The second milestone has been finished!
With successful completion of Phase 2 (the FPGA core; green box) I updated the image to the current planning.
- A slower STM32F429 microcontroller for Phase 4 will be replaced by a Cortex A5 which can run natively Linux. It’s faster (500MHz instead of 168MHz) and it will have more memory (128MB instead of 32MB). It also has very nice security features like ARM TrustZone and it can boot signed Linux Kernels. It also has a MMU, what means, it runs full Linux with dynamic memory management (which makes it much easier!).
- Troika was originally not planned. Short after Troika was announced a hashing core was built and integrated in the FPGA. The core is pretty fast — it needs a single clock cycle per hashing round and therefore is as fast as Curl-P81 and Keccak384. That’s very nice because (optimized) software implementations on CPUs are x10 slower than Keccak384/SHA3. Additionally, a smaller spin-off also was developed in hope small microcontrollers could make use of it.
- Also not planned from the beginning was the development of a HAT (hardware attached on top) for the Arty S7 FPGA board. It is equipped with a connector for a SWD-Debugger, a Secure-Element, SPI-Flash and a W5500 Ethernet controller. The ETH-controller was nice to play with. It proved the FPGA could run stand-alone — doing everything alone except getting Tips from the Tangle — at a speed of 1TPS (including PoW). Unfortunately, the gTTA (getTransactionsToApprove) API call to IRI still is very slow but recently there have been already major improvements on IRI and certainly gTTA also will be improved short term.
- The interface to the FPGA has not been decided in the proposal. Currently it uses a virtual COM-port interface with binding for a fork of the iota.go library for generating seeds and addresses, signing and PoW— in hope the results of the second milestone can be used by others more easily. This probably will change to something faster. One candidate is SPI with a binary protocol. Transmitting and receiving data to and from the FPGA would be more than 100 times faster.
There are good and bad news.
The good news: The concept of embedding ROM-code of the Cortex M1 in an encrypted Bitstream works very well. The code can’t be changed and keys (e.g. for decrypting transmits from the secure element) which are embedded in the ROM-code are protected by Bitstream encryption but the FPGA core still can be updated very easily by simply providing a new Bitstream. The FPGA only accepts Bitstreams which are encrypted with the correct AES key — of course only if encryption is enabled which would have to be done when using the FPGA in an unsecured environment.
The bad news: The FPGA provides an API which can be used for signing transactions. A seed- and key-index together with the BundleHash is sent to the FPGA and depending on the security-level one to three signature fragments are returned. An additional hash which is built from the parameters and an secret API key can be verified by the FPGA to make sure data hasn’t been altered on the way to the Crypto Core. Unfortunately this doesn’t make a private key completely safe because it could be brute-forced by sending different BundleHashes to the Core. This is a known problem of Winternitz OTS (one-time-signatures) which expose parts of the private key for a given key-index with every signature. An attacker could send multiple different BundleHashs to the FPGA which weakens the private key security significantly and it would be possible to calculate the key from signatures — the seed itself is pretty safe though. The consequence is that the system using the FPGA for signing transactions has to be as secure as the FPGA itself. This was one reason for switching to a more secure application controller for Phase 4 because RDP (read-out-protection) of STM32 microcontrollers are (partly) hacked.