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 8BMode / 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: 32
Hidden: 4,096
KV heads: 8 × 128
KV / token: 128 KB

B / Estimated VRAM

Inference

5.5GB

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

Model weights8.03B params × 0.5 B4.0 GB
KV cache8,192 tokens × batch 11.1 GB
Runtime overhead≈10% of weights0.40 GB

Memory headroom

Free after estimated load

Utilization

Target ≤ 90% of 94.0 GB

Decode ceiling

~971 tok/s

Theoretical, batch 1, bandwidth-bound

VRAM usage

5.5 GB / 94.0 GB

0 GB94 GB

Pinned configuration

1 × NVIDIA H100 NVL 94GB

Total VRAM: 94.0 GB
Topology: Single GPU
Interconnect: NVLink bridge · 600 GB/s
Board power: 400 W

Compatibility / 1 × H100 NVL

Inference
5.5 GB · 6% of 1 × H100 NVL
Ready
Fine-tuning
13.4 GB · 14% of 1 × H100 NVL
Ready
Training
149.6 GB · Needs 2 × H100 NVL
Limited

Alternative configurations

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