4 ms·
GPU Platforms Set to Lengthen Deep Learning Reach
- amelius 11y agoBut why do we (they) keep on calling them GPUs? And why not improve their form factor? Currently, a GPU hardly fits into a server slot. And a typical desktop can host at most 2-3 of them. As a radical idea: why not design this type of auxiliary computing power as separate boxes that can be attached to a desktop or server computer by a cable?
- mrb 11y agoWe call them GPUs because they are built from GPU chips. The Tesla M40 for example is based on a GM200 chip which is the same as in a GeForce GTX Titan X GPU. Also Nvidia and other assemblers have done exactly that: separate boxes that you attach to servers by a PCIe cable. See for example the Tesla S2050 server: http://www.nvidia.com/docs/io/43395/nv-ds-tesla-s2050-june10-final-lores.pdf http://www.nvidia.com/docs/io/43395/nv-ds-tesla-s2050-june10... A good list of Tesla systems is at https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units#Tesla https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_proces... (which, coincidentally, I have been heavily updating these last few days).
- gok 11y agoSome older Tesla products were separate boxes: http://www.nvidia.in/object/product-tesla-S2050-in.html http://www.nvidia.in/object/product-tesla-S2050-in.html The form factor isn't a choice... The cards have a lot of stuff on them that need to go somewhere. For applications that need it, single servers can fit 8 GPUs these days. As for why they're still "GPUs"; there is nothing particularly "central" about modern CPUs either. One could argue that neural networks are a form of "graph" though.
- llukas 11y agoTesla M4 is what you are looking for - it is designed to fit sever slot.
- GFK_of_xmaspast 11y ago> attached to a desktop or server computer by a cable? Data bandwidth is already an issue in GPUs, and sending bits over a cable isn't going to help.
- p1esk 11y agoNvidia DIGITS DevBox is what you're looking for.
- p1esk 11y agoIt makes no sense to make "x10", or "x40" speedup claims, without mentioning a specific benchmark. I'm guessing the code that scales well with 200 or even 12 GPUs is fairly rare. For example, problems like ImageNet classification with convolutional NNs don't scale well beyond 2-4 GPUs. At some point, you will actually get a slowdown, because of the communication overhead.