3 ms·
As far as accelerators go, NVidia CUDA architecture seems to be the way. The Power9 Summit Supercomputer at Oak Ridge National Labs really is just an architectu
by dragontamer 9y ago
As far as accelerators go, NVidia CUDA architecture seems to be the way. The Power9 Summit Supercomputer at Oak Ridge National Labs really is just an architecture to communicate more quickly with the massive number of NVidia cards that they're shoving into the machine.
I'm a bit bearish on the Xeon Phi myself. The rival supercomputer: Aurora Supercomputer at Argonne National Labratory, was supposed to be Xeon Phi based. But its late! Intel and Cray dropped the ball.
At least in the near future, it means that Summit (Power9 + NVidia CUDA) is beating out Intel (Xeon + Xeon Phi) with regards to USA Supercomputers.
- rbanffy 9y agoMy desire for the Xeon Phi comes from the hunch programming it feels a lot like programming a top-of-the-line Xeon Platinum from 2028. Core counts have been steadily rising. And since the latest models even have virtualization (which I need to work) so, at least for me, Phi is just one Intel GPU away from the perfect desktop workhorse. Seymour will have to forgive me, but I think we're all plowing our fields with chickens now. That and my side project of putting a dozen Octavo SoCs on a PCB along with a ethernet switch chip to build a single-board cluster. If only Intel had an Atom part with ethernet, RAM and USB in a single package...
- dragontamer 9y ago>> That and my side project of putting a dozen Octavo SoCs on a PCB along with a ethernet switch chip to build a single-board cluster. If only Intel had an Atom part with ethernet, RAM and USB in a single package... What's wrong with Intel NUC Board DE3815TYBE ?? Aside from the fact that you can get an i7 like NUC7i7DNBE. I'd imagine the main benefit of GPUs / Xeon Phi accelerators are their communication structures and RAM. AMD Vega64 has HBM2 RAM. NVidia V100 has HBM2 RAM as well, and also is pushing the envelope with regards to "dynamic threading" for its SIMD cores. Learning to use the LDS (AMD / OpenCL) or "Shared Memory" region (NVidia) is very different than Ethernet. You only get 32kB (AMD) or 48kB (NVidia) per compute unit, but it supports atomics. Furthermore, multiple "true" threads as well as SIMD-threads can read and write to the region at speeds that are absolutely bonkers. LDS / Shared Memory is comparable to L1 speeds. Xeon Phi also has a shared memory region between cores, but I don't fully understand it. As far as I can tell, the "future" is about extending atomic operations above and beyond the current norms. For example, PCIe allegedly supports some atomic operations now (allowing CPUs and Accelerators to efficiently share data in a way that supports concurrency). In OpenCL, this is "Coarse Grain Shared Virtual Memory", and I know NVidia has an equivalent (I forgot its name however). In theory, Ethernet (1 Gbit / s to 10Gbit/s) could support atomic operations with sockets or something. But that's way slower than the LDS of an GPU Accelerator (8-bytes per cycle per shader. On the AMD Vega64, there are 64-compute units with a total of 4096 shaders. That gives the LDS a bandwidth of 32768 bytes per clock tick) There are all sorts of restrictions on this theoretical limitation: memory banks, memory channels, and whatnot. Its also heavily SIMD-based (shaders within a compute unit are "ganged" and operate in lock step). And finally, resources contend with each other: LDS is shared between workgroups and ideally you want multiple workgroups functioning at a time. But clearly this methodology supports a ridiculous amount of parallelism that truly cannot be emulated with a bunch of CPUs connected by Ethernet.
- rbanffy 9y ago> What's wrong with Intel NUC Board DE3815TYBE ?? It's for the cluster thing. With a 24-port switch chip I can fit 23 of these SoCs on a Mini ITX with one outgoing Ethernet port. On the Phi, all cores see all of the memory. It really looks like a puny version of a future Core i9 (or the 2028 Xeon Platinum). And it was the first one to have AVX 512, which is kind of nice. I guess I like my computing less asymmetric.