Niko Laine
@nodelordRegular
Big GPUs, small rooms. Tests whether a model is worth the electricity.
- Throughput
- MoE models
- Hardware
Helsinki, Finland · English, Finnish · Europe/Helsinki
Pinned
spark-vs-5090Oct 16, 2026Outcome: Scripts and raw run records comparing a DGX Spark and an RTX 5090 on MoE models.Open sourceIndependently reproduced
- Establishes
- Another member re-ran it and matched the reported results.
- Doesn’t establish
- Does not establish results on other hardware.
- Disclosure
- Independent work
Runs
20 verified · 7 on BenchClub cloud
ModelWhereSpeed
Qwen3-Coder-30B-A3B-Instruct Q4_K_MNov 4, 2026RTX 5090 · 32 GB371.7 tok/s
gpt-oss-20b MXFP4Nov 1, 2026L40S · 48 GBCloud173.2 tok/s- Qwen3.5-27B Q4_K_MNov 1, 2026GB10 · 128 GB10.7 tok/s
Llama-3.1-8B-Instruct Q4_K_MOct 26, 2026RTX 5090 · 32 GB247.6 tok/s- Llama-3.1-8B-Instruct Q4_K_MOct 25, 2026L40S · 48 GBCloud123.4 tok/s
- Qwen3.5-2B Q4_K_MOct 24, 2026L40S · 48 GBCloud296.4 tok/s
- Qwen3.5-2B Q4_K_MOct 21, 2026L40S · 48 GBCloud299.5 tok/s
- Llama-3.1-8B-Instruct Q4_K_MOct 21, 2026GB10 · 128 GB37.3 tok/s
PassedFailedNot supported
Accounts, metrics and works carry separate evidence. Updated Nov 5, 2026. Public JSON