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 405BMode / 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: 126
Hidden: 16,384
KV heads: 8 × 128
KV / token: 504 KB

B / Estimated VRAM

Inference

227.5GB

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

Model weights405.9B params × 0.5 B202.9 GB
KV cache8,192 tokens × batch 14.2 GB
Runtime overhead≈10% of weights20.3 GB

Memory headroom

Free after estimated load

Utilization

Target ≤ 90% of 384.0 GB

Decode ceiling

~52 tok/s

Theoretical, batch 1, bandwidth-bound

VRAM usage

227.5 GB / 384.0 GB

0 GB2 × 192 GB

Pinned configuration

2 × AMD Instinct MI300X 192GB

Total VRAM: 384.0 GB
Topology: Single node · TP
Interconnect: Infinity Fabric
Board power: 1,500 W

Compatibility / 2 × MI300X

Inference
227.5 GB · 59% of 2 × MI300X
Ready
Fine-tuning
305.8 GB · 80% of 2 × MI300X
Ready
Training
7,191 GB · Needs 48 × MI300X (6 nodes)
Cluster

Alternative configurations

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