By Flex Logix Technologies, Inc.
There are multiple eFPGA suppliers in the market today: Achronix, Adicsys, Efinix, Flex Logix™, Menta, QuickLogic.
There are 3 different business models and engineering approaches to eFPGA which you should understand to assess how it will impact your success in using their eFPGA IP and their viability as a supplier long term.
FPGA Chip Companies Providing eFPGA IP
FPGA chip companies generally build a new generation of FPGAs every ~3 years when there is a major advance in process technology.
They pick one foundry, one node, one variation of that node and do full-custom circuit design with typically the maximum or near-maximum number of metal layers in order to get the highest density FPGA they can. It takes them most of the 3 years to do the complex engineering required.
Since FPGA customers want a range of sizes and some variation in the ratio of options like DSP/RAM, the FPGA chip companies will construct their FPGAs from some modular pieces: a block of LUTs, a DSP block, and typically a block-RAM (dual port). The 3-10 different sizes of the FPGA are put together from the blocks with circuit designers tuning the mesh interconnects and I/O’s for the array size.
Their business model is to optimize to make the best FPGAs. What happens when they provide embedded FPGA IP?
1st There are typically dozens of metal stacks that a foundry supports. The bottom 4-7 layers, depending on the process, are generally common because of foundation IP like standard cells and memories. Above that, some customers want fewer layers for lower cost for simpler circuits; others want maximum layers for large, complex circuits. There are many variations of thicknesses/widths by layer to optimize for each customers design. FPGA companies usually design their chip with maximum or near-maximum metal layers which significantly limits the supported metal stacks to one or two. If a customer wants a different metal stack, they have to re-route. If the customer wants the same number of metal layers but with variations in thickness for some of the layers, timing will have to be redone and likely re-routing of the whole design along with circuit changes to offset timing/DRC issues with thicker/thinner metal. If the customer wants fewer metal layers, it may be impossible: presumably the FPGA chip uses the number of layers it does because it was not possible to route with fewer layers. The time to do all this work is likely 4-6 months with significant engineering expense.
2nd Foundries continually improve each of their process nodes for yields, fewer metal layers and shrinks with a new variation every year or so. Since FPGA chip companies do full-custom design, they will need to re-simulate and likely re-design multiple portions of their chip to support the incremental changes to the process. (Whereas standard cells are generally useable across 2 or 3 incremental variations because they use less aggressive logic design rules AND the Copyright 2017, Flex Logix Technologies, Inc. Flex Logix and EFLX are trademarks of Flex Logix. Achronix, Adicsys, Menta, QuickLogic, Efinix are trademarks of their respective owners. foundries try to keep the standard cells the same for their customers to migrate easily to the newer process variation).
3rd Supporting a range of array sizes and options (DSP, RAM) requires custom engineering: the blocks may be modular, but the connections between them and most importantly the interconnect will need to be redone especially since the amount of interconnects grows with N2 for mesh interconnect designs. And the I/O ring is custom for each different array size.
4th Since the GDS changes for every metal stack and array size and process variation, it is uneconomical to do a validation chip for every GDS change.
Figure 1: Traditional 2D-Mesh Interconnect diagram illustrates the non-uniformity of a mesh network across the FPGA. Any change to the area or configuration will require re-implementing the interconnect, effectively creating a new embedded FPGA implementation.
The FPGA chip companies have been in business >10 years but offer eFPGA on only a few nodes.
The engineering investment to support different array sizes and options and metal stacks within an existing node/variation are significant; the engineering investment to port a full-custom design to a new node are much greater (that’s why FPGA companies usually only do a generation every few years).
This is probably why the big FPGA chip companies don’t bother with eFPGA: it is a costly distraction to their primary business with, for them, a low return on investment.
Pure eFPGA IP Companies: Soft IP
eFPGA soft IP companies offer a software tool that will generate RTL for an array based on inputs such as array size, I/O count, etc. The customer can then use EDA tools with a standard cell library to implement the eFPGA in any process – but the density is very low: FPGAs are very regular and benefit from structured placement. This approach has some use in test chips or very low volume products such as aerospace/defense.
One of these companies now offers hard IP on a couple of foundries/nodes. For that company, their maximum array size is <<10K and there are only a handful of sizes/option combinations to choose from. Density for the smallest arrays is ~0.5x of a full-custom FPGA; and for the largest arrays ~1/3 of a full-custom FPGA. Presumably what is happening is the N2 complexity growth in interconnect for larger arrays. Their largest array is 2x the LUTs and Flip-Flops of their midsize array, but is ~3x the silicon area! This trend in interconnect complexity growth is probably why there are no large arrays offered. The number of metal layers required or the range of metal stacks they are compatible with is not public.
