Xilinx, High Performance Computing Using FPGAs, Sep 2010, here.

The shift to multicore CPUs forces application developers to adopt a parallel programming model to exploit CPU performance. Even using the newest multicore architectures, it is unclear whether the performance growth expected by the HPC end user can be delivered, especially when running the most data- and compute- intensive applications. CPU-based systems augmented with hardware accelerators as co-processors are emerging as an alternative to CPU-only systems. This has opened up opportunities for accelerators like Graphics Processing Units (GPUs), FPGAs, and other accelerator technologies to advance HPC to previously unattainable performance levels.

I buy the argument to a degree. As the number of cores per chip grow, the easy pipelining and parallelization opportunities will diminish. The argument is stronger if there are more cores per chip. 8 cores or under per general purpose chip it’s sort of a futuristic theoretical argument. More than a few programmers can figure out how to code up a 4 to 8 stage pipeline for their application without massive automated assistance. But the FPGA opportunity does exist.

The convergence of storage and Ethernet networking is driving the adoption of 40G and 100G Ethernet in data centers. Traditionally, data is brought into the processor memory space via a PCIe network interface card. However, there is a mismatch of bandwidth between PCIe (x8, Gen3) versus the Ethernet 40G and 100G protocols; with this bandwidth mismatch, PCIe (x8, Gen3) NICs cannot support Ethernet 40G and 100G protocols. This mismatch creates the opportunity for the QPI protocol to be used in networking systems. This adoption of QPI in networking and storage is in addition to HPC.

I buy the FPGA application in the NIC space. I want my NIC to go directly to L3 pinned pages, yessir I do, 100G please.

Xilinx FPGAs double their device density from one generation to the next. Peak performance of FPGAs and processors can be estimated to show the impact of doubling the performance on FPGAs [Ref 6], [Ref 7]. This doubling of capacity directly results in increased FPGA compute capabilities.

The idea proposed here is that you want to be on the exponentially increasing density curve for the FPGAs in lieu of clock speed increases you are never going to see again. Sort of a complicated bet to make for mortals, maybe.

I like how they do the comparisons though. They say here is our Virtex-n basketball player and here is  the best NBA Basketball player … and they show you crusty old Mike Bibby 2012. Then they say watch as the Virtex-n basketball player takes Mike Bibby down low in the post, and notice the Virtex-n basketball player is still growing exponentially. So you can imagine how much better he will do against Mike Bibby in the post next year. Finally they say that Mike Bibby was chosen as the best NBA player for this comparison by his father Henry, who was also a great NBA player.

FPGAs tend to consume power in tens of watts, compared to other multicores and GPUs that tend to consume power in hundreds of watts. One primary reason for lower power consumption in FPGAs is that the applications typically operate between 100–300 MHz on FPGAs compared to applications on high-performance processors executing between 2–3 GHz.

Silly making Lemonade out of Lemons argument, the minute I can have my FPGAs clocked at 3 GHz I throw away the 300MHz FPGAs, no?

Intel, An Introduction to the Intel QuickPath Interconnect, QPI, Jan 2009, here.

Xilinx Research Labs/NCSA, FPGA HPC – The road beyond processors, Jul 2007, here. Need more current references but  I keep hearing the same themes in arguments for FGPA HPC, so let’s think about this for a bit:

FPGAs have an opening because you are not getting any more clocks from microprocessor fab shrinks: OK.

Power density: meh. Lots of FinQuant code can run on a handful of cores. The Low Latency HFT folks cannot really afford many L2 misses. The NSA boys are talking about supercomputers for crypto not binary protocol parsing.

Microprocessors have all functions that are hardened in silicon and you pay for them whether you use them or not  and you can’t use that silicon for something else: Meh, don’t really care if I use all the silicon on my 300 USD microprocessor as long as the code is running close to optimal on the parts of the silicon useful to my application. It would be nice if I got more runtime performance for my 300USD, no doubt. This point is like Advil is bad because you don’t always need to finish the bottle after you blow out your ankle. Yeah, I understand the silicon real estate is the most expensive in the world.

Benchmarks: Black Scholes 18msec FPGA @ 110 Mhz Virtex-4 203x faster than Opeteron – 2.2 Ghz: You Cannot be Serious! 3.7 microseconds per Black Scholes evaluation was competitive performance at the turn of the century. The relative speedup slides and quotations make me nervous. Oh, Celoxica provided the data – hey Black Scholes in 36 Nanoseconds on a single core of a dual core off-the-shelf general microprocessor from 2007. So the Virtex-4 does 1M Black Scholes evaluations in 18 milliseconds flat to competitive code on a dual core general purpose off-the-shelf microprocessor in 2007.

Make it easy for the users to use this hardware and get „enough of a performance‟ increase to be useful: meh, it’s for applications that do not need to go fast, for now (2007)?

Do not try to be the fastest thing around when being as fast with less power is sufficient: meh, really do not care so much about the power thing

FPGA: Different operations map to different silicon allows massive pipelining; lots of parallelism: OK. So, why bother with the previous two points?

Eggers/ U. Washington, CHiMPS, here. Eggers is reasonable.

There have been (at least) two hindrances to the widespread adoption of FPGAs by scientific application developers: having to code in a hardware description language, such as Verilog (with its accompanying hardware-based programming model) and poor FPGA memory performance for random memory accesses. CHiMPS, our C-to-FPGA synthesis compiler, solves both problems with one memory architecture, the many-cache memory model.

Many-cache organizes the small, distributed memories on an FPGA into application-specific caches, each targeting a particular data structure or region of memory in an application and each customized for the particular memory operations that access it.

CHiMPS provides all the traditional benefits we expect from caching. To reduce cache latency, CHiMPS duplicates the caches, so that they’re physically located near the hardware logic blocks that access them. To increase memory bandwidth, CHiMPS banks the caches to match the memory parallelism in the code. To increase task-level parallelism, CHiMPS duplicates caches (and their computation blocks) through loop unrolling and tiling. Despite the lack of FPGA support for cache coherency, CHiMPS facilitates data sharing among FPGA caches and between the FPGA and its CPU through a simple flushing of cached values. And in addition, to harness the potential of the massively parallel computation offered by FPGAs, CHiMPS compiles to a spatial dataflow execution model, and then provides a mechanism to order dependent memory operations to retain C memory ordering semantics.

CHiMPS’s compiler analyses automatically generate the caches from C source. The solution allows scientific programmers to retain their familiar programming environment and memory model, and at the same time provides performance that is on average 7.8x greater and power that is one fourth that of a CPU executing the same source code. The CHiMPS work has been published in the International Symposium on Computer Architecture (ISCA, 2009), the International Conference on Field Programmable Logic and Applications (FPL, 2008), and High-Performance Reconfigurable Computing Technology and Applications (HPRCTA, 2008), where it received the Best Paper Award.