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 3.3 70BMode / 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: 80
Hidden: 8,192
KV heads: 8 × 128
KV / token: 320 KB

B / Estimated VRAM

Inference

41.5GB

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

Model weights70.6B params × 0.5 B35.3 GB
KV cache8,192 tokens × batch 12.7 GB
Runtime overhead≈10% of weights3.5 GB

Memory headroom

Free after estimated load

Utilization

Target ≤ 90% of 192.0 GB

Decode ceiling

~227 tok/s

Theoretical, batch 1, bandwidth-bound

VRAM usage

41.5 GB / 192.0 GB

0 GB192 GB

Pinned configuration

1 × NVIDIA B200 192GB

Total VRAM: 192.0 GB
Topology: Single GPU
Interconnect: NVLink 5 · 1.8 TB/s
Board power: 1,000 W

Compatibility / 1 × B200

Inference
41.5 GB · 22% of 1 × B200
Ready
Fine-tuning
64.0 GB · 33% of 1 × B200
Ready
Training
1,262 GB · Needs 8 × B200
Limited

Alternative configurations

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