GTX 1080ti

http://www.anandtech.com/show/11172/nvidia-unveils-geforce-gtx-1080-ti-next-week-699

Some odd numbers for this GPU, in particular they seem to love the number 11;

11 GB RAM

11 Gbs DDR5X memory clock

352-bit bus width (32*11=352)

11.5 TFLOPS

Odd numbers indeed. Was that pun intentional or unintentional? There is an obvious connection between the first and the third item (11 channels comprising 1 GB each), not sure about the rest; maybe NVIDIA employs a hobby numerologist in the marketing department :-)

i guess they keep lucky 12 number for updated titan

Titan is dead, long live GTX Ti !!11!!!

Looks a bit like just showing off how their platform can handle these odd numbers, but thinking about it it indeed seems rather unlikely that consumer behavior model predicted extremal profits always align with powers of two.

Now that the Titan line doesn’t offer any real benefit over GTX, it seems time to reenable full double precision throughput again (i.e. GP-100-based Titan). Unless someone can think of another fused-off feature useful to CUDA programmers.

Double speed FP16 anyone?!

AMD’s upcoming consumer RX Vega GPU has FP16x2, (they call it RPM: Rapid Packed Math) and they demoed its utility in hair dynamics. If AMD shows there are more more actual consumer applications, maybe we’ll see FP16x2 appear in NVidia’s consumer cards too.

Interestingly, FP16x2 is in the Tegra X1 mobile part, but it’s unclear why.

It’s obviously no technical problem for NVidia to include FP16x2; it’s just an optimization problem of where the GPU’s transistors are best spent for the best user needs. GP100’s fp16x2 and fp64 transistors means there’s less available to spend on fp32 ALUs, which is why GP102 cards beat it for single precision even though GP100 has a larger die.

Im wondering Can the GT1080 Ti be considered as an alternative to the titan X pascal for deep learning? Im pretty new to CUDA / GPU processing and coming at this from a developers angle that builds neural networks. I’m unhappy paying out for server space constantly to get experiments turned around faster when i could really just take the plunge and buy some hardware (if it was the right price!)

I’m also a casual gamer so i thought i might be able to have a nice compromise piece of hardware in this at an acceptable price!

I have had a look around but most of the talk so far is from a gaming perspective rather than anything related to deep learning?

It appears that the GTX 1080ti will be faster than the Pascal Titan X, though without that extra 1 GB of memory and it probably will not be able to be put into Windows TCC driver mode. Other than that it will be great for 32-bit deep learning.

http://www.nvidia.com/download/driverResults.aspx/115887/en-us

http://www.nvidia.com/download/driverResults.aspx/115886/en-us

Drivers released for GTX 1080 Ti.

http://www.anandtech.com/show/11180/the-nvidia-geforce-gtx-1080-ti-review/14

Some compute benchmarks for GTX 1080 Ti.

Unfortunately the linked compute comparisons do not include Titan XP. I have only seen gaming benchmarks so far this morning, showing GTX 1080 Ti performance ahead of GTX Titan XP by 2%-5%. It seems reasonable to assume that results for compute tasks would be in a similar range.

If so, CudaaduC’s expectations in #9 would be fulfilled, although it is unlikely a single-digit percentage performance increase would be noticeable outside of careful measurements :-)

I may be interested in buying one of these, but first I want to know if these cards have the same bottleneck/performance/stutter/micro-stutter problems as GTX 970 with disabled chips and all that.

So I want to see some tests performed and tests results as was done with GTX 970.

No, because Nvidia disabled 1 32bit memory controller and the 256KB L2 cache on the GPU and removed the 1 memory chip from the card. Hence why it has 11GB of GDDR5X memory.

Tried to buy a GTX 1080ti at 10 AM per the invite and they were sold out in less than 2 minutes.
Guess I will wait for one of the EVGA custom versions.

Have any memory bandwidth benchmarks been done yet?

Seeing how nvidia now seems confident enough to increase the memory clock speeds, I’m really curious if any issues with GDDR5x for CUDA workloads have been resolved.