Workbench / Online

BUILD YOUR COMPUTE STACK

Start with the model. MuseBoard maps the memory and hardware. Settings are mirrored to the URL, so any configuration can be shared.

Workbench / Online

Model / Llama 3BMode / EstimateUnits / GB (10⁹ B)

A / Model configuration

ModelMeta

Parameters (billions)

Editing the count switches to a custom model with an estimated architecture.

Precision0.5 B / param

Context length8,192 tokens

Batch sizeConcurrent sequences

Workload

Target hardwarePinned

Architecture / published config

Layers: 28
Hidden: 3,072
KV heads: 8 × 128
KV / token: 112 KB

B / Estimated VRAM

Inference

2.7GB

Serving: weights + KV cache + runtime overhead. Estimate — not a guarantee.

Model weights3.21B params × 0.5 B1.6 GB
KV cache8,192 tokens × batch 10.94 GB
Runtime overhead≈10% of weights0.16 GB

Memory headroom

Free after estimated load

Utilization

Target ≤ 90% of 24.0 GB

Decode ceiling

~583 tok/s

Theoretical, batch 1, bandwidth-bound

VRAM usage

2.7 GB / 24.0 GB

0 GB24 GB

Pinned configuration

1 × NVIDIA GeForce RTX 3090

Total VRAM: 24.0 GB
Topology: Single GPU
Interconnect: PCIe 4.0 · NVLink bridge (2-way)
Board power: 350 W

Compatibility / 1 × RTX 3090

Inference
2.7 GB · 11% of 1 × RTX 3090
Ready
Fine-tuning
9.2 GB · 38% of 1 × RTX 3090
Ready
Training
63.6 GB · Needs 4 × RTX 3090
Limited

Alternative configurations

MuseBoard gives first-order estimates for planning conversations. It does not replace profiling, capacity testing or detailed ML infrastructure planning.