Economics of FPGAs

[diep]

Originally Posted by xilman

As you wish.

ll I can say is that my experience is not the same as yours. I am not at liberty to go into much more detail.

Paul

I assume you're speaking about a more limited chip; one that can multiply a couple of thousands of bits in just hardware without RAM.

That's an utter useless chip to the million bit primes most search for here :)

Note such limited chip got produced also as a coprocessor low power and you can buy it. That's not a fpga. I encountered one on the homepage of the company www.fox-it.com. Not that i can advice doing business at that company; when i spoke there it had a bunch of braindead people working there.

[fivemack]

It is generally incivil to ask for more information from people who've said they're not allowed to reveal it.

Generally FPGA has an enormous advantage over GPU in cases where you're processing huge amounts of data within time constraints - having the high-speed ADCs connected straight to the FPGA pins rather than going through a PCIe bus and an OS is certainly helpful.

The Australians working for the Square Kilometre Array pathfinder have built a GPU-based radio-interferometry correlator: http://local.wasp.uwa.edu.au/~chris/...ion_thesis.pdf
but the front-end is still made of FPGAs.

I haven't seen (though I'm sure it exists) an implementation of bit-sliced DES on GPUs.

DES is of course unusually hardware-friendly; http://www.copacobana.org/ have an FPGA-based DES-cracker (note the 136MHz there means it tests 136 million keys per pipelined-DES-unit per second; a single i7 core does about 16 million keys per second). Copacobana has 120 FPGAs with two units per FPGA, and costs I believe about 60k euros; 1.8 euro per megakey-per-second is a lot better than Nehalems with all their surrounding motherboard &c.

[bsquared]

Originally Posted by diep

I assume you're speaking about a more limited chip; one that can multiply a couple of thousands of bits in just hardware without RAM.

No. While I can't speak for xilman (and he can't tell us anyway), I assume he's talking about a chip that is working on a different problem. That doesn't mean that it is more or less limited. GPUs are only good at problems which map to what GPUs are good at. There are lots of interesting problems for which GPUs will be useless. For people who still want to solve these problems quickly, there are FPGA. For people who still want to solve these problems very quickly, there are ASICs.

[diep]

Originally Posted by bsquared

No. While I can't speak for xilman (and he can't tell us anyway), I assume he's talking about a chip that is working on a different problem. That doesn't mean that it is more or less limited. GPUs are only good at problems which map to what GPUs are good at. There are lots of interesting problems for which GPUs will be useless. For people who still want to solve these problems quickly, there are FPGA. For people who still want to solve these problems very quickly, there are ASICs.

Give an example of a problem where a FPGA can help, as we are running in circles here.

The only real disadvantages of gpu's is the slow support and the limited amount of RAM/caches.

Yet at a fpga you'll have less cache as you have less transistors available.

A couple of mllions versus the gpu's now, as well as cpu's are at over 2.5 billion transistors. A factor 1000 difference.

[bsquared]

Originally Posted by diep

Give an example of a problem where a FPGA can help, as we are running in circles here.

The only real disadvantages of gpu's is the slow support and the limited amount of RAM/caches.

Yet at a fpga you'll have less cache as you have less transistors available.

A couple of mllions versus the gpu's now, as well as cpu's are at over 2.5 billion transistors. A factor 1000 difference.

Surround the FPGA with memory controllers and QDRII+ SRAMs and you can essentially construct a 100MB L1 cache with a half a Tb/s of low latency bidirectional memory bandwidth. Use your imagination from there...

[diep]

Originally Posted by bsquared

Surround the FPGA with memory controllers and QDRII+ SRAMs and you can essentially construct a 100MB L1 cache with a half a Tb/s of low latency bidirectional memory bandwidth. Use your imagination from there...

Such project is some tens of millions of dollars. Idiotic to use a FPGA for that. You can use a full blown hot cpu for that that uses hundreds of millions of transistors.

See my first post on that there.

SRAM is bloody expensive, not to mention 100MB L1.

That's a project that's *really* expensive.

Completely nuts to use fpga for that instead of a real custom chip design.

Regards,
Vincent

p.s. single GPU's already has to the *shared* memory a bandwidth of 1 TB/s. To local memory it has roughly bandwidth 20TB/s.

[diep]

There was a posting from Paul DeMone, a year or 10-11 ago, referring to a full custom cpu, that was really huge (nearly a 1000MM^2, so obviously he didn't refer to itanium which was at the time far smaller than 500 MM^2), and with lots of memory controllers.

I don't see how your FPGA can compete with that chip.

Regards,
Vincent

[fivemack]

Originally Posted by diep

Give an example of a problem where a FPGA can help, as we are running in circles here.

I gave you the DES-cracker above. Cisco routers used to be made mostly out of FPGAs; mobile-phone base stations are made mostly out of FPGAs. As soon as you're manipulating perceptible amounts of I/O, FPGAs are the way to go.

FPGA is really useful when you have a completely different compute/control balance from CPU problems ... 64-bit adders on a Virtex 5 run at 400MHz and you've got space for hundreds of them. Or if you want something which looks very unlike an ALU (GF(2) polynomial multiplication, for example).

Yet at a fpga you'll have less cache as you have less transistors available

You can, as bsquared points out, attach large SRAMs ($60 for a 36Mbit 200MHz SRAM from Digikey for delivery tomorrow) to the FPGA through its thousand or so general-purpose I/O lines ... and an FPGA tends to be full of fast memory banks anyway: a large Virtex 5 has 256 block RAMs (512 36-bit words in each) which clock at 500MHz. The largest FPGA in a family will be the largest (in square-millimetre terms) chip that TSMC are prepared to give Xilinx a quote to make; transistor counts are very comparable with CPUs and GPUs, though it takes an awful lot of transistors to make one FPGA LUT.

[bsquared]

Originally Posted by diep

That's a project that's *really* expensive.

Completely nuts to use fpga for that instead of a real custom chip design.

You're completely backwards. The reason FPGAs exist in the first place is because of the enormous front end cost of ASICs. Such an approach is completely unjustified for any hobbyist project. The NRE for a FPGA solution, on the other hand, is meerly huge, and one can approach ASIC-like performance for a reasonable entry fee.

[diep]

Originally Posted by bsquared

You're completely backwards. The reason FPGAs exist in the first place is because of the enormous front end cost of ASICs. Such an approach is completely unjustified for any hobbyist project. The NRE for a FPGA solution, on the other hand, is meerly huge, and one can approach ASIC-like performance for a reasonable entry fee.

The project described above has a cost of many millions. A custom chip has that as well.

You don't create a fpga for prime number crunching in order to just produce then just 1 fpga card. That would never justify costs of carrying out the project.

To print fpga's you need to print at least a 1000-2000 chips. Add a pci card for each as well and otherwise fpga is not interesting at all; in this case that pci-card also is completely custom designed.

So we already speak here about a project in the many millions.

A custom CPU is similar price. Only if you want latest process technology for such chip it'll be more expensive.



Regards,
Vincent

[bsquared]

I thought we were talking about single unit hobbyist systems. Mass production brings in economies of scale and ASICs start to become more attractive, I agree. Fundamentally, whether you should go FPGA/ASIC, general purpose CPU, or pre-existing but special purpose IC (i.e., GPU), depends on the problem you are trying to solve. For prime-crunching (swerving back to thread topic), leveraging GPUs makes a lot of sense. For other problems it may not.