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 32BMode / EstimateUnits / GB (10⁹ B)

A / Model configuration

ModelDeepSeek

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: 64
Hidden: 5,120
KV heads: 8 × 128
KV / token: 256 KB

B / Estimated VRAM

Inference

20.2GB

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

Model weights32.8B params × 0.5 B16.4 GB
KV cache8,192 tokens × batch 12.1 GB
Runtime overhead≈10% of weights1.6 GB

Memory headroom

Free after estimated load

Utilization

Target ≤ 90% of 24.0 GB

Decode ceiling

~18 tok/s

Theoretical, batch 1, bandwidth-bound

VRAM usage

20.2 GB / 24.0 GB

0 GB24 GB

Pinned configuration

1 × NVIDIA L4 24GB

Total VRAM: 24.0 GB
Topology: Single GPU
Interconnect: PCIe 4.0
Board power: 72 W

Compatibility / 1 × L4

Inference
20.2 GB · 84% of 1 × L4
Ready
Fine-tuning
33.9 GB · Needs 2 × L4
Limited
Training
590.2 GB · Needs 32 × L4 (4 nodes)
Cluster

Alternative configurations

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