Google Groups no longer supports new Usenet posts or subscriptions. Historical content remains viewable.
Dismiss
Die size of BRAM/DSP48 in CLBs
66 views
Skip to first unread message
Kevin Neilson
May 22, 2013, 12:43:53 PM5/22/13
to
Does anybody know how big a BRAM or DSP48 is in terms of CLBs? (In a Virtex 5/6/7 part.) I was wondering, in terms of die size, whether it's better to use a multiplier or, say, four 36-bit fabric adders?
Andy Bartlett
May 22, 2013, 2:15:47 PM5/22/13
to
"Kevin Neilson" <kevin....@xilinx.com> wrote in message
news:2aa27460-910a-4af2...@googlegroups.com...
> Does anybody know how big a BRAM or DSP48 is in terms of CLBs? (In a
> Virtex 5/6/7 part.) I was wondering, in terms of die size, whether it's
> better to use a multiplier or, say, four 36-bit fabric adders?
I would have thought it is more a question of what you have available - go
> Virtex 5/6/7 part.) I was wondering, in terms of die size, whether it's
> better to use a multiplier or, say, four 36-bit fabric adders?
for the fastest solution if both are going spare!
jone...@comcast.net
May 22, 2013, 2:16:37 PM5/22/13
to
I would not evaluate them based on area, since you do not have the freedom to exchange one for the other in silicon.
For an arbitrary design, I would use the ratio of multipliers to CLBs on the FPGA device. If the FPGA device has a hundred multipliers and and 10,000 CLBs, then a multiplier is worth 100 CLBs on that device.
The information is available on Xilinx's website to evaluate the ratios.
For your specific design, when contemplating an optimization or additional functionality, I would use the ratio of available (remaining) multipliers & CLBs. If the virgin ratio is 1:100, but the remaining ratio is 1:1000, it would indicate that multipliers are a little more precious than CLBs for your design on that device than perhaps for other designs on that device.
Andy
For an arbitrary design, I would use the ratio of multipliers to CLBs on the FPGA device. If the FPGA device has a hundred multipliers and and 10,000 CLBs, then a multiplier is worth 100 CLBs on that device.
The information is available on Xilinx's website to evaluate the ratios.
For your specific design, when contemplating an optimization or additional functionality, I would use the ratio of available (remaining) multipliers & CLBs. If the virgin ratio is 1:100, but the remaining ratio is 1:1000, it would indicate that multipliers are a little more precious than CLBs for your design on that device than perhaps for other designs on that device.
Andy
RCIngham
May 23, 2013, 4:59:14 AM5/23/13
to
>Does anybody know how big a BRAM or DSP48 is in terms of CLBs? (In a
Virtex 5/6/7 part.) I was wondering, in terms of die size, whether it's
better to use a multiplier or, say, four 36-bit fabric adders?
>
What else are you going to use the multipliers for?
Virtex 5/6/7 part.) I was wondering, in terms of die size, whether it's
better to use a multiplier or, say, four 36-bit fabric adders?
>
The BRAMs and DSP blocks are on the die whether you use them or not.
The 'trick' is to not run out of them or CLBs when fitting your design.
---------------------------------------
Posted through http://www.FPGARelated.com
Kevin Neilson
May 23, 2013, 8:26:25 PM5/23/13
to
Yes, I see the point that since the number of CLBs and DSP48s isn't variable, like it would be on an ASIC, I should just see how much I have left of each. I guess I was thinking more in terms of a rule of thumb for reusable IP that I might want to use in different designs on different parts, like "If you're multiplying by more than a 4-bit number, a DSP48 is better." But then it still depends on that particular design.
A similar question would relate to power consumption. How many CLBs equals 1 DSP48?
A similar question would relate to power consumption. How many CLBs equals 1 DSP48?
glen herrmannsfeldt
May 23, 2013, 9:14:37 PM5/23/13
to
Kevin Neilson <kevin....@xilinx.com> wrote:
> Yes, I see the point that since the number of CLBs and DSP48s isn't
> variable, like it would be on an ASIC, I should just see how much
> I have left of each.
I believe some of the earlier families had versions available of about
> Yes, I see the point that since the number of CLBs and DSP48s isn't
> variable, like it would be on an ASIC, I should just see how much
> I have left of each.
the same size, with and without DSP48 or other processors. Maybe also
with/without BRAM and multipliers. Otherwise, you might be able to tell
from a photograph of the chip or mask. DSP48, BRAM, and multipliers
have to fit in rectangles in the array, unless they are all on the edge.
-- glen
0 new messages