eGPU is quite portable and similar price to conventional graphics card. I am wondering if the CPU-GPU connection is fast enough to write down output on hard-disk and not make so big delay.
My question: is it applicable for MD simulation using Gromacs or Amber???
I haven't personally tried, but by looking at gaming benchmarks (check egpu.io) you can expect 25-5% performance drop just by using the ThunderBolt interface (assuming Thunderbolt 3, the fastest available).
Interestingly, performance drops decrease when the benchmark resolution is higher. For example, at 1080p, it drops at about 20-25%, but at 4K the drop is simply 5%. Take it with a pint of salt, though. MD might require different calculations and bandwidth usage.
If you are still willing to try, make sure your laptop has direct TB3->CPU connection. Some share the bus with other devices like WiFi and so on. Check this discussion as an example: https://egpu.io/forums/pc-setup/gtx-1080-ti-egpu-or-built-in-gtx-1080/#post-10570
That said, I wouldn't recommend eGPUs for MD simulations. Portability would stop being an advantage if you have to leave the eGPU connected almost permanently for the sims to run. And if it is for sporadic usage, maybe you can rent GPUs online. This option is getting increasingly cheap: https://vectordash.com/
GROMACS still separates the calculations between the CPU and the GPU. This means moving a lot of data between the GPU and the CPU, which could mean that thunderbolt becomes a bottleneck. Although the larger the system, the less important the bottleneck should be.
The molecular dynamics package OpenMM can on the other hand run the entire simulation on the GPU. This also is likely to be important when using MD on multi-GPU stations.
- Badges
- Science topic