RTX PRO 6000 BLACKWELL · SLICES
from $0.30/hr billed per second · $0 while parked
One NVIDIA RTX PRO 6000 Blackwell 96GB, sliced by the g4-standard machine types into 1/8, 1/4, or 1/2 — linear price scaling, per-second metering, and the same park/wake state freeze as every runaii GPU VM. Fine on torch, JAX, raw CUDA, axolotl, vLLM, sglang, ComfyUI.
| Slice | VRAM | Per hour | Deploys |
|---|---|---|---|
| Workstation 1/8 | 12 GB | $0.30$0.00008/sec | Launch g4-slice-8th → |
| Workstation 1/4 | 24 GB | $0.60$0.00017/sec | Launch g4-slice-4th → |
| Workstation 1/2 | 48 GB | $1.20$0.00033/sec | Launch g4-slice-half → |
| Full card (Workstation) | 96 GB | $2.40 | The whole card → |
Workstation 1/8 · 12 GB
8B-class models in FP16, LoRA adapters, embeddings, quick experiments.
Workstation 1/4 · 24 GB
~13B dense at FP8 plus video pipelines — the catalog's price/perf sweet spot class.
Workstation 1/2 · 48 GB
13-30B dense with KV headroom; mid-size LoRA fine-tunes run comfortably.
$0.30/hr for the 12GB 1/8 slice, $0.60/hr for the 24GB 1/4, and $1.20/hr for the 48GB 1/2. All are metered per second with $0 charged while parked — linear slice scaling, the same rate ladder the launch wizard bills by.
A fraction of one NVIDIA RTX PRO 6000 Blackwell 96GB. The g4-standard machine types slice the card into 1/8 (12GB), 1/4 (24GB), or 1/2 (48GB); the whole card rents at $2.40/hr as the g4-1x preset.
12GB: 8B-class models in FP16 and LoRA adapters. 24GB: ~13B dense at FP8 plus video pipelines. 48GB: 13-30B dense with KV headroom and mid-size fine-tunes. Rule of thumb: about one byte per parameter at FP8 weights.
Park the deployment and wake it on bigger hardware — state restores byte-exact. The fleet's park-wake receipts include cross-hardware wakes, so work continues across the hop.