What are the changes in CryptoNight v2 PoW algorithm?

Question

According to the last Dev Meeting, PoW changes to the upcoming fork are almost finished. What exactly will be changed and how will it affect speed on CPUs and GPUs but also potential ASICs or FPGAs?

Answer 1

The main changes are:

• reading a whole 64 byte chache line at a time: many recent processors will read 64 bytes at a time when fetching from RAM, so access to the whole set is virtually free after it's cached. Since ASICs don't have to behave like this, this will increase the memory bandwidth an ASIC needs by 4, since Cryptonight acts on 16 bytes at a time. A simple addition is done to ensure an ASIC can't just multiplex the data.

• adding integer math operations (division and square root): on a CPU, these are "hidden" by the memory access latency. This will cause an ASIC to have to have extra circuitry to compute those, as well as prevent part of the pipelining an ASIC may be able to perform due to the serialization of data use and calculation.

On most recent enough CPU/GPU hardware (apart from ARM), performance was found to be from 85% to 95% of the performance of Cryptonight, see reports from https://github.com/monero-project/monero/pull/4218.

Answer 2

I wouldn't say it's "almost finished". There is a lot of discussion on this particular PR here. On reading, it looks like there is a lot to iron out still on this particular proposal if it were to ever get merged.

There are alternative PoW changes being worked on also. Any merged change would ensure CPU/GPU mining would have negligible performance changes whilst also bricking any in development ASICs. Right now, there is no evidence of any ASIC or FPGA mining on the network - the network difficulty has remained very stable since April of this year to now.