The market for accelerators and APUs

The PeakStream news, raising $17M in a B round was wonderful to hear about. I am happy for them, and wish them success. Recently I read that Linux Networx raised money as well. LNXI is also an interesting company. Maybe this is the harbinger of good things to come.

I don’t know. In either case, PeakStream’s product has some limitations as an accelerator, due to the single precision focus. Linux Networx also announced its own accelerators recently, though I don’t know how much has been released publically at this point about them, or if customers have them and have reactions yet. I have spoken with a wide range of customers, and a few vendors, and system builders. Everyone seems to agree that this is an important direction for the future of HPC. The story on heterogeneous processing basically indicated that this “new” direction was exciting and made for interesting possibilities. Unfortunately most users want pre-packaged systems, with already running code, and do n’t want to do the code ports themselves. So how do you make this happen? PeakStream is trying to solve this in one method using its tools and some compiler bits. It is an interesting idea, one we support. You provide an API, a common API, an open API, that people can write to. This API makes handing calculations off to the APU very easy. All manner of support functions around it, add in compiler/library support. Make this programming easy and fast. This is what PeakStream tries to do. And it looks like it might work. If you use single precision. And your code is basically a long streaming vector-like calculation. That is, it isn’t sufficiently general yet. But it is still IMO the right direction. The PeakStream people, and to a degree IBM, disparage FPGA computing. I think they may be wrong though. Again, if you can make it easy, and you can make it fast, and you can do so at a reasonable price, then there will be lots of interest. The FPGA people either have to create a VM system as in Mitrion C and others, in which case you are simply abstracting the hardware, at a cost, to be able to program it. Or you have to build “compilers” that can generate reasonably optimized VHDL, or optimizable VHDL. This is a SMOP, which if you remember the old Trace MultiFlow as a prototypical VLIW machine, sort of set the standard for how hard the SMOP of VLIW compiler and “optimal” code generation is. SMOPs (simple matters of programming) are generally never simple. Whether they are in fact feasible in finite time periods is also debatable. The FPGA is both harder in some aspects and simpler in others. Harder in the sense that you have a much larger range of optimization possibilities, in that the hardware design is not “preset” with fixed resources requiring a specific schedule. You can adjust the hardware design, which will impact the performance as well as the output. Simpler in the sense that you can build up a library of “nearly” optimal structures, with defined “code paths” to work with. If you move too much to the simpler model, then you are just doing processor code design with fixed immutable building blocks. If you move too much to the complex model, then you might need a cluster to run your combinatorial optimizer. What I am getting at is that this is a SMOP that is not simple by any stretch of the imagination. We might need different ideas to start thinking about optimization for such systems. Regardless, current FPGA programming can be simple or fast. Pick one. If you can make it simple and fast, you have a winner. I know of a company recently formed which will try to do this for a certain group of applications. But I have rambled far afield. The main point of this text is to note that companies wouldn’t be formed, products wouldn’t be developed, if there were not a meaningful economic incentive to do so. That is, if there was no market, an investment would not be a wise move. I am hoping that there are more VCs out there who recognize a good thing like the accelerator and APU market. I have pointed out before that I think it may be larger than the HPC market. Lets hope the investors agree and that these deals were not the only ones.