Upcoming OpenMM version doubles fahcore17 performance

[JimF]

we might just be talking past each other, but my prediction is that a HD 5870 will not see the same percentage speedup as a AMD 7970.

The results of my lowly HD 7770s on about three OpenCL projects (only one of which was Core_17) seem to confirm that. I think the 7000 series does a lot better than the earlier ones; the new drivers and applications may help everyone, but they help the newer GCN architecture more.

And while we are on the subject, QRB helps the fast ones the most. So save your pennies for a high-end AMD card seems to be the message for the moment (rather than two low-to-mid-range cards as I usually buy), though it could be a somewhat different message next month.

[SodaAnt]

I am curious exactly what causes this magnitude of a speedup. Whatever it is though, good on everyone involved with OpenMM there!

[Rattledagger]

The speedup is shown using the examples of a GTX680 and a HD7970.
I hope it will speedup also the "older" HD5- and HD6-series . . .
A HD 5870 performing 25K PPD - awesome

While based on the "raw" performance, going by Wikipedia an old 5850 manages 2.088 TFLOPS while 7950 manages 2.867 TFLOPS, so the 7950 is roughly 1.4x faster...

But, in reality the difference is much larger. Using HCC1 as an example (this project will finish in roughly 23 days time so for anyone wants to test their hardware they'll need to do this ASAP), my old 5850 uses 9-11 minutes per wu, using around 95% of GPU, while the 7950 uses 8 minute/8 wu, meaning on average finish one wu per minute. Meaning, in HCC1, the actual difference is roughly 10x, nearly an order-of-magnitude larger than the theoretical difference. This clearly shows it's not just some "faster" hardware, but completely different performance from the hardware.

A 2x speedup in FAH due to OpenMM on 7970 will therefore likely also give a 2x speedup on 7950, but this tells absolutely nothing about performance on older cards like the 5xxx and 6xxx-cards.

[mdk777]

Well, you could try CAT 13.3 for latest open cl support.
but, as I mentioned, previous generation of cards were a different architecture, so the result doesn't surprise.

EDIT: OK, responding to now missing post.

I guess it was deleted for discussion of specific beta results outside of beta forum.

[bollix47]

The post was not deleted ... merely moved to the beta forum for the reason you guessed: viewtopic.php?p=240635#p240635

[Jesse_V]

Yep. Just to clarify, OpenMM 5.1 powers the version of FahCore 17 currently in beta testing. It is faster than the version powered by 5.0, but interestingly enough some classes of cards haven't seen a 2x speedup.

[PantherX]

Please note that there were many off-topic comments so I have created two different threads which are:

Implications of FahCore_17 on Current GPU WUs

New Information Sources For F@H