Folding Forum

The new version is broken. Is freezes my whole system and I have to reset it! Tested on i5-3570k with onboard HD4000 and newest SDK/driver. Also the Intel typo isn't fixed in this version.

I did inform Proteneer about the issues and he said that he will look into them. Nonetheless, thanks for your report.

vidia GeForce GTX 670 SC @ Stock
OpenCL Explicit DP -> 5.25485 ns/day
OpenCL Implicit DP -> 6.89259 ns/day

mrk wrote:The new version is broken. Is freezes my whole system and I have to reset it! Tested on i5-3570k with onboard HD4000 and newest SDK/driver. Also the Intel typo isn't fixed in this version.

Hi mrk, which drivers do you have? The HD4000 isnt' support at the moment due to driver issues. I will test on my end as well. I tested it on my i7 3770K and it seems to be OK.

EDIT: updated spreadsheet.

For fun, here's a Excel Web APP spreadsheet I created to estimate GPU QRB points based on FAHBench v1.2.0 CPU/GPU results, once core_17 gets released to full FAH. Hopefully the spreadsheet is more self-explanatory now. Take note of the disclaimer though. It's just a prediction which isn't tied to the official SMP benchmark machine. Furthermore, it has been mentioned somewhere that FAHBench results should not be related linearly to live GPU QRB - once it is in place.

I'd like a confirmation from proteneer himself on that, though. What's the point of "official FAH benchmark" if it bears only vague relationship to live WUs? Tech demo only?

PS:
PantherX,
my 018.8992 ns/day OCL SP implicit result is for the exact same NVidia GT430 card & clocks & conditions as the 003.87542 ns/day OCL SP explicit result. Please insert GT430 to the appropriate place, next time you update the OP. TIA.

something similar was asked at evga forums.

The answer was that we are comparing apples to oranges. The smp client running gromacs is very mature and efficient. With FAHBENCH, you are comparing OpenCL run on the cpu to OpenCL run on the GPU. What we need to accurately predict points is a comparison of the SMP gromacs run on the cpu to OpenCl run on the GPU...something we (as the community) don't have access to right now.

Yep, it is apples-to-oranges! I wrote something to that effect in the spreadsheet disclaimer, and indeed, that 2.5x GROMACS / OCL ratio came from the EVGA forum topic. So, take with a grain of salt.

EDIT:

Except that the 2.3 - 2.5 comment came from proteneer himself. So...

EDIT2:

Look a bit closer at the spreadsheet. The parameters are educated guesses, but the emphasis is on the word educated. Unless there's some terrible bug in the formulae. In that case, I'd appreciate if someone would point it out.

OK, no problem....I came to a 400k ppd back of the envelope calculation myself for a AMD 7970.
EDIT:

A 30 ns/day GPU (Titan or fast AMD GPU) should net about 150 000 PPD.

see, I was confabulating explicit with implicit.

Like you, I see the improvement to date as tremendous...but still not seeing the equal pay for equal work paradigm coming to fruition yet.
Always hope.

My spreadsheet IS the "equal pay for equal work" prediction based on FahBench results (OCL), PantherX's strictly unofficial "reference CPU" and SMP values, and a GPU - any GPU - with x ns/day.

Linear SMP -> GPU TPF scaling may not be fair, but then again, it's a prediction. Core_17 isn't out yet. Download the spreadsheet, inspect the formulae carefully and *then* provide criticism.

It *can't* be anything but apples-to-oranges at this point. Please keep in mind that this was just a though excercize.

Peace, please...

Explicit vs explicit (CPU SMP vs GPU ZETA) FAHBench is fair, just like it was said in the EVGA forums.

Edited a spreadsheet comment to reflect that.

EDIT: thanks for the feedback. I should've been more explicit in the disclaimer comments of the spreadsheet.

proteneer wrote:

mrk wrote:The new version is broken. Is freezes my whole system and I have to reset it! Tested on i5-3570k with onboard HD4000 and newest SDK/driver. Also the Intel typo isn't fixed in this version.

Hi mrk, which drivers do you have? The HD4000 isnt' support at the moment due to driver issues. I will test on my end as well. I tested it on my i7 3770K and it seems to be OK.

Tested with drivers 9.18.10.3071 and 9.18.10.3131. What driver are you using?

First post updated and corrected.

mrk wrote:...Tested with drivers 9.18.10.3071 and 9.18.10.3131. What driver are you using?

I don't have any issues running it on my Intel i7-3840QM by using driver version 9.18.10.3071 (https://downloadcenter.intel.com/Detail ... 5&lang=eng). However, it does crash whenever I try to run it on my iGPU HD 4000 which is a driver issue as stated by proteneer. Is your CPU overclocked or not? If so, try to reduce the overclock and run it again.

I'm talking about my onboard graphics HD4000. You don't need a graphics driver for your CPU.

I was under the impression that the GPU Drivers provided support for OpenCL on Intel CPU/iGPU thus is needed. So if no iGPU drivers are installed, the application would still run on the CPU using OpenCL?

There's a miscommunication here. There is no support for Intel iGPUs at this point. All the discussion in this thread referencing Intel pertains to running FAHBench on the CPU.