Flex Logix has announced a design win for their embedded FPGAs, DARPA. What interests SemiAccurate about this is that the win isn’t for a product per se, it is for a class of products.
Since DARPA doesn’t make chips you are probably wondering why they are licensing embedded FPGA IP like Flex Logix makes. The answer is sort of simple if you know what kind of chips they require for projects, namely low volume parts. If you are making a satellite, how many copies of that sensor and accompanying image processor do you need? Off the shelf is right out because it probably needs to do things no consumer part needs to do, plus it probably needs to be rad hardened.
You could count to the total needed parts count on both hands for some projects, sometimes only on one, but they also need to be perfect. Low volume does not play to semiconductor economics’ strengths in this era of 8-digit mask costs. That said given the cost of getting the result into high earth orbit and the teams of people needed to run it, silicon cost isn’t exactly a deal-breaker but it isn’t welcome either.
What is less welcome is if you spend a few hundred million dollars to make a device that officially doesn’t exist, launch it into orbit, and run it for a few weeks only to discover a bug. Or a security flaw. Or something else that on a PC would require a new card or board. A satellite requires the same but getting it to high earth orbit tends to be a bit less economical, as is the process of swapping the boards when you get there.
Satellites may be an extreme case of what DARPA does, but high reliability, low volume parts with obscure performance requirements are a hallmark of many of their devices. Then there is security which, by the very nature of the things DARPA does tend to be constantly under attack, mostly of the hacking type but often in the meatspace sense of the term. All of these goals and requirements put an undue pressure on getting things done right the first time and if not then upgradability is the only choice.
Since nothing is ever done perfectly out of the box, especially on the security side, upgradability is effectively mandatory. Even if things are done perfectly a leaked key means your hardware security is about as useful as the proverbial screen door on a submarine unless it can be upgraded. Software patches tend to be frowned upon in many arenas so that leaves one path forward, make the chip upgradable in the field.
If you remember this article is about Flex Logix not the oddities of high value low volume quasi-military silicon, or at least the intersection of the two worlds. If you want upgradability what better way is there than to have an FPGA as part of your system. External FPGAs have a place but having an embedded FPGA is more useful still, if you can spare the die area.
Anyone who has a clue about silicon economics will realize that the costs of any silicon for the markets above are effectively immune to area concerns, or at least immune to their costs. If a chip costs $100M to develop, the requirements for the lifetime of the project are <1000 units, and each wafer costs $5000 to run for 100 die candidates, does adding an extra 100mm to the size really matter? Even if yields are in the single digits the cost of extra area is effectively rounding error.
Flex Logix EFL-X2.5K tiles
So back to Flex Logix, what is DARPA buying exactly? Flex Logix’s EFLX-2.5K cores have been licensed so that anyone building chips for the agency and the US government in general can incorporate the technology in future projects without licensing it individually. This saves cost, overhead, and most importantly shaves weeks if not months off time to market. Flex Logix says there is one active project now with two more in planning stages.
It also means DARPA gets all of their designs from here on out with a single embedded FPGA to code against so they have common libraries and pre-vetted blocks of logic to use. The obvious candidates are security and authentication, if the keys leak or the algorithm is cracked, upload a new one. It also allows critical IP/RTL to be kept in-house until after the chip is designed, then uploaded once the silicon is back in-house.
For things that go into space it also makes sense to put in lots and lots of LUTs just because, who knows what bugs will be found over the decade that makes up a satellite’s working life. Having a single FPGA makes debugging easier too, it is a win/win unless you are an embedded FPGA maker who is not Flex Logix that is bidding for DARPA business.
So this win is the first SemiAccurate is aware of for any embedded FPGA, if there are any others public, please let us know because we can’t find one. It covers TSMC’s 16nm node and the Flex Logix EFLX-2.5K FPGA products for any governmental projects. For some reason Flex Logix would not comment on what the three projects in question were though, and we couldn’t get them to say giant killer bunny-looking hopping drones either. It is also highly unlikely there will be any sort of fanfare when these projects are finished either but that is expected. If you are designing silicon for a giant killer bunny-looking hopping drone, at least you won’t have to negotiate a license for your embedded FPGA IP now.S|A
Latest posts by Charlie Demerjian (see all)
- More on Intel’s 10nm process problems - Sep 17, 2018
- Intel puts out another 14nm 2020 server platform - Sep 11, 2018
- Why Can’t Intel Supply Enough 14nm Xeons? - Sep 10, 2018
- Intel can’t supply 14nm Xeons, HPE directly recommends AMD Epyc - Sep 7, 2018
- AMD reintroduces the Athlon name with two CPUs - Sep 6, 2018