Hey JCE,
Can you estimate how much of a hash-rate drop we will have on CNv2?
Thank You!
Hey!
I cannot give any precise number, even if the official Monero dev says about -10%.
What's new in V8:
* The
tweak on variable B after the AES is removed
* But replaced by a simple XOR
* There's a new 64 bits divide/remain
* There's a new square root
* There's a new 64-bits multiply
* There are
three new accesses to the scratchpad
* The main loop now have variables A, B, C and new D, while previous had only A and B, C being a temporary. It doubles the need of registers.
So i expect it to be a bit slower on 64-bits CPU since they have enough registers to handle all this, and the 64-bits divide/remain and the squareroot are all native instructions.
On 32-bits, there will be a higher performance drop, i expect, since they will run out of registers.
On GPU, both the divide and the triple access to Memory will have a big impact, i expect. Not benched yet. Because GPU memory is always asynch and such accesses to VRAM cost a lot of time. It will increase the gap between good and bad memory-equipped cards, and tighten the performance gap on cards having the same memory controller, like RX550 against RX560
Again, all this is theorical impact, i haven't finished my assembly implementation. When done i'll release the 0.33 for Windows and Linux CPU first, then the GPU version.
