Question about GPU RTX 2080

[Johan Diaz Tovar]

Hello,

I would like to know if someone has used the new nvidia GPU GeForce RTX series, because we are interesting to purchase that specific GPU and we really want to know if it is going to work properly for MCX.

Thank so much for your time,

Best,

Johan.

[Qianqian Fang]

On 2/11/19 11:07 AM, Johan Diaz Tovar wrote:

Hello,

I would like to know if someone has used the new nvidia GPU GeForce RTX series, because we are interesting to purchase that specific GPU and we really want to know if it is going to work properly for MCX.

hi Johan

I ordered an RTX2080, hopefully will get it by Friday, will post my experience here.

You can also see a related discussion in a separate thread:

https://groups.google.com/forum/?hl=en#!searchin/mcx-users/titan$20v%7Csort:date/mcx-users/wG236-dJg6U/MmnNoal9GgAJ

Qianqian

Thank so much for your time,

Best,

Johan.

--
You received this message because you are subscribed to the Google Groups "mcx-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to mcx-users+...@googlegroups.com.
To post to this group, send email to mcx-...@googlegroups.com.
Visit this group at https://groups.google.com/group/mcx-users.
For more options, visit https://groups.google.com/d/optout.

[Qianqian Fang]

hi Johan

my 2080 (Gigabyte GAMING OC 8G ) has arrived today, and did some quick benchmarks. You can see the results at

http://mcx.space/gpubench/

just search for today's date "2019-2-22".

All tests were performed using the latest development branch of mcx (etherdome) on github. I compared 3 cards, 2080, Titan V, and 1080Ti. For 2080, I also tried 3 different cuda versions to see if there is any difference.

In comparison, 2080 performs somewhere between Titan V and 1080Ti, as expected. it is roughly 20% slower than TitanV, but about 50% faster than 1080Ti (with the latest version of mcx and nvidia driver). This is without any specific optimization for Turing, which we could certainly do.

One thing you may notice, is that there are 1080Ti benchmark entries that performed better than the 2080 scores, this is due to two things:

1. the previous 1080Ti tests were performed using different driver versions. CUDA 7.5 combined with driver versions between 375-381 gave the highest speed in the past. With the newer cuda and driver releases, mcx speed was reduced due to the compiler differences, see some previously discussions

https://devtalk.nvidia.com/default/topic/925630/cuda-programming-and-performance/cuda-7-5-on-maxwell-980ti-drops-performance-by-10x-versus-cuda-7-0-and-6-5/post/4994759/#4994759

2. the previously 1080Ti benchmarks were obtained from older releases of mcx, which had less features and different sets of optimizations.

If you revert the driver back to 37x series, I expect the 1080Ti speed to be closer to 2080 (won't be faster, but at most 10-20% slower). The current 50% speed difference is slightly in 2080's favor because of the newer driver support.

Hope this helps.

Qianqian

[Jeremy Ford]

Hi Dr. Fang,

I am glad that you have been able to test out the new RTX card. Past the simulation speeds, are you (or anyone else testing out multiple GPU architectures) able to comment on any changes in the "accuracy" of the results? Do you any significant changes in the output of the same simulation across GPU architectures?

Thank you for thoughts on this.

Jeremy

To unsubscribe from this group and stop receiving emails from it, send an email to mcx-...@googlegroups.com.

To post to this group, send email to mcx-...@googlegroups.com.
Visit this group at https://groups.google.com/group/mcx-users.
For more options, visit https://groups.google.com/d/optout.

--
You received this message because you are subscribed to the Google Groups "mcx-users" group.

To unsubscribe from this group and stop receiving emails from it, send an email to mcx-...@googlegroups.com.

[Qianqian Fang]

On 4/12/19 4:40 PM, Jeremy Ford wrote:

Hi Dr. Fang,

I am glad that you have been able to test out the new RTX card. Past the simulation speeds, are you (or anyone else testing out multiple GPU architectures) able to comment on any changes in the "accuracy" of the results? Do you any significant changes in the output of the same simulation across GPU architectures?

Thank you for thoughts on this.

hi Jeremy

My short answer to your question is "no", there is no significant impact to MCX simulation accuracy on different generations of GPUs.

here is why (the long version):

First, the "errors" in an MC simulation result are (strongly) dominant by its stochastic noise (shot-noise) due to limited photon numbers. The magnitude of such noise is so big that numerical round-off errors is insignificant. We have characterized the stochastic noise in our recently published denoising paper on JBO (see Section 2.3 and Fig. 4). Even for 10^8 photons, a big portion of the simulated volume has an SNR lower than 40dB (i.e. the std of the stochastic noise is >1% of the averaged intensity). In comparison, the round-off error typically appears on the 6th-7th digit of the data (single precision) which is orders of magnitude lower than the stochastic noise.

As I mentioned previously, my student Leiming Yu is current working on an expanded study to investigate the impact of storage precision to MCX output. The particular interests are to answer two questions 1) whether half-precision math can lead to faster speed without loosing accuracy, and 2) whether double-precision is necessary for MC simulations. Our answers are "yes", and "no". Will share more details about this study once once he completes the draft.

Second, the numerical "accuracy" (and also related, "repeatability") is not guaranteed in parallel computing, even with high-precision storage. The reason is that +/* are not associative anymore if operands are in limited precision, in other words - (a1+a2)=a3 is not the same as a1+(a2+a3). In parallel computing (particularly massively parallel ones), when accumulating values by multiple threads, the order matters, and typically, such order is not guaranteed and is governed by the scheduler. For MMC, if you run multiple threads, your result will always be slightly different; for MCX, the CUDA scheduler is surprisingly deterministic and produce nearly repeatable results, but if you change thread numbers or block numbers, the results will be different again. However, given point#1, even there is a small variation due to floating-point round off, it is too small to be worried compared to the stochastic noise.

Third, we have applied relaxed math in both of our CUDA (-use_fast_math) and OpenCL (-DMCX_USE_NATIVE) code (can be disabled). These flags do not guarantee IEEE 754 compliant. The magnitude of the errors produced by these flags can be found in their reference manuals. However, they are on the same scale as the numerical round-off error, and orders of magnitude smaller than the impact of changing the number of photons or seed.

So, in short, you don't need to worry about it.

Qianqian

To unsubscribe from this group and stop receiving emails from it, send an email to mcx-users+...@googlegroups.com.

[Jeremy Ford]

Thank you.  This is very helpful. Our group is buying a new GPU and the Nvidia RTX is a lower price point for more cores and equivalent video RAM compared to the GTX, so I am glad the configuration change shouldn't significantly alter the simulation output.

Regards,

Jeremy