Each array size is a different design so a validation chip for one does not prove the others. Doing a validation chip for each array size is uneconomical.
Pure eFPGA Hard IP: Flex Logix
Flex Logix is the youngest of the companies providing eFPGA but offers eFPGA on more process nodes/variations (7 foundry and 1 captive) and over a wider range of sizes than any competitor.
We started the company based on Cheng Wang’s revolutionary interconnect which he developed working with others at UCLA while doing 5 different FPGA test chips of increasing complexity over multiple process nodes prior to starting Flex Logix.
In traditional FPGAs, the FPGA fabric is 70-80% interconnect and only 20-30% of the area is logic/LUTs.
Cheng’s test chips were limited in size by budget suppliers: To get more logic on the chip he came up with a new interconnect that was much denser than the traditional mesh. And its’ complexity grows more slowly than mesh for larger array sizes.
Figure 2: Illustration of Flex Logix interconnect resulting in a much more efficient interconnect compared to a traditional interconnect.
Cheng’s interconnect as developed at UCLA was the subject of a paper that won the Outstanding Paper Award at ISSCC in 2014 and of a patent recently issued to UCLA, of which Flex Logix is the exclusive licensee. Since starting Flex Logix, Cheng has made numerous improvements to the interconnect, some of which resulted in two recently issued patents.
The end result is that Flex Logix can use standard cells for rapid implementation and portability across incremental process variations, while achieving density essentially the same as eFPGA from full-custom FPGA chips: the increased density of the interconnect offsets the lesser density of the standard cells. The effective density of EFLX eFPGA is further increased by the higher utilization we achieve compared to traditional FPGA interconnect.
In 40nm our IP is compatible with two variations; in 28nm our IP is compatible with two variations; and in 16nm our IP is compatible with 3 variations.
Further, Flex Logix’ interconnect does not need maximum metal layers: in 40nm, 28nm and 16nm our eFPGA IP is compatible with almost all metal stacks.
Array Size Scalability
When we started Flex Logix, we realized in early talks with potential customers that customers wanted both silicon proven and a wide range of array sizes. We realized that designing say 10 different array sizes would mean needing to have 10 different validation chips: this wasn’t feasible and not validating in silicon would mean unacceptable risk for the customer.
So Cheng came up with an innovative solution: make an eFPGA IP core which is a complete FPGA on its own which can be tiled, without GDS changes, to make a very wide range of array sizes.
For example, the EFLX®4K IP core is a complete embedded FPGA of 4K LUT4s with >600 inputs and >600 outputs. But the EFLX4K also has a top-layer interconnect, not shown in the block diagram to the right, which automatically extends between cores when abutted enabling ~50 array sizes up to 200K LUTs. Any array configuration needed for a chip can be generated within a few days: most of the time is generating the .LIB files across numerous process corners, since timing is done at the array level.
Figure 3: Diagram of an EFLX Array. By arraying a validated embedded FPGA core, arrays of different sizes are validated by design thus eliminating risks associated with new embedded FPGA implementations on a given process.
We have an EFLX150 core for smaller array requirements; and someday we can implement an EFLX16K core when customers need up to 800K (or larger) LUT4 arrays.
Plus we have two versions of each IP core: all Logic and Logic with ~20% of the area replaced by Multiplier-Accumulators: the two versions are exactly the same dimensions so they can be intermixed to give customers the ratio of DSP-to-Logic they need.
Silicon Proven IP for Your Node/Variation, Your Metal Stack and Your Array Size
When Flex Logix ports an EFLX IP core to a new process, it builds a validation chip with at least a 2x2 array (this verifies that the top level interconnect extension works on all 4 sides) on a die with PLLs, PVTs and attached SRAM so testing can be done at high speeds (>1GHz on 16nm), at exact voltages and extreme temperatures to validate silicon function and performance. We have built 4 validation chips so far in 40nm, 28nm and 16nm (two: one for the EFLX150 and one for EFLX4K).
Since an EFLX array is built by abutting EFLX cores with no GDS changes, the EFLX Array you get is known good for your size and your node/variation.
We have demonstrated this by fabricating a 7x7 array of EFLX4K cores (EFLX200K = 200K LUT4s) in 16nm which is in validation now. It is fully functional, was demonstrated at ARM TechCon running our Flex Micro architecture. Now performance measurements are being done over temperature and voltage.
Flex Logix gives you the highest density, most flexibility on process variations/metal stacks, and widest range of array sizes while ensuring the GDS you use has been proven in silicon.
We are happy to go into more detail on any of the above analysis.
If you wish to download a copy of this white paper, click here