Mike Treseler <mike.treseler@flukenetworks.com> wrote in message news:<3FE77A3C.9020805@flukenetworks.com>...> Ray Andraka wrote: > > The Altera parts still don't have a workable equivalent of the Xilinx SRL-16, > > which can be used to huge advantage in DSP applications...if you code to it. > > Otherwise, I'd agree. > > What's the downside of using the stratix block memory shifters? > > -- Mike TreselerMay be, the fact that $tratix M512 blocks unavailable in the Cheaplon?
Spartan3 availability
Started by ●December 18, 2003
Reply by ●December 23, 20032003-12-23
Reply by ●December 23, 20032003-12-23
They don't make very dense DA LUTs for distributed arithmetic filters. The M512s only give you 9 taps vs 4 with an SRL16, and there are a lot less of them than LEs. They also don't help you for reloadable LUTs, which I use as a poor-man's reconfiguration (I use SRL16's for the DA LUTs in xilinx to permit me to make a reloadable DA filter, as well as for a DA adaptive filter). The shifters work fine if it is just a shift register you are after, however the power of the SRL16 is not so much the shift, but the ability to use them as a reloadable LUT and for the dynamic shift, which is useful for reordering, for instance in an FFT. Mike Treseler wrote:> Ray Andraka wrote: > > The Altera parts still don't have a workable equivalent of the Xilinx SRL-16, > > which can be used to huge advantage in DSP applications...if you code to it. > > Otherwise, I'd agree. > > What's the downside of using the stratix block memory shifters? > > -- Mike Treseler-- --Ray Andraka, P.E. President, the Andraka Consulting Group, Inc. 401/884-7930 Fax 401/884-7950 email ray@andraka.com http://www.andraka.com "They that give up essential liberty to obtain a little temporary safety deserve neither liberty nor safety." -Benjamin Franklin, 1759
Reply by ●December 24, 20032003-12-24
Ray Andraka <ray@andraka.com> wrote in message news:<3FE8CEB1.ECF2D342@andraka.com>...> They don't make very dense DA LUTs for distributed arithmetic filters. The M512s > only give you 9 taps vs 4 with an SRL16, and there are a lot less of them than > LEs. They also don't help you for reloadable LUTs, which I use as a poor-man's > reconfiguration (I use SRL16's for the DA LUTs in xilinx to permit me to make a > reloadable DA filter, as well as for a DA adaptive filter). > > The shifters work fine if it is just a shift register you are after, however the > power of the SRL16 is not so much the shift, but the ability to use them as a > reloadable LUT and for the dynamic shift, which is useful for reordering, for > instance in an FFT. > > Mike Treseler wrote: > > > Ray Andraka wrote: > > > The Altera parts still don't have a workable equivalent of the Xilinx SRL-16, > > > which can be used to huge advantage in DSP applications...if you code to it. > > > Otherwise, I'd agree. > > > > What's the downside of using the stratix block memory shifters? > > > > -- Mike Treseler > > -- > --Ray Andraka, P.E. > President, the Andraka Consulting Group, Inc. > 401/884-7930 Fax 401/884-7950 > email ray@andraka.com > http://www.andraka.com > > "They that give up essential liberty to obtain a little > temporary safety deserve neither liberty nor safety." > -Benjamin Franklin, 1759I don't follow you point about reloadability. How are M512 less reloadable than SRL16? For me, they are more reloadable, in the sense that unlike SRL16 M512 allows random-access reload. The point about density is also questionable. In the absence of hard data I would guess that one M512s is physically smaller than 20 SRL16 it replaces in DA applications. Of course the point about moderate total amount of M512 vs. enormous amount of SRL16 is hard to argue.
