sech1: which CFROUND tweak?

calling it less frequently

because right now it's 10% of loop time

so if it's called 1/16th of a time (for example), it will boost hashrate by 10%

on AMD CPUs

we can use something like "some 4 bits of src register == mod.cond" to make it run 1/16th of a time

we'd have to be careful with this not to underutilize the nonstandard rounding modes

there are basically 2 options for this "ASIC" if it turns out to be real: 1. custom board with a bunch of ARM/RISC-V cores 2. VLIW core that implements the RandomX ISA in silicon

2. should be possible to brick

I don't think it will be possible to brick

If they weren't stupid

If they were, simple program size increase can brick it (it will not fit into internal buffers)

Regarding CFROUND, it will still be executed 2048/16=128 times per program

at best they could have implemented some kind of microcode to handle new instructions, which means we can't brick it, but we can make it slow

There is a 1.4 GH/s miner on Hashvault now

pretty odd they are going after that 0.6 xmr block reward, but i guess ethereum mining is dead

<g​ingeropolous:monero.social> i hope they actually sell this one. i wanna see what it is

this miner looks more like another cloud server hacker

why do we want to reduce CFROUND and "boost hashrate" just for the sake of boosting hashrate?

hmm. 2023-08-18 17:36:43.294 W There were 94 blocks in the last 90 minutes,

@hyc because it's 10% more RandomX IPC for all existing Ryzen and EPYC CPUs?

less impact on Intel, but it's the point of RandomX - to be as close to real CPUs as possible

so if some tweak gives better performance/watt on real CPUs and doesn't hurt ASIC resistance, it's a good tweak

that solo miner dropped from 1.4 GH/s to 950 MH/s

CFROUND tweak will make it 10% more efficient on Ryzens, 0-1% more efficient on Intel, and presumably will have the same ASIC resistance

<h​bs:matrix.org> Has it been considered to have opcodes shuffled from one run to the other? Could that possibly affect a putative ASIC?

the CFROUND tweak makes sense to implement; I also have some ideas for new instructions that could be added