Overclocking — measured, not guessed

why the usual OC playbook does not apply to a cache-resident proof-of-work

updated 2026-07-29 23:50:58 UTC

GPU ceiling

494

H/s, RTX 3060 — measured

data bus used

24%

of peak — not bandwidth-bound

L2 read hit

0.11%

every access goes to DRAM

huge pages

+18.5%

the one real CPU lever

SMT off

+4.7%

one thread per physical core

AVX-512

−34%

bit-exact but slower

All figures measured on our own hardware in July 2026 with a bit-exact validation gate.

DON'T

the honest answer is that there is nothing to gain here — we publish the measurements instead of a table

🔧 overclocking — the measured answer

this page says: do not overclock for this coin

On most coins an OC table gains you 5-15%. On Sandglass it gains you almost nothing, and we measured why rather than assuming it.

The GPU is not the bottleneck you think. Our CUDA kernel tops out at 494 H/s on an RTX 3060. Hardware counters show the memory data bus running at only 24% of peak and an L2 read-hit rate of 0.11%. Reading 32 bytes instead of 4 changes nothing (−3%). The wall is the number of random DRAM transactions — about 2.6 billion per second — not bandwidth and not clocks. Raising core or memory clocks does not raise that number meaningfully.

Independent confirmation: the production CUDA miner shipped by the largest pool on this network reaches 476 H/s on the same card. Two independent implementations land within 4% of each other. That is a hardware ceiling, not a tuning gap.

On CPU, frequency matters and cache does not. Zen 3 and Zen 2 cost the same number of cycles per step (7.83 vs 7.71) — the per-core advantage of newer chips comes from clock speed. The hash is bound by an L2 load-to-use dependency chain, and our miner already runs within 2-3% of that floor.

What actually helps: 2 MB huge pages (+18.5% measured), one thread per physical core with SMT off (+4.7%), and pinning. All three are enabled by default in our miner — you have nothing to tune.

Power limits: leave them where your operator set them. On a memory-latency-bound workload, a lower power limit costs very little hashrate — but that is a decision for the machine owner, not for this page.