This website requires many clicks. For intelligable graphs, click here

http://add-for.com/topics/benchmarks/caffe-vs-neon-vs-nvcaff...

Then make sure to change from stacked to grouped.

It makes me sad that stacked bar graphs are even an option here.

They are using different CPUs for these tests. It seems that the CPU used for Pascal has a higher GHz speed. I hope that isn't confounding on the results.

They're using different CPUs[0], different amounts of RAM, different base OSes, different drivers and even one different product version (Tensorflow RC). As far as benchmarks go this is pretty awful. It does show some useful top-level comparisons, but it's about as un-rigorous as you can get.

[0] http://ark.intel.com/compare/82930,94188,77780

The test setup for the 1080 also had half the RAM of the others, and used an older ubuntu.

Not a great test of GPUs in isolation.

I guess that sometimes you have to work with what you have and I prefer these results to no results.

I would assume that it definitely affects the result if the procedure involves moving data between GPU and main memory, but I do not know the size of the effect.

I am still pretty confident that the difference in performance comes mostly from differences between different GPU architectures.

Baidu Research also recently announced DeepBench: https://svail.github.io/DeepBench/

From eyeball perf/pricing, it looks roughly like the Pascal is 50% faster for twice the price. Which isn't surprising from the mass-market to specialized transition.

You're also paying for the Titan XPs 12GB of VRAM, which is overlooked by the common canned benchmarks that easily fit in the 1080s 8GB.

50% more RAM for 2x more. Same formula!

so ... 50% more RAM and 50% more speed for 100% more money?

honestly, this doesn't seem like a big issue.

Just bought a GTX 1070 mostly for reasons other than Deep Learning but it will be interesting to see how it stacks up compared to a 1080. Just hard to justify a 55% price increase for a 25% performance gain. Maybe SLI will get better in the future?

SLI is not for compute. SLI scaling has nothing to do with compute performance, unless you are hitting a PCIe bandwidth bottleneck or have to jump over QPI you can have 1:1 scaling with multiple GPUs.

What a shitty way to display the numbers. Combined Bar charts are not made for this.

You can switch to "grouped" above the charts to change it.

...and you could argue that it does make sense for showing how X fares across categories Y,Z,T... The height of the bar is proportional to the average performance across the different frameworks.

seriously...

Double speed 16bit inference on the Titan X too

Does anybody know what is being used at big companies like Facebook and Google?

Treating big companies as monoliths never works well. Different teams will inevitably choose to use different hardware, but the nVidia Tesla Kxx is popular among a lot of these "big four" types.

Likely not these GPUs unless it's for small scale projects.

You need a Tesla card for most enterprise scale compute features.