interesting...

Maybe. Should make no difference to mfaktc/mmff which relies on integer multiplies. It might make a big difference to CUDALucas. Or it might make little difference if CUDALucas is mostly bottlenecked on memory bandwidth. Benchmarks would be nice.

[Ethan (EO)]

Changes from the 680 generation of some relevance to various gpu efforts here:

"Space and bandwidth for both the register file and the L2 cache have been greatly increased for GK110. At the SMX level GK110 has 256KB of register file space, composed of 65K 32bit registers, as compared to 128KB of such space (32K registers) on GF100. Bandwidth to those register files has in turn been doubled, allowing GK110 to read from those register files faster than ever before. As for the L2 cache, it has received a very similar treatment. GK110 uses an L2 cache up to 1.5MB, twice as big as GF110; and that L2 cache bandwidth has also been doubled."

And memory access patterns outside of those handled well by the other caches and memory now have a cache home:

"...it’s also worth noting that NVIDIA has reworked their texture cache to be more useful for compute. On GF100 the 12KB texture cache was just that, a texture cache, only available to the texture units. As it turns out, clever programmers were using the texture cache as another data cache by mapping normal data at texture data, so NVIDIA has promoted the texture cache to a larger, more capable cache on GK110. Now measuring 48KB in size, in compute mode the texture cache becomes a read-only cache, specializing in unaligned memory access patterns."

New low level instructions are introduced:

"NVIDIA has added a number of new instructions and operations to GK110 to further improve performance. New shuffle instructions allow for threads within a warp to share (i.e. shuffle) data without going to shared memory, making the process much faster than the old load/share/store method. Meanwhile atomic operations have also been overhauled, with NVIDIA both speeding up the execution speed of atomic operations and adding some FP64 operations that were previously only available for FP32 data."

And, very significantly, Compute Model 3.5 adds the ability for GPU kernels to launch other GPU kernels! For many compute tasks the latency of host-gpu-host-gpu transfers is a big issue, further complicated by interactions between compute scheduling and host communication. "Dynamic Parallelism" as introduced on the GK110 means that work units can be generated and dispatched by processes running entirely on the GPU with correspondingly much lower overhead. For certain tasks this will make a big difference.

And, yeah there's the significant detail of FP64 speed back to a more reasonable 1/3 of SP speed if you downclock to stay within thermal limits. Practically speaking this means > 1Tflop/s DP on a single card, with 6GB of 288GB/s memory and a huge register file. So this can indeed be very fast for some tasks.


-Ethan

[Ethan (EO)]

It should also be remembered that the GK110 design was informed by contract requirements for a certain customer order that ate the first ~19000 boards off the line:

http://www.olcf.ornl.gov/titan/

And there are several academic papers out which examine GK110 performance characteristics and optimization demands:

http://scholar.google.com/scholar?q=...&as_sdt=0%2C48

[TheJudger]

Originally Posted by Ethan (EO)

Changes from the 680 generation of some relevance to various gpu efforts here:

"Space and bandwidth for both the register file and the L2 cache have been greatly increased for GK110. At the SMX level GK110 has 256KB of register file space, composed of 65K 32bit registers, as compared to 128KB of such space (32K registers) on GF100. Bandwidth to those register files has in turn been doubled, allowing GK110 to read from those register files faster than ever before. As for the L2 cache, it has received a very similar treatment. GK110 uses an L2 cache up to 1.5MB, twice as big as GF110; and that L2 cache bandwidth has also been doubled."

BUT a SMX in Kepler consists of 192 cores while a SM on Fermi has only 32 cores. So twice the size and twice bandwidth the to the register file for SIX times compute cores...

Oliver

[msft]

https://developer.nvidia.com/sites/d...erformance.pdf
FFT Performance

[Prime95]

http://www.anandtech.com/show/6774/n...nce-unveiled/3

Looks like 2.5x faster than a GTX 580 (for CUDALucas)

[fivemack]

GK110 apparently has full-precision FMA in DP; is that enough to make it reasonable to do double-precision arithmetic with FP rather than integer operations, or does the smaller number of DP units still bite you?

[Redarm]

my 680 in 2 weeks for sale (serious)

[nucleon]

Check out the FFT dp benchmark:

http://www.anandtech.com/show/6774/n...nce-unveiled/3

This card looks like to be one to use for LL testing.

-- Craig

[chalsall]

Originally Posted by nucleon

This card looks like to be one to use for LL testing.

Yeah... I read the whole article after George pointed us to it.

It's interesting... This SKU almost seems like it's targeted more to Compute than Gaming. Gamers are bemoaning the $1000 MSRP, but at approximately 1/7th the price of a Tesla K20X I suspect many will be more than happy to forgo ECC memory.

I can't wait to see some benchmarks from those here who have the budget (and/or the connections) to see how our favorite programs preform on this....