One of the features in Ethereum is the use of a PoW function specially designed to be memory-hard. While this may be true (though a formal proof is missing and the design document is quite incomplete), the authors completely forgets another key property a PoW function must provide: it must be
sequential-memory hard. This means that not only the function should require large amounts of RAM, but it must not allow easy parallelization. Dagger seems to provide almost the best possible scenario for parallelization. In Dagger, a certain amount of RAM is filled by pseudo-random data derived from the header and the nonce. This data is produced in rounds. Each round, a number of elements from the previous round outputs are hashed together to produce the elements of the following round. These hashes can be performed in parallel. An optimized implementation for an ASIC (or FPGA) is evident for anyone with some discrete logic design background. A speedup from 256X to 2560X seems possible.
I posted more on this issue here:
http://bitslog.wordpress.com/2014/01/17/ethereum-dagger-pow-is-flawed/My own proposal (SeqMemoHash) solves this problem (
http://bitslog.wordpress.com/2013/12/31/strict-memory-hard-hash-functions/)
Best regards,
Sergio.
