4 ms·

Amazon is offering instances with 16 Tesla K80 GPUs. I don't know how that is realized under the hood, though.

wtallis 10y ago
SLI is only for realtime rendering with Direct3D and OpenGL. Tesla GPU products are intended primarily for CUDA and OpenCL use, which have always required the application to be aware of the underlying hardware configuration. There's no need to pretend that the 16 GPUs are one super-powerful accelerator device.
ansible 10y ago
Some of the high end Xeon systems have gobs of PCIe lanes. If you have a four CPU socket system, you can hang four x16 PCIe cards off of each processor.
- llukas 10y ago
  Or get specialized board with PCIe switch ie. http://www.cirrascale.com/blog/index.php/welcome-big-sur/ http://www.cirrascale.com/blog/index.php/welcome-big-sur/ so GPUs get better bandwidth (P2P etc).
- svacko 10y ago
  4CPU sockets !=> 4x x16 cars. It always depends on the number of PCI lanes for each processors. Check i.e. these 2 E7 Xeons: http://www.intel.com/content/www/us/en/products/compare-products.html?productIds=93811,93804 http://www.intel.com/content/www/us/en/products/compare-prod... Both have max 32 PCI Express lanes - so you can theoretically hang max 2 PCIx x16 cards off each CPU, but as few of the lanes is reserved for other peripherals, in 4GPU/CPU setup they will practically run at lower speeds, ie 8/8/4/4. However, using system with PCIe root complex [1] can improve the GPU<->GPU communication speed. [1] https://www.microway.com/hpc-tech-tips/pci-express-root-complex-confusion/ https://www.microway.com/hpc-tech-tips/pci-express-root-comp...
masklinn 10y ago
They're treated as independent compute units and the host program load-balances compute work between them. You can even do P2P memory copy or access between your CUDA devices without going through the CPU. See https://www.nvidia.com/docs/IO/116711/sc11-multi-gpu.pdf https://www.nvidia.com/docs/IO/116711/sc11-multi-gpu.pdf for a high-level overview.

SLI is only for realtime rendering with Direct3D and OpenGL. Tesla GPU products are intended primarily for CUDA and OpenCL use, which have always required the application to be aware of the underlying hardware configuration. There's no need to pretend that the 16 GPUs are one super-powerful accelerator device.

ansible 10y ago

Some of the high end Xeon systems have gobs of PCIe lanes. If you have a four CPU socket system, you can hang four x16 PCIe cards off of each processor.

llukas 10y ago
Or get specialized board with PCIe switch ie. http://www.cirrascale.com/blog/index.php/welcome-big-sur/ http://www.cirrascale.com/blog/index.php/welcome-big-sur/ so GPUs get better bandwidth (P2P etc).
svacko 10y ago
4CPU sockets !=> 4x x16 cars. It always depends on the number of PCI lanes for each processors. Check i.e. these 2 E7 Xeons: http://www.intel.com/content/www/us/en/products/compare-products.html?productIds=93811,93804 http://www.intel.com/content/www/us/en/products/compare-prod... Both have max 32 PCI Express lanes - so you can theoretically hang max 2 PCIx x16 cards off each CPU, but as few of the lanes is reserved for other peripherals, in 4GPU/CPU setup they will practically run at lower speeds, ie 8/8/4/4. However, using system with PCIe root complex [1] can improve the GPU<->GPU communication speed. [1] https://www.microway.com/hpc-tech-tips/pci-express-root-complex-confusion/ https://www.microway.com/hpc-tech-tips/pci-express-root-comp...

masklinn 10y ago

They're treated as independent compute units and the host program load-balances compute work between them. You can even do P2P memory copy or access between your CUDA devices without going through the CPU. See https://www.nvidia.com/docs/IO/116711/sc11-multi-gpu.pdf https://www.nvidia.com/docs/IO/116711/sc11-multi-gpu.pdf for a high-level overview.