3 ms·
DrDreams: it looks like a comment you made 41 days ago resulted in your account being disabled (see https://news.ycombinator.com/threads?id=DrDreams https://new
by nitrogen 14y ago
DrDreams: it looks like a comment you made 41 days ago resulted in your account being disabled (see https://news.ycombinator.com/threads?id=DrDreams https://news.ycombinator.com/threads?id=DrDreams). I'm quoting this sibling comment here for readability by those who do not enable showdead, as it is relevant to the FPGA discussion.
DrDreams 29 minutes ago | link [dead]
Speaking as an embedded developer, I see a number of other embedded devs hobbying around with FPGAs. However, I very rarely see convincing use cases for FGPAs. This article seems to lean toward the belief many of my colleagues have, that FPGAs are right around the corner in terms of general usefulness. However, I disagree strongly. I find that they are highly-specialized devices.
Before reading the rest of my writing, consider that at this time, brilliant hardware designers are putting similar amounts of work into both general purpose CPUs and into FPGAs. However, CPUS are comprised of dense blocks of special-purpose silicon for common purposes, such as floating-point math. FPGAs always have to match that dense silicon through configurable silicon, which is less dense. Furthermore, the routing in CPUs is a known entity at manufacturing time. On FPGAs, the routing is highly variable and must be re-negotiated at nearly every compile cycle. That's a huge time sink, both in terms of build time and in predicting performance. Especially since those short routes that you get at the beginning of a project, typically end up being longer by the end of it. Nowadays, we are seeing more FPGAs with dedicated, pre-made hardware blocks inside of them, such as FPUs and even CPU cores. These have more of a chance of catching on for general purpose computing. Notice however, that on these devices, it's the general-purpose CPU dominating, leaving the FPGA as a configurable peripheral, subordinate to the dense, pre-designed silicon.
Although one may be able to match GPU performance with an FPGA, it's usually just not worth it. It will take dozens of hours of FPGA coding and simulation. Compiling and fitting and the rest of the FPGA dev chain is very time-consuming and resource-intensive, compared to the speed and elegance of gcc. Speaking of standard development practices, FPGA code is not nearly as portable as C. It often has special optimizations done for the sake of the device implemented. http://opencores.org http://opencores.org has a number of more generic modules available, but still, FPGA code does not scale as well as C code. There are add-on packages that help write FPGA synthesis code - code synthesizers, but they make matters especially complicated. The syntax of Verilog and VHDL is not well-designed for scaling. Speaking of these languages, if you are used to languages written to be parsed easily, such as lisp, python, or even C and Java to some extent, you will be very appalled at the structure of Verilog and VHDL. There are many redundant entities, lots of excess verbiage and all kinds of special cases. It really has evolved very little since the days of the Programmed Array of Logic (PAL).
Another problem with FPGAs is the additional hardware on board needed to configure them. It's one more component or interface that is not needed when using CPUs. It's an additional software image to maintain, revision, store in source control, etc. FPGAs also often require more power supplies and better power supply conditioning than a regular CPU and often a separate clock crystal. They are high-maintenance.
FPGAs do shine though in a few specific instances: 1. When there is a particular, uncommon high-speed bus protocol you need to communicate with and can not buy pre-designed silicon for. This does not mean, e.g. USB. It means something like a non-standard digital camera interface or embedded graphic display. 2. Software Radio. 3. Obscure, but computationally-intensive algorithms like Bitcoin.
I hope my words have convinced some people to cool their lust for FPGAs, because I feel they're a bit of a dead-end or distraction for many who are attracted to the idea of "executing their algorithm extremely fast" or "becoming a chip designer." I have seen many students and professionals burn up hours and hours of their time getting something to run on an FPGA which could just as easily have been CPU-based. For example, one student implemented a large number of PWM oscillators on an FPGA where it would have been much simpler to use the debugged, verified, PWM peripherals on microcontrollers. Another guy I work with is intent on running CPU cores on FPGAs. This is an especially perverse use of the FPGA. Unless you've got some subset of the CPU which adds incredible value to the process, you're exchanging the the density of the VLSI/ASIC version of the chip for the flexible, less dense version on FPGA. This may be useful in rare situations, such as adding an out-of-order address generator to an existing core for speeding up an FFT, but it suffers an incredible performance and developer time hit to get to this point.
- powertower 14y ago> DrDreams: it looks like a comment you made 41 days ago resulted in your account being disabled (see https://news.ycombinator.com/threads?id=DrDreams https://news.ycombinator.com/threads?id=DrDreams). I guess the lesson is - don't talk about gently fucking math-hipsters with chainsaws on HN.