Reply by ●December 24, 20032003-12-24
A DA LUT contains all the possible sums of the input taps, which for a serial filter is one tap per address bit. As you increase the number of taps, the corresponding table size grows exponentially. With LUTs, you get 4 taps per LUT (which you can do with LE's provided you don't need to reload the LUTs). If you have to reload them, then using an M512 only gives you 9 taps per M512, and each M512 is associated with a number of LABs, each containing 10 LEs. Thus the density of a reloadable DA LUT using the DA_LUTs is considerably lower than what you'd get if using LEs. Yes, the M512 is reloadable, I am not disputing that. My point is that it is not a substitute for reloadable LUTs in the LEs. The densities simply do not compare. Michael S wrote:> Ray Andraka <ray@andraka.com> wrote in message news:<3FE8CEB1.ECF2D342@andraka.com>... > > They don't make very dense DA LUTs for distributed arithmetic filters. The M512s > > only give you 9 taps vs 4 with an SRL16, and there are a lot less of them than > > LEs. They also don't help you for reloadable LUTs, which I use as a poor-man's > > reconfiguration (I use SRL16's for the DA LUTs in xilinx to permit me to make a > > reloadable DA filter, as well as for a DA adaptive filter). > > > > The shifters work fine if it is just a shift register you are after, however the > > power of the SRL16 is not so much the shift, but the ability to use them as a > > reloadable LUT and for the dynamic shift, which is useful for reordering, for > > instance in an FFT. > > > > Mike Treseler wrote: > > > > > Ray Andraka wrote: > > > > The Altera parts still don't have a workable equivalent of the Xilinx SRL-16, > > > > which can be used to huge advantage in DSP applications...if you code to it. > > > > Otherwise, I'd agree. > > > > > > What's the downside of using the stratix block memory shifters? > > > > > > -- Mike Treseler > > > > -- > > --Ray Andraka, P.E. > > President, the Andraka Consulting Group, Inc. > > 401/884-7930 Fax 401/884-7950 > > email ray@andraka.com > > http://www.andraka.com > > > > "They that give up essential liberty to obtain a little > > temporary safety deserve neither liberty nor safety." > > -Benjamin Franklin, 1759 > > I don't follow you point about reloadability. How are M512 less > reloadable than SRL16? For me, they are more reloadable, in the sense > that unlike SRL16 M512 allows random-access reload. > The point about density is also questionable. In the absence of hard > data I would guess that one M512s is physically smaller than 20 SRL16 > it replaces in DA applications. > Of course the point about moderate total amount of M512 vs. enormous > amount of SRL16 is hard to argue.-- --Ray Andraka, P.E. President, the Andraka Consulting Group, Inc. 401/884-7930 Fax 401/884-7950 email ray@andraka.com http://www.andraka.com "They that give up essential liberty to obtain a little temporary safety deserve neither liberty nor safety." -Benjamin Franklin, 1759
Reply by ●December 24, 20032003-12-24
Ray Andraka <ray@andraka.com> wrote in message news:<3FE9C209.66051EED@andraka.com>...> A DA LUT contains all the possible sums of the input taps, which for a serial filter is > one tap per address bit. As you increase the number of taps, the corresponding table size > grows exponentially. With LUTs, you get 4 taps per LUT (which you can do with LE's > provided you don't need to reload the LUTs). If you have to reload them, then using an > M512 only gives you 9 taps per M512, and each M512 is associated with a number of LABs, > each containing 10 LEs. Thus the density of a reloadable DA LUT using the DA_LUTs is > considerably lower than what you'd get if using LEs. >There are two possibilities: 1. I am totally clueless about DA. 2. You have never read Stratix datasheet and have no idea what is M512. Understandably, I prefer the 2nd option. Nobody in his right mind will use M512 for DA in 512x1b configuration, like you suggest. The natural configuration would be 32x16b or 32x18b. In such configurations one M512 provides 25% more DA bandwidth than 16 (or 18, depending on required precision) SRL16 cells. So we can say that one M512 replaces 20 (or 22.5) SRL16 cells. Additionally, each shift-accumulate block is amortized over 25% more taps, providing additional space saving. Taking all this into account, I don't believe that M512 is less dense than SRL16. Less available - probably yes, but not less dense.
