From: glen herrmannsfeldt on 27 Mar 2010 13:39 John_H <newsgroup(a)johnhandwork.com> wrote: > On Mar 26, 11:39?pm, Randy Yates <ya...(a)ieee.org> wrote: >> I'm looking for a device that will perform something on >> the order of hundreds of millions of 12x12 multiplies >> per second, and I need it small. I only need about >> 30-40 pins of I/O. (snip) > How do you expect to get hundreds of millions of operands onto > and off the chip through 30-40 IO? That would be worth knowing, otherwise we have to guess. I think you can do I/O at 200MHz through many current FPGA, so he could do one multiplier at 200MHz. Maybe he doesn't need all the product bits, so 36 I/Os. My guess, though, is a digital filter doing a succession of multiplies and adds, in which case the I/O rate could be much lower. A systolic array multiplier in FPGA logic, latched at every step, would be very fast and you get pretty many of them on even a small FPGA. That is, not even using the block multipliers. -- glen
From: Randy Yates on 27 Mar 2010 14:29 glen herrmannsfeldt <gah(a)ugcs.caltech.edu> writes: > John_H <newsgroup(a)johnhandwork.com> wrote: >> On Mar 26, 11:39?pm, Randy Yates <ya...(a)ieee.org> wrote: > >>> I'm looking for a device that will perform something on >>> the order of hundreds of millions of 12x12 multiplies >>> per second, and I need it small. I only need about >>> 30-40 pins of I/O. > (snip) > >> How do you expect to get hundreds of millions of operands onto >> and off the chip through 30-40 IO? > > > That would be worth knowing, otherwise we have to guess. > > I think you can do I/O at 200MHz through many current FPGA, > so he could do one multiplier at 200MHz. Maybe he doesn't > need all the product bits, so 36 I/Os. > > My guess, though, is a digital filter doing a succession > of multiplies and adds, in which case the I/O rate could > be much lower. A systolic array multiplier in FPGA logic, > latched at every step, would be very fast and you get pretty > many of them on even a small FPGA. That is, not even using > the block multipliers. Hey glen! Nice to see you here! Sorry, I didn't mean to ignore your (valuable) suggestions. I've heard the term "systolic array." I guess I'm gonna have to learn about them now. :) That is, if John's lookup table approach won't work - that would be the best, and since the input data is just two bits, it'll probably be the way to go. -- Randy Yates % "With time with what you've learned, Digital Signal Labs % they'll kiss the ground you walk mailto://yates(a)ieee.org % upon." http://www.digitalsignallabs.com % '21st Century Man', *Time*, ELO
From: Randy Yates on 27 Mar 2010 14:31 Symon <symon_brewer(a)hotmail.com> writes: > On 3/27/2010 2:07 PM, Randy Yates wrote: >> >>> The coolrunner most probably won't do you any good. There are no >>> embedded multipliers and no storage beyond the macrocells (at least >>> per my recollection). >> >> That was my feeling too, but I wanted to get a more professional >> opinion. > > Hi Randy, > > I don't have much experience of CPLDs, but you may well be able to get > the multiplier performance you need. Even though a CPLD probably won't > have dedicated multipliers, there's more than one way to skin a > cat. Check out distributed arithmetic solutions. > > http://www.andraka.com/distribu.htm > > Also, you mention FEC. This might use up a fair chunk of hardware, but > again you can serialise it, if timing permits. > > As for your size requirements, I thought the smallest Coolrunner was > 6x6 = 36mm², too big for your spec. I was being very loose on the size spec. In fact I don't really (and my customer doesn't) really have a hard requirement on it, other than this, the D/As, a PIC, and other support stuff has to fit on a 2.5-in diameter PCB. -- Randy Yates % "Rollin' and riding and slippin' and Digital Signal Labs % sliding, it's magic." mailto://yates(a)ieee.org % http://www.digitalsignallabs.com % 'Living' Thing', *A New World Record*, ELO
From: glen herrmannsfeldt on 27 Mar 2010 16:41 Randy Yates <yates(a)ieee.org> wrote: (someone wrote) >> As for your size requirements, I thought the smallest Coolrunner was >> 6x6 = 36mm?, too big for your spec. > I was being very loose on the size spec. In fact I don't really (and > my customer doesn't) really have a hard requirement on it, other than > this, the D/As, a PIC, and other support stuff has to fit on a 2.5-in > diameter PCB. You could use a soft processor, such as picoblaze in a Spartan, instead of the PIC. You might do with less 'support stuff' if some of that could go in the FPGA. -- glen
From: -jg on 27 Mar 2010 16:56 On Mar 28, 2:07 am, Randy Yates <ya...(a)ieee.org> wrote: > > It's a SOQPSK modulator, so the input data is 2 bits per baud. OK, > that takes 2 bits. The output is 14-bit I/Q. Ok, thats total 30 bits. > Add a few for control. Done. You still have not stated the actual IP and OP data rates, and the maths-ops needed per delivered output. Will you be using BGA, or is this QFP and maybe single-row QFN ? -jg
First
|
Prev
|
Next
|
Last
Pages: 1 2 3 4 Prev: Version of Xilinx ISE for Spartan 6 FPGAs Next: Maximum output rate |