Skip to content
MUSEBOARD

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.

Compute workbench

Workbench / Online
Model / Mixtral 8x7BUnits / GB (10⁹ B)

A / Model configuration

Mistral AI
12.9B active / token
B

Editing the count switches to a custom model with an estimated architecture.

Precision0.5 B / param
Context length8,192 tokens
Batch sizeConcurrent sequences
Workload
Pinned

Architecture / published config

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

B / Estimated VRAM

Inference

26.8GB

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

  • Model weights23.4 GB
  • KV cache1.1 GB
  • Runtime overhead2.3 GB

Memory headroom

229.2 GB

Free after estimated load

Utilization

10%

Target ≤ 90% of 256.0 GB

Decode ceiling

~930 tok/s

Theoretical, batch 1, bandwidth-bound

VRAM usage

26.8 GB / 256.0 GB

0 GB256 GB

Pinned configuration

1 × AMD Instinct MI325X 256GB

Total VRAM
256.0 GB
Topology
Single GPU
Interconnect
Infinity Fabric
Board power
1,000 W

Compatibility / 1 × MI325X

  • Inference

    26.8 GB · 10% of 1 × MI325X

    Ready
  • Fine-tuning

    34.6 GB · 14% of 1 × MI325X

    Ready
  • Training

    826.7 GB · Needs 4 × MI325X

    Limited

Alternative configurations

  • Mixture-of-experts: all 46.7B parameters stay resident in memory; only ~12.9B are active per token.