Skip to main contentThe Local Bench

2× RTX 3090 (used) 48 GB

NVIDIA · Multi-GPU · as of 2026-10-12

Specs sourced · speeds estimated
Multi-GPU tower

  • Data checked: 2026-10-12. Review interval shown below. Prices need review after 30 days; specifications and benchmark collections after 90 days. Dates describe the recorded check, not a live retailer quote.

    • Verified Checked against the primary source (for example the model’s config.json on Hugging Face).
    • Sourced Taken from a cited source; not yet independently checked.
    • Estimate Calculated with the published formulas; not a measurement.
    • Sample Placeholder data for layout; not a real figure.
    • Not yet measured No public, citable score found for this benchmark yet. We never fill the gap with a guess.

    Verified metadata is not a measured performance result. Only cited runs matching model, device, quant and recorded context can replace speed estimates. Estimates assume one stream and F16 KV cache, with a ±25% uncertainty band. A published run applies only to its recorded runtime and conditions; missing settings are unknown.

    Evidence labels · Freshness policy · Calculation methodology

48GB
memory, 47.0 GB usableSourced
936GB/s
memory bandwidthSourced
$2,400
price, USDEstimate
9of 10
catalogue models fit at 8k

Layer split adds capacity, not bandwidth, for a single stream (brief §7.3).

Decode-speed ceiling for a model is 936 GB/s ÷ GB read per token. For a 16.6 GB file that is 57 tok/s; real runtimes reach roughly 60% of it. Why bandwidth is the limit.

In builds: Dual-GPU 48 GB

See every model on this device

Specifications

Memory
48 GBSourced2 × 24 GB (nvidia.com). Source
Bandwidth
936 GB/sSourcedPer card; layer split does not add bandwidth for a single stream. Source
FP16 dense compute
71 TFLOPSSampleSample FP16 dense compute; pending vendor-sheet citation.
Host RAM (offload)
128 GBSampleAssumed host RAM for offload.
Vendor TDP / board power
700 WSourced2 × 350 W Graphics Card Power, nvidia.com. Source
Load power at wall
780 WSampleSample wall power under load.
Price, USD
$2,400Estimate2 × the eBay-sold median for a used RTX 3090 (≈US$1,200, Oct 2026); excludes the second-PSU/board upgrade. Source
Price, CAD
CA$3,425EstimateUSD price × 1.4271 (Bank of Canada USD/CAD, 2026-10-09); no Canadian retail listing captured. Source

Models that run

Default quant, single stream, 8k context. Fastest first; estimates are ±25% until measured.

Model
Quant
Memory
Verdict
Decode, tok/s
Max context
gpt-oss-20bMXFP412.9 GBFits245.3Estimate128k
Qwen3 30B-A3BQ4_K_M19.9 GBFits199.8Estimate32k
Llama 3.1 8B InstructQ4_K_M6.6 GBFits93.7Estimate128k
Qwen3 8BQ4_K_M6.8 GBFits90.0Estimate32k
Qwen3 14BQ4_K_M10.9 GBFits54.3Estimate32k
gpt-oss-120bMXFP464.3 GBNeeds offload36.1Estimate—
Mistral Small 3.2 24BQ4_K_M16.2 GBFits35.8Estimate128k
Gemma 3 27B ITQ4_K_M18.2 GBFits31.8Estimate128k
Qwen3 32BQ4_K_M22.5 GBFits25.6Estimate32k
Llama 3.3 70B InstructQ4_K_M45.8 GBTight12.4Estimate8k

Usable memory: discrete VRAM − 0.5 GB reserve; unified RAM × 0.75; fits when total ≤ 90% of usable (concept brief §7.3). Markdown · JSON