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 / Gemma 9BMode / EstimateUnits / GB (10⁹ B)

A / Model configuration

ModelGoogle

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: 42
Hidden: 3,584
KV heads: 8 × 256
KV / token: 336 KB

B / Estimated VRAM

Inference

7.9GB

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

Model weights9.24B params × 0.5 B4.6 GB
KV cache8,192 tokens × batch 12.8 GB
Runtime overhead≈10% of weights0.46 GB

Memory headroom

Free after estimated load

Utilization

Target ≤ 90% of 40.0 GB

Decode ceiling

~337 tok/s

Theoretical, batch 1, bandwidth-bound

VRAM usage

7.9 GB / 40.0 GB

0 GB40 GB

Pinned configuration

1 × NVIDIA A100 40GB

Total VRAM: 40.0 GB
Topology: Single GPU
Interconnect: NVLink 3 · 600 GB/s
Board power: 400 W

Compatibility / 1 × A100 40GB

Inference
7.9 GB · 20% of 1 × A100 40GB
Ready
Fine-tuning
18.9 GB · 47% of 1 × A100 40GB
Ready
Training
175.7 GB · Needs 8 × A100 40GB
Limited

Alternative configurations

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