3 ms·
If they can leverage the GPU on the Pis it could be quite a bit of compute power spread across that array. Even if they're purely CPU bound that is 4240 cores
by jandrese 7y ago
If they can leverage the GPU on the Pis it could be quite a bit of compute power spread across that array. Even if they're purely CPU bound that is 4240 cores which is not nothing.
Power wise each Pi going flat out burns around 5W, so all told you're burning ~6kW, which is quite modest for a big computing setup.
- willis936 7y agoIt looks like you can get 400 single precision MFLOPS for those 5W. So 80 MFLOPS/W +/- 20%. It isn’t the worst and it isn’t the best, but it is respectable. It would be a big hassle to make and maintain a cluster of thousands of these. https://www.arnabkumardas.com/platforms/nvidia/nvidia-jetson-nano-review-and-benchmark/ https://www.arnabkumardas.com/platforms/nvidia/nvidia-jetson... https://raspi.tv/2019/how-much-power-does-the-pi4b-use-power-measurements https://raspi.tv/2019/how-much-power-does-the-pi4b-use-power... https://en.wikipedia.org/wiki/Performance_per_watt https://en.wikipedia.org/wiki/Performance_per_watt
- ramses0 7y ago...but you could experiment with "disposable computing" (ie: similar to original google datacenter racks: never replace a node, only replace a rack at a time if 51% is failing).
- jacquesm 7y agoThat and interesting configurations (systolic arrays, message passing architectures, Erlang on a large number of nodes and so on). Lots of fun with stuff like this.
- willis936 7y agoOh jeez. Out of curiosity I did a comparison against a gaming GPU. A GTX 1080 uses 200W of power for 9 TFLOPS, which is roughly 500x better performance per watt than a raspberry pi. Outside of the fun of it, and maybe a cheap way to prototype larger systems, I can't see a serious reason to make a raspi cluster.