Reply by ●December 26, 20032003-12-26
I'm speaking of density on the device scale. I agree, if you look just at the M512, then you have a denser filter than if you look at the equivalent number of LUTs. When you look at the device scale, you don't have very many M512s relative to the number of LUTs. I am familiar with the Stratix, and did a careful comparative analysis for one of the FPGA vendors between tbe Stratix and Virtex architectures. That vendor also was not aware of all that could be done with the SRL16. Also, yes, you would use it in the widest configurations to get the most efficient use. As a 32x16 you are effectively equivalent to 30 SRL16s plus a 16 bit adder. With DA, you only use one shift-accumulate after summing all of the partial products. In the limit, using SRL'16s you use 2 LUTs per coefficient bit for every 4 taps (one lut for the DA LUT, one for part of your partial adder). In equivalent terms, the M512 replaces about 46 LE's in a non-programmable DA filter. Additionally, you'll still need your tap delays, which means either additional memories or flip-flops in the LEs. My point is for a given marketing gate sized device, you can get a considerably larger DA filter in a device with SRL16's than you can in a device with M512's at the current M512 to LUT densities. Michael S wrote:> Ray Andraka <ray@andraka.com> wrote in message news:<3FE9C209.66051EED@andraka.com>... > > A DA LUT contains all the possible sums of the input taps, which for a serial filter is > > one tap per address bit. As you increase the number of taps, the corresponding table size > > grows exponentially. With LUTs, you get 4 taps per LUT (which you can do with LE's > > provided you don't need to reload the LUTs). If you have to reload them, then using an > > M512 only gives you 9 taps per M512, and each M512 is associated with a number of LABs, > > each containing 10 LEs. Thus the density of a reloadable DA LUT using the DA_LUTs is > > considerably lower than what you'd get if using LEs. > > > > There are two possibilities: > 1. I am totally clueless about DA. > 2. You have never read Stratix datasheet and have no idea what is > M512. > Understandably, I prefer the 2nd option. > > Nobody in his right mind will use M512 for DA in 512x1b configuration, > like you suggest. The natural configuration would be 32x16b or 32x18b. > In such configurations one M512 provides 25% more DA bandwidth than 16 > (or 18, depending on required precision) SRL16 cells. So we can say > that one M512 replaces 20 (or 22.5) SRL16 cells. Additionally, each > shift-accumulate block is amortized over 25% more taps, providing > additional space saving. > Taking all this into account, I don't believe that M512 is less dense > than SRL16. Less available - probably yes, but not less dense.-- --Ray Andraka, P.E. President, the Andraka Consulting Group, Inc. 401/884-7930 Fax 401/884-7950 email ray@andraka.com http://www.andraka.com "They that give up essential liberty to obtain a little temporary safety deserve neither liberty nor safety." -Benjamin Franklin, 1759
Reply by ●January 6, 20042004-01-06
jerry1111 <stop_this_spam_jerry1111_remove@remove.wp.pl> wrote in message news:<iga6uv87trtn31ftng213diohtdegv2d6m@4ax.com>...> >Aye, but to be fair, how long has Cyclone been out now? > > But you were able to buy small quantities early... not after > a year - IMHO.Also, dont forget that the Cyclone performs better in benchmarking than SpartanIII in terms of power and raw clock speed, despite not being 90nm...
Reply by ●January 6, 20042004-01-06
Richard Temple <rtemple@arrowuk.com> wrote: : jerry1111 <stop_this_spam_jerry1111_remove@remove.wp.pl> wrote : in message news:<iga6uv87trtn31ftng213diohtdegv2d6m@4ax.com>... : > >Aye, but to be fair, how long has Cyclone been out now? : > : > But you were able to buy small quantities early... not after : > a year - IMHO. : Also, dont forget that the Cyclone performs better in benchmarking : than SpartanIII in terms of power and raw clock speed, despite not : being 90nm... How can you judge power and raw clock speed of Spartan III, while the parts are not available yet? -- Uwe Bonnes bon@elektron.ikp.physik.tu-darmstadt.de Institut fuer Kernphysik Schlossgartenstrasse 9 64289 Darmstadt --------- Tel. 06151 162516 -------- Fax. 06151 164321 ----------
Reply by ●January 6, 20042004-01-06
Uwe Bonnes wrote:> > Richard Temple <rtemple@arrowuk.com> wrote: > : jerry1111 <stop_this_spam_jerry1111_remove@remove.wp.pl> wrote > : in message news:<iga6uv87trtn31ftng213diohtdegv2d6m@4ax.com>... > : > >Aye, but to be fair, how long has Cyclone been out now? > : > > : > But you were able to buy small quantities early... not after > : > a year - IMHO. > > : Also, dont forget that the Cyclone performs better in benchmarking > : than SpartanIII in terms of power and raw clock speed, despite not > : being 90nm... > > How can you judge power and raw clock speed of Spartan III, while the parts > are not available yet?I thought that information was in all the press releases! :)
