2 ms·

all for open source, but these are two different products. from their website: > To fully utilize Meshroom, a NVIDIA CUDA-enabled GPU is recommended. The bina

by kmlx 5y ago

all for open source, but these are two different products. from their website: > To fully utilize Meshroom, a NVIDIA CUDA-enabled GPU is recommended. The binaries are built with CUDA-10 and are compatible with compute capability 3.0 to 7.5. Without a supported NVIDIA GPU, only "Draft Meshing" can be used for 3D reconstruction. versus an easy to use iphone app.

Aardwolf 5y ago
Somewhat off topic from the article, but I'm curious about GPU vs CPU computing here: So it looks like meshroom disables a feature (or features) when not having a Nvidia CUDA-enabled GPU, rather than e.g. still have the feature, but have it be slower. Any idea, if it were computed (or emulated) on CPU, how much slower it would be? (like 10x slower, 100x slower, 1000x slower, even more?!) And also, ignoring speed but instead looking at implementation effort, if it's feasible to emulate on CPU at all, or is GPU really so different for this that it's too much work to implement the same thing for a general purpose processor? At least matrix multiplication libraries (the main thing GPUs can do well afaik) exist for CPU too...
- dr_zoidberg 5y ago
  From my experience with deep learning I can eyeball it: one or two orders of magnitude. The fine details may be different on this particular problem. In DL in general, if you have sort of equivalent CPU/GPUs (think a 5600 vs GeForce 3070) you'll find yourself in the range of 10 to 30x faster training with GPU. If you have substantially different platforms, you can start looking at 100x speedups or more. For example, my mobile 4600H CPU trains about 100x slower than a desktop GeForce 2060. And close to 300x faster than the old Intel 7200u I use as a "slow system" to test things. Which also has a GeForce 940MX, that trains about 30x slower than the 2060 (if you manage to make your model small enough to fit in 2GB of VRAM).

by kmlx 5y ago

Aardwolf 5y ago

Somewhat off topic from the article, but I'm curious about GPU vs CPU computing here: So it looks like meshroom disables a feature (or features) when not having a Nvidia CUDA-enabled GPU, rather than e.g. still have the feature, but have it be slower. Any idea, if it were computed (or emulated) on CPU, how much slower it would be? (like 10x slower, 100x slower, 1000x slower, even more?!) And also, ignoring speed but instead looking at implementation effort, if it's feasible to emulate on CPU at all, or is GPU really so different for this that it's too much work to implement the same thing for a general purpose processor? At least matrix multiplication libraries (the main thing GPUs can do well afaik) exist for CPU too...

dr_zoidberg 5y ago
From my experience with deep learning I can eyeball it: one or two orders of magnitude. The fine details may be different on this particular problem. In DL in general, if you have sort of equivalent CPU/GPUs (think a 5600 vs GeForce 3070) you'll find yourself in the range of 10 to 30x faster training with GPU. If you have substantially different platforms, you can start looking at 100x speedups or more. For example, my mobile 4600H CPU trains about 100x slower than a desktop GeForce 2060. And close to 300x faster than the old Intel 7200u I use as a "slow system" to test things. Which also has a GeForce 940MX, that trains about 30x slower than the 2060 (if you manage to make your model small enough to fit in 2GB of VRAM).