Wormius

[Wormius @gatowormius]

@GerardoOdriozo1 @icstcf2020 This happens because the algorithms don't Use any optimizations like neighbors list or anything else. It's a simple relative raw comparation that expose the potential of the gpu computing and the relevance of the grid and memory handle.

[Wormius @gatowormius]

Hello! I present you my work entitled "GPU-Acceleration of force decomposition in molecular dynamics" at
@icstcf2020 #ICSTCF20Poster #GPUprogramming #NvidiaCUDA #UAMI

Please visit https://t.co/8lNrYReCD8 to see an animated version.

[Wormius @gatowormius]

@Alpixels @icstcf2020 About the time, for example, the algorithm A* with 50000 take near 43s to calculate the force matrix , where 3s of this time are for data transfer (to device and from device).

[Wormius @gatowormius]

@Alpixels @icstcf2020 Thank you. The time in the graph shows the average time to finish the compute of the force matrix by gpu. In fact, this involve the data transfer between host and device, with the aim of getting a (almost rudimentary) comparation with the cpu algorithm.

[Wormius @gatowormius]

@ernie2k @icstcf2020 These programs used a basic Lennard-Jonnes potential. Due to this simplicity, we can scale easily to other applications. Even if the model doesn't allow optimizations as Newton's 3rd law.

[Wormius @gatowormius]

@ErickSa56721343 @icstcf2020 Tank you for your comment. Absolutely: The use of the appropriate memory spaces can improve considerably the algorithm performance. This happens too with every release of a new upgrade of the nvidia's graphics cards.

[Wormius @gatowormius]

An static version of the poster can be found here: https://t.co/qW856MyzYR

[Wormius @gatowormius]

Share this & earn points to win a MobiScribe Origin! https://t.co/0r5Zyj4fnn