Benchmarking PPA of eFPGA

When comparing eFPGA between vendors, it is important to be sure to do the comparison with the same assumptions:  same foundry, same process, same variant, same array size, same voltage, same temperature range, same process corners.  

Some vendors have a significant amount of information on their websites.  Others have very little, but you may be able to get more information from those vendors under NDA.

The only metric where significant information is available for comparison is area: see a comparison of Achronix/Menta/Flex Logix in the section below performance benchmarking.

Performance Benchmarking eFPGA

Unfortunately other vendors do not have any performance information on their website other than generalities.

Flex Logix has an application note, written by our Director of Solutions Architect Tony Kozaczuk, with details of benchmarking EFLX eFPGA performance at 16nm and 28nm.  You can view and download it HERE.  Note that all of the examples are given with the exact measurement assumptions: Tj, Vj, process, process corner conditions and the compiler version.  We are constantly improving our EFLX Compiler's algorithms, so future versions may provide slightly different results.

Screenshot 2018-06-07 09.37.37.png

You can also watch a video of Tony discussing performance benchmarking HERE.

If you would like to use our performance benchmarks to compare with other eFPGA, you may access all of the Verilog files that Tony used in a Dropbox file HERE.  When you use these to compare with other eFPGA, be sure you use the exact same conditions for the other eFPGA.  If you have questions or need help, Tony would be happy to assist you: email him at tony@flex-logix.com.

What we have been told by customers who have done detailed performance benchmarking of Flex Logix versus others is that Flex Logix' performance is similar to Achronix and faster than Menta and QuickLogic.

Area Benchmarking eFPGA: Achronix, Flex Logix, Menta

Area information for some eFPGA implementations are available for three eFPGA vendors.  

For Menta, there is information to directly compare similar array sizes with EFLX for TSMC28HPM/HPC/HPC+:

Menta vs EFLX density.png

Menta's maximum array size is about 4K LUT6 equivalents.  The reason for this is apparent in the chart above which shows that the LUTs/mm2 for Menta's arrays decreases with large array sizes - presumably this is because of the exponential area increase required for a traditional mesh interconnect.  In all cases, the EFLX array has higher density than Menta by 2-3x.

For Achronix, there is no array to array comparison possible since they have no array level area information on their website.  But there is information available that allows us to compare functional blocks of the arrays.  First we'll show the area information for the EFLX4K for key functional blocks then compare that to Achronix':

EFLX4K area by block.png
achronix area by block.png
achronix screen shot 2017.png

From this information we can see that for a block of 1K LUT4 equivalents (and the associated flip flops, carry logic, including local interconnect), in TSMC16FF EFLX takes 0.19mm2 whereas Achronix takes from 0.2-0.215mm2: Achronix takes 5-13% greater area.

For the DSP MAC in TSMC16FF, EFLX takes 0.0067mm2 whereas Achronix takes from 0.0116-0.012mm2, more than 70% larger than EFLX (they do have a bigger multiplier but that doesn't explain all of the area increase).

Since EFLX is less area for LUTs and MACs, the array area should be somewhat less as well: this is confirmed by what we hear from customers who have done comparisons for various array sizes, that Achronix and Flex Logix area similar in area.