Skip to content

Vast Benchmarking v0.1.0

Choose a tag to compare

View all tags
@kyleboddy kyleboddy released this 22 Aug 20:13
· 2 commits to main since this release
v0.1.0
f119ca5

Vast Benchmarking v0.1.0

The first packaged release provides a bounded Python benchmark, safe Vast.ai controller,
SQLite result store, and Flask dashboard for comparing container-visible hardware
capacity.

Included

  • Concurrent CV and FP16 GPU stress across every visible CUDA device.
  • Effective CPU concurrency, single-thread SHA-256, memory copy, and durable disk tests.
  • A 540-second standard-profile ceiling with partial-run validation.
  • Single-offer and parallel Vast runners with spend limits and exact-instance cleanup.
  • Category rankings normalized to the measured leader, with historical price context and
    performance-per-dollar sorting.
  • A professional light dashboard with self-hosted Geist Sans and Geist Mono.
  • Wheel and source distributions containing the CLI, dashboard, generated artwork, and
    favicon family.
  • Modular public guides for the measured demo, methodology, Vast.ai execution, and
    generic self-hosting, with automated local-link checks.
  • MIT licensing under Driveline Research.

Demonstrated snapshot

The sanitized demo contains 19 stored runs, 15 accepted machines, and 60 tested GPUs. Its
measured leaders include 118,420.80 CV images/s from eight RTX 4070 Super GPUs, 368.64
effective CPU cores on an EPYC 9654 rental, and 5.333 GB/s single-thread SHA-256 on a
Core Ultra 9 285K rental. The sanitized campaign report and machine annotations are
included in the source archive.

See the versioned demo snapshot
for exact category leaders, campaign cost, acceptance caveats, and cleanup confirmation.

Rental rates shown in the dashboard and reports are historical observations captured at
benchmark time, not current Vast.ai quotes.

Assets 5
Loading