Nvidia Titan V

[foldy]

It would be interesting to have also a FahBench beta with the Core_22 so we could test its performance on a wide variety of user GPUs even before the real core gets released as beta.

There are also more 11xxx work units out now, so this could give the Titan V top PPD

[AndyE]

I just received my copy of the recently released 6th edition of Computer Architecture: A Quantitive Approach. For your information, Chapter 4 covers over 70 pages "Data-Level Parallelism in Vector, SIMD and GP Architectures". Nice writeup of the evolution since the 5th edition covered the then new and emerging GPGPU model 10 years ago. Some updated information about Intel's AVX512 and NVidia's Pascal architecture. Also a quick explanation by NVidia of the evolution of the native instruction sets of new GPUs underneath the PTX instruction set.

Interestingly, they ran (identical, recompiled) code from a paper, Intel researchers published in 2010 where they compared the Core i7-960 CPU with the GTX 280. Hennessy and Patterson took those comparisons forward and compared them to contemporary CPU'S (Intel Platinum 8180 with 28 cores) and NVidia's P100 Pascal architecture. Their based on those exampels, they concluded, that the speed difference between CPU and GPU now vs. back then got rather narrower than wider (CPU performance progressed faster than GPU's).

The book: https://www.elsevier.com/books/computer ... 2-811905-1
The 2010 Intel paper: http://sbel.wisc.edu/Courses/ME964/Lite ... PU2010.pdf

Would be interesting to see if the "good ol' days" of CPU driven big WU's might come back . (Probably not, as the comparison was not including economic efficiency. Both parts compared are approx in the 10k USD range and a 1080Ti has a better ratio to a price-comparable Intel CPU)

Nevertheless, allways nice to look things up in the seminal CA-QA book.
Andy

[foldy]

Scientific software gave an error in some of the Titan V. A molecular dynamics package called Amber – that reportedly is affected by the Titan V weirdness.
https://www.theregister.co.uk/2018/03/2 ... ucibility/

[Nathan_P]

Amber cores have not been used by F@h in years, I started in 2008 and they had already been retired then, also amber was designed for cpu, if someone has adapted for gpu then anything could have happened.

[bruce]

If you're getting FAH error messages about amber, what project were you folding?

[foldy]

I did not get errors in FAH, I just found this article about someone having problems in his software using Amber and Titan V, maybe a driver issue.

[artoar_11]

Amber cores have not been used by F@h in years, I started in 2008 and they had already been retired then, also amber was designed for cpu, if someone has adapted for gpu then anything could have happened.

http://ambermd.org/gpus/index.htm#

GPU accelerated PMEMD has been implemented using CUDA and thus will only run on NVIDIA GPUs at present although we are working on supporting AMD GPUs. Due to accuracy concerns with pure single precision the code uses a custom designed hybrid single / double / fixed precision model termed SPFP ........

Mar 2018: Titan-V reliability concerns. We have received conflicting reports about Titan-V cards failing the validation tests. Early reports suggested problems, but many subsequent tests have failed to reproduce this. You should probably treat Titan-V cards with care, making sure that repeated runs give identical results; but this is also good advice for all GPU cards: we know of examples where brand new cards failed, and also of cases where failures began to show up after years of use.

Mar 2017: GTX-1080TI GPUs successfully validated with AMBER. A single GTX-1080TI GPU provides an incredible 624.78 ns/day for the DHFR NVE 4fs benchmark. At $699 per GPU this is a price/performance ratio of $1.12 per ns/day. Compare this with the 16GB Tesla P100 which retails for ~$6000 and achieves only 568.10 ns/day on the same benchmark, a price performance ratio of almost 10x worse at $10.56 per ns/day. 4 and 8 GPU GTX-1080TI based GPU nodes are in stock and available, as turnkey AMBER certified systems with 3 year warranties, from Exxact Corp.

Fees:
Academic/non-profit/government: $500 (US dollars).
Industrial (for-profit): $20,000 for new licensees, $15,000 for licensees of Amber 14.

[bruce]

Amber does provide a set of simulation tools that has been widely used. So does OpenMM.

From what I've read, it's not clear whether they've found a problem that somehow is related to using Amber force fields, from building a molecular model using constructs using Amber terms, from simulating using packages provided by amber. It's also not clear whether it has anything watsoever to do with Amber, but rather that the Titan V has a hardware defect or the CUDA API has a software defect or there's a undiagnosed driver bug. Until additional testing begins to differentiate between those and other possible bugs that are actually related to the reports, we actually don't know much. Experts everywhere are attempti9ng to sort this all out.

OpenMM (used for the GPU simulations) does provide a turnkey version of at least one Amber force-field but there is actually up to the project owner to choose from a wide variety of forcefields -- most with more obscure names. OpenMM does provide a script that converts a model written using Amber terms to a model that's operable in OpenMM. (They also provide conversion scripts for CHARMM Files, for GROMACS files, an probably for others.)