From: Pawel Kedzierski (pawel.kedzierski_at_pwr.edu.pl)
Date: Thu Oct 11 2018 - 06:13:47 CDT

Thank you John for the very informative answer.
Greetings,
Pawel Kedzierski

W dniu 10.10.2018 o 17:21, John Stone pisze:
> Hi Pawel,
> I'll address the RTX and Turing parts of your question separately
> since they're independent but related topics.
>
> RTX refers to the new ray tracing acceleration (thus the 'R', 'T', 'X')
> hardware found in Turing and subsequent RTX-capable GPUs. Presently,
> the only part of VMD that will benefit from RTX are the TachyonL-OptiX
> rendering engines built-into VMD.
> The RTX hardware I've benchmarked to date can give VMD as much as
> 6x to 8x performance gain when compared with the previous Volta GPU hardware,
> and Volta was (I forget, maybe 1.3x?) faster than previous Pascal hardware.
>
> I've been testing pre-production RTX GPU hardware and RTX-enabled OptiX
> and driver software since early August. That said, there have been
> some unforeseen delays in the availability of public versions
> of some of the RTX software components that VMD uses, which are going
> to delay the availability of RTX-enabled VMD builds to at least the
> end of this year, so I am now planning on releasing the non-RTX VMD 1.9.4
> builds after a short beta testing period. Once the RTX software components
> from NVIDIA can be made publicly available, I will then release RTX-enabled
> VMD 1.9.4 builds and/or a new version of VMD that adds further RTX features
> beyond what I have built-into VMD 1.9.4 today. The VMD 1.9.4 source code
> already contains RTX support, so my expectation is that as soon as
> NVIDIA makes the software components publicy available, anyone would
> be able to build RTX-enabled VMD binaries from the source.
>
> With regard to the non-RTX features of Turing as compared with Pascal,
> I think that the greatest advances in Turing for general purpose use
> in VMD are the increased memory bandwidth (~600 GB/sec, GDDR6)
> and increase arithmetic throughput. There is an informative set
> of tabulated performance data available in the Turing White Paper here:
> https://www.nvidia.com/content/dam/en-zz/Solutions/design-visualization/technologies/turing-architecture/NVIDIA-Turing-Architecture-Whitepaper.pdf
>
> The Turing hardware has much better integer arithmetic performance than any
> GPU NVIDIA has made to date, so I'm expecting it to perform very well
> on some of VMD's CUDA-accelerated cryo-EM density map processing algorithms,
> e.g. image segmentation, cross-correlation, rigid body fitting of structures
> with maps, and general density map manipulation operations.
>
> Down the road further when we start getting OpenXR and native
> HMD support into VMD, I expect that the special single-pass stereoscopic
> rendering features of Turing will become very beneficial for using
> VMD with VR HMDs. I believe that those features are also described
> in the architecture white paper document PDF above.
>
> Best regards,
> John Stone
> johns_at_ks.uiuc.edu
>
>
> On Wed, Oct 10, 2018 at 10:55:13AM +0200, Pawel Kedzierski wrote:
>> Dear All,
>>
>> I'm about to buy new hardware and I'm already aware that the new VMD
>> will support real-time raytracing (aka RTX) on Turing GPUs, as
>> announced in this message:
>> https://www.ks.uiuc.edu/Research/vmd/mailing_list/vmd-l/30255.html
>>
>> Are there any other advantages of the new Turing GPUs for VMD use
>> and CUDA calculations, in comparison to Pascal? As my budget is
>> fixed I want to make an fair balance between number of cores, RAM
>> and GPU capability.
>>
>> Thank you,
>>
>> Pawel