Core or Memory: What to Tune for Different Mining Algorithms

TL;DR (for those skipping the read)

The two algorithm types

Memory-bound. The algorithm constantly streams data through VRAM: the core spends most of its time waiting for memory to respond. Hashrate scales almost linearly with memory clock and barely depends on core clock. Classic examples are the Ethash/Etchash family.

Core-bound. The algorithm computes on the chip itself, with modest memory traffic. Hashrate grows with core clock, while hot memory just adds watts and degrees. Examples include Autolykos, kHeavyHash and most "compute-heavy" algorithms.

There are mixed cases where both matter — but one knob almost always matters more. A simple way to tell your coin's type: look at the core clocks other owners of the same GPU use for it. If everyone runs the core low and memory high, it's memory-bound — and vice versa.

What to tune in each case

For memory-bound: raise memory in small steps up to the stability edge, and lower the core. Cutting the core by 200–300 MHz from stock often leaves hashrate untouched while removing tens of watts and several degrees. Afterwards the power limit can go down too — the card will stop leaning against it (we cover the limit and throttling in a separate article).

For core-bound: keep memory moderate — extra memory clock gives no hashrate here but heats the VRAM chips (especially GDDR6X, see the hot-memory article). Raise the core while hashrate grows and no errors appear, and watch the watts: past a certain point, every extra megahertz costs more than the last.

A shared rule for both types: change one knob at a time and validate with a long run, not a five-minute glance. Instant hashrate lies — see the article on hashrate variance for why.

The economics: watts in the wrong place are just a bill

A typical memory-bound mistake is leaving the core at stock. That's 25–40 W of extra heat per card producing zero hashes. At $0.07/kWh: 30 W × 24 h = 0.72 kWh/day ≈ $0.05/day ≈ $1.50/month per card. An eight-GPU rig wastes about $12 a month heating the room.

The mirror mistake on core-bound coins is overheated memory. The cost isn't only watts: VRAM chips live at their limit, fans spin faster, wear accelerates. A preset tuned for the specific algorithm is almost always cooler, quieter and cheaper than a "universal" one — and more stable in hashrate too.

If you mine several coins in rotation, keep several presets — one per algorithm. Switching between them is cheaper than a compromise that is bad for all of them (we covered this in the article on tuning modes).

Where automation fits

Figuring out an algorithm's type and finding its working point is exactly what OCBot does instead of manual trial and error: it sweeps core and memory clock combinations for the specific coin, discards unstable ones, and keeps the variant with the best hashrate per watt. When the algorithm changes, it re-tunes instead of reusing an old preset on hope.