Skip to content

benchmarks.md

Latest commit

 

History

History
100 lines (80 loc) · 5.66 KB

benchmarks.md

File metadata and controls

100 lines (80 loc) · 5.66 KB

Benchmarks — what the tuning actually buys you

RigForge's pitch is "stock XMRig, but with the setup and tuning that are fiddly to get right by hand." So how much does that tuning actually move the needle? Every number here is measured on real hardware, mining live, not synthetic --bench runs, so it reflects what you'd see in the wild.

TL;DR — across two very different CPUs, RigForge's one-command tuning beats stock XMRig by +3.5% (desktop Ryzen 7800X3D) to +6.6% (48-core EPYC 7642) in hashrate, and +7.6% / +6.0% in efficiency (H/s per watt). On the EPYC it also matched an expert's hand-tuned config, and auto-dodged a CPU-specific landmine: a prefetch setting that halves RandomX there but wins on the X3D. All in one command, with the nuance kept in.

How it's measured

Each configuration runs as its own XMRig process mining to the live pool. After a warm-up to steady state, we sample the hashrate (XMRig's HTTP API, 60-second average) and CPU-package power (RAPL energy counter) over several-minute windows, repeated across rounds. RandomX is low-variance and both rigs were thermally steady, so the means are solid (hashrate within ~0.1%).

The baselines:

  • Stock XMRig — upstream ./xmrig on a fresh box: no explicit HugePages, prefetcher MSRs at firmware default, default governor. (Transparent HugePages stay at Ubuntu's madvise default, as a real user would have.)
  • RigForge — setup's kernel work (2 MB + 1 GB HugePages, the per-family MSR prefetcher preset, performance governor) plus the winning knobs from a full live tune.

Rig 1 — Ryzen 7800X3D (desktop)
CPU AMD Ryzen 7 7800X3D — 8C/16T, 96 MiB L3 (V-Cache), 5050 MHz boost, AVX-512
RAM 2 × 16 GB G.Skill DDR5-6000 (EXPO, dual-channel)
OS / algo Ubuntu 24.04, kernel 6.8 · live Monero rx/0
Configuration Hashrate (H/s) Power (W) Efficiency (H/s per W)
Stock XMRig (no tuning) 10,416 86.8 120.1
RigForge — system tuning only 10,756 83.6 128.7
RigForge — tuned (performance) 10,779 83.5 129.2
RigForge — tuned (efficiency) 10,779 83.5 129.2
Stock → tuned +3.5% −3.8% +7.6%

Stock XMRig here burns more watts for less work. Without HugePages the CPU stalls on memory, drawing ~87 W to produce fewer hashes; tuned is faster and cooler. Performance and efficiency tuning landed on the same config: RigForge measured the power and found this chip pins ~84 W in any all-core setup, so there's no hashrate-for-watts trade-off to make, and it correctly didn't invent one.

Rig 2 — EPYC 7642 (48-core server) · RigForge vs an expert hand-tune

This box is the interesting one: it was already running a hand-tuned miner (the worker an operator had configured for this EPYC by hand) at 36,860 H/s. So it's not just "RigForge vs naive XMRig"; it's RigForge's one-command auto-tune against a human who tuned it themselves.
CPU AMD EPYC 7642 — 48C/96T, 256 MiB L3, 4 NUMA nodes, 2300 MHz
RAM / OS / algo 62 GB · Ubuntu 24.04, kernel 6.8 · live Monero rx/0
Configuration Hashrate (H/s) Power (W) Efficiency (H/s per W)
Stock XMRig (no tuning) 34,599 228.4 151.5
Expert hand-tuned worker (XMRig 6.25) 36,860 230.2 160.1
RigForge — tuned (XMRig 6.26) 36,866 229.5 160.6
Stock → RigForge +6.6% ~flat +6.0%
  • RigForge matched the human expert (within 0.02%), and did it with a newer XMRig (6.26 vs 6.25), cpu.yield off (the expert left it on), and 5× fewer HugePages (266 vs 1,280 reserved, both hitting 100%) — i.e. the same result with ~2 GB less RAM tied up.
  • The auto-tune dodged a landmine. On this EPYC, prefetch mode 2 halves the hashrate (~17,900 H/s), the exact opposite of the 7800X3D, where mode 2 is the winner. A fixed "golden profile" would get one of these two chips badly wrong; the per-CPU live tune measured it and stayed on the right mode for each.
  • The HugePages win is bigger here (+6.6% vs the X3D's +3.5%): a 4-NUMA EPYC with a per-node dataset leans much harder on huge pages than a single-die desktop chip. Efficiency and performance tuning again converged (power ~230 W in any config).

Caveats (read before quoting a number)

  • Two CPUs, two systems, and RandomX gains vary a lot. A desktop X3D and a 48-core EPYC already differ by more than 3× in raw hashrate; your CPU, RAM speed, NUMA layout, and kernel all matter. Treat the percentages as illustrative, not a guarantee.
  • Modern kernels narrow the stock gap. Ubuntu 24.04's Transparent HugePages (madvise) back some of even the "stock" allocation with 2 MB pages, so the stock baseline is closer to tuned than on an older kernel or with THP off. Don't expect the 20–30% some older write-ups quote.
  • setup does the heavy lifting; tune refines. Most of the win is the system tuning; the knob search confirms you're at the optimum (and, as the EPYC shows, keeps you off the landmines) rather than adding a big jump on top.

Reproduce it

# bare baseline: stock upstream XMRig, no system tuning
./xmrig -o <pool> -u <wallet>           # no HugePages reserved, MSRs default

# RigForge: one command does the setup + tuning
sudo ./rigforge.sh                       # build + HugePages + MSR + governor + service
sudo ./rigforge.sh tune --now --long     # full live tune (or just let setup's defaults ride)
sudo ./rigforge.sh doctor                # confirms HugePages 100%, MSR applied, governor, clocks