Member

Basement Cluster

Used 3090s and loud fans.

@basement-cluster · joined 13 Sept 2026 · 2 public runs · 1 public reviews

Sample content. This preview has no sign-in, so nothing below was written by a real member: the accounts, runs and reviews are fictional seed data, kept to show how contributions will work.

Benchmark runs

36.5 tok/s generation throughput

Qwen2.5 32B Instruct AWQ 4-bit

on AMD Ryzen 9 7950X + 3× NVIDIA GeForce RTX 3090 (Triple 3090 rack, 128 GB RAM) · vLLM 0.7.2 · cuda

@basement-cluster · 20 Feb 2025 unverified

Full environment and measurements

Benchmark: Interactive throughput
Generation throughput: 36.5 tok/s
Peak memory: 44 GB
Prompt throughput: 1,780 tok/s
Time to first token: 240 ms
Artifact: Qwen2.5 32B Instruct AWQ 4-bit
Context: 16K
OS: Debian 12
Driver: 560.35
Parameters: tensor_parallel_size=2

17.1 tok/s generation (128)

Llama 3.3 70B Instruct Q4_K_M

on Dual RTX 3090 (128 GB DDR5) · llama.cpp b4610 · cuda

@basement-cluster · 5 Jan 2025 verified

“Both cards power limited to 280W.”

Full environment and measurements

Benchmark: llama-bench
Prompt processing (512): 372 tok/s
Generation (128): 17.1 tok/s
Artifact: Llama 3.3 70B Instruct Q4_K_M (LM Studio Community)
Context: 4K
KV cache: f16
Flash attention: on
OS: Debian 12
Driver: 560.35
Parameters: split_mode=layer power_limit_w=280

Reviews

Still the value pick for VRAM per dollar

@basement-cluster on NVIDIA GeForce RTX 3090 · 13 Sept 2026

Bought three used. Power limits at 280W lose very little generation speed. Budget time for riser cables and airflow.

Speed4/5

Reliability4/5

Ease of setup3/5

Value5/5