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
B / Estimated VRAM
Inference79.4GB
Serving: weights + KV cache + runtime overhead. Estimate — not a guarantee.
- Model weights70.5 GB
- KV cache1.9 GB
- Runtime overhead7.1 GB
Memory headroom
Free after estimated load
Utilization
Target ≤ 90% of 96.0 GB
Decode ceiling
Theoretical, batch 1, bandwidth-bound
VRAM usage
79.4 GB / 96.0 GB
Pinned configuration
4 × NVIDIA GeForce RTX 3090
- Total VRAM
- 96.0 GB
- Topology
- Single node · TP
- Interconnect
- PCIe 4.0 · NVLink bridge (2-way)
- Board power
- 1,400 W
Compatibility / 4 × RTX 3090
- Ready
Inference
79.4 GB · 83% of 4 × RTX 3090
- Limited
Fine-tuning
99.2 GB · Needs 8 × RTX 3090
- Cluster
Training
2,491 GB · Needs 120 GPUs — exceeds one consumer node
Alternative configurations
- Mixture-of-experts: all 141B parameters stay resident in memory; only ~39B are active per token.
- 01SERVE / READY
Inference
Estimate memory requirements for serving models.
weights + KV cache + overhead
Open in workbench - 02ADAPT / LORA
Fine-tuning
Explore memory requirements for adapting existing models.
frozen base + adapters + activations
Open in workbench - 03TRAIN / ADAM
Training
Understand large-scale compute requirements.
weights + grads + optimizer + activations
Open in workbench
MuseBoard gives first-order estimates for planning conversations. It does not replace profiling, capacity testing or detailed ML infrastructure planning.