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 / DeepSeek V3Mode / EstimateUnits / GB (10⁹ B)

A / Model configuration

ModelDeepSeek

Parameters (billions)37B active / token

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: 61
Hidden: 7,168
KV heads: MLA
KV / token: 69 KB

Uses multi-head latent attention (MLA), which compresses the KV cache substantially.

B / Estimated VRAM

Inference

369.6GB

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

Model weights671B params × 0.5 B335.5 GB
KV cache8,192 tokens × batch 10.58 GB
Runtime overhead≈10% of weights33.6 GB

Memory headroom

Free after estimated load

Utilization

Target ≤ 90% of 768.0 GB

Decode ceiling

~1,730 tok/s

Theoretical, batch 1, bandwidth-bound

VRAM usage

369.6 GB / 768.0 GB

0 GB4 × 192 GB

Pinned configuration

4 × NVIDIA B200 192GB

Total VRAM: 768.0 GB
Topology: Single node · TP
Interconnect: NVLink 5 · 1.8 TB/s
Board power: 4,000 W

Compatibility / 4 × B200

Inference
369.6 GB · 48% of 4 × B200
Ready
Fine-tuning
442.8 GB · 58% of 4 × B200
Ready
Training
11,824 GB · Needs 72 × B200 (9 nodes)
Cluster

Alternative configurations

Mixture-of-experts: all 671B parameters stay resident in memory; only ~37B are active per token.

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