- DanWaterworth 14y agonitrogen: Thank you for reposting. DrDreams: Furthermore, the routing in CPUs is a known entity at manufacturing time. The 'routing' of a CPU is much more variable than the routing of an FPGA. Data moves around a CPU based on the program that is executing. The control logic of a CPU is the equivalent of the routing logic of an FPGA. On FPGAs, the routing is highly variable and must be re-negotiated at nearly every compile cycle. The 'routing' of a CPU is the same. The compiler has to perform register allocation afresh on every compilation. There's obviously a tradeoff between fast compilation vs most efficient use of resources. Both problems are NP-complete I believe. Nowadays, we are seeing more FPGAs with dedicated, pre-made hardware blocks inside of them, such as FPUs and even CPU cores. You just contradicted yourself. Previously, you said "FPGAs always have to match that dense silicon through configurable silicon". Your next paragraph talks about toolchain issues. This is hardly an insurmountable problem. Someone just needs to design a high level language that can be synthesised; something akin to a python of the FPGA world if you will. Another problem with FPGAs is the additional hardware on board needed to configure them. I don't quite understand, do you mean the hardware that reads the bitstream, etc or the hardware that is required in order for the FPGA to be configurable, like routing, LUTs, etc. Another guy I work with is intent on running CPU cores on FPGAs. I do agree with you here, this is a weird perversion if the purpose is not eventually to create an ASIC. I also don't believe that future processors will be FPGAs, but I do believe they will be a lot closer to FPGAs than CPUs.
- DrDreams 14y agoThanks for clearing up some of my blurrier points :) I should not have included "always" in that statement about matching the density of silicon. That breaks my own rule about absolutes. I hadn't considered the routing of a CPU at compile time similar to the routing of an FPGA. They're at different scales and have different challenges... I guess it's because I have mainly seen FPGAs in time-critical situations (where CPUs can't be used) in which it was very difficult to predict their performance and they required lots of hand-tweaking. That was in SONET routing, btw. CPUs on the other hand, usually have some time to spare, whether it is because they are overspecced, or because they are used in applications which tolerate variation in execution time (user interfaces) in comparison to FPGAs used for e.g. translating data between communications busses. It is simple to measure, reproduce and predict specs for common algorithms like the FFT, convolution, etc. for CPUs. I believe this is because the operations in the "routing" of CPU algorithms are based on the sum of known, discrete, orthogonal events like memory fetches, ALU operations, etc. Inside FPGAs though, the timing is not as uniformly discretized. It's about taking resources of routing distance from a large pool with a large geographical component, making it highly non-linear and tricky to predict. I think we all agree a better language would be the key to getting more out of FPGA technology! And I would like to see more FPGA-elements on traditional CPUs like ARM, x86, AVR, PIC, etc. I wonder what elements an improved hardware description language would use? It could certainly be trivially parseable by tools like antlr while still giving bit-level access... W/r/t additional hardware, I meant hardware that configures the bitstream, such as a header or CPU interface, as well as power supplies. For example, this schematic which I grabbed as an example: http://upload.wikimedia.org/wikipedia/en/3/3f/WillWare_Usb_fpga_sch.png http://upload.wikimedia.org/wikipedia/en/3/3f/WillWare_Usb_f... You can see the FPGA has THREE power supplies (1.2V, 2.5V, 3.3V). This particular design doesn't need additional clock sources, fortunately, but it's rather common. Understand that this isn't a fatal flaw with FPGAs, it's simply a disadvantage- one doesn't add an FPGA alone, one adds an FPGA, power supplies, possibly a crystal, and a programming interface. It means using an FPGA incurs a bunch of overhead. As for getting my account banned due to my reaction to a bogus article about a way of multiplication fraudulently claimed to be taught to Japanese school kids: Well if you folks in this little community don't need me around and don't feel what I have to say counts for anything and you don't like the way I say what I have to say, and you can't tolerate differences in people who are different from you, then go ahead and enjoy gassing each other up and blowing smoke up each other's asses in your sealed-off little echo chamber without me. I don't need you, either. I will still profit and benefit from the information here without contributing anything back if you're all so sure you know all the answers. And I'll be sure to let everyone I encounter know what an open and swell bunch of folks they can expect here, after everyone with a differing or seasoned opinion is silently silenced and banned. Keep slurping up your daily vomit from Jacques Mattheij. http://24.media.tumblr.com/tumblr_lfiferFT7c1qa55edo1_400.jpg http://24.media.tumblr.com/tumblr_lfiferFT7c1qa55edo1_400.jp...