Efficient · NVFP4
$40/hr $0.0111/sec · $29,200/mo at 24/7
Great for ≤70B dense models. ≤70B dense models, evals.
VRAM pool
180GB ×4 = 720GB
Position
vs L40S 4×: ~2.5× inference speed
Compute die
NVIDIA B200 180GB — NVFP4 tensor cores, tuned for ≤70b dense models, evals.
HBM3e memory
720 GB of HBM3e across the 4 GPUs.
NVLink 5 interconnect
NVLink 5 across the 4 GPUs — one memory domain.
Per-second metering
$40/hr → 0.0111/sec. Pause the deployment and the meter hits $0.
Scale-to-zero
Autoscale + region pins (GLOBAL / US / EU / APAC) on the control plane.
Memory budget
720 GB of VRAM on this preset. 70B+ dense checkpoints fit with room for KV cache (rule of thumb: about half a byte per parameter at NVFP4).
Typical loads
multi-tenant inference, parallel eval harnesses, 70B+ dense serving. Fine on: torch, JAX, raw CUDA, axolotl, vLLM, sglang, ComfyUI.
Billing math
$40/hr is $0.0111/sec while running and $0 while parked; park/wake freezes and restores state instead of paying for idle.
Best fit
720 GB across four B200 180 GB cards on NVLink 5 at $40/hr, $10 per GPU-hour. The catalog lists this row at NVFP4 and sizes it for ≤70B dense models; at FP8 weights a 70B checkpoint is ~70 GB, leaving ~650 GB of the 720 GB pool for KV cache and batch concurrency (rule of thumb).
Trade-off
b300-4x-fp4 is $48/hr — $8/hr more — for the same four-card shape with B300 288 GB cards, 1,152 GB total instead of 720 GB, still at NVFP4. Per GPU that is $12/hr against $10/hr, so the step up buys 60% more VRAM per card.
$40/hr on runaii.cloud, metered per second: $0.0111/sec while the VM runs and $0 while parked. A month of 24/7 uptime would be about $29,200, and park/wake means busy-fraction pricing in practice.
Yes. Per-second metering with park/wake: idle freezes the VM state into a signed artifact and stops the meter the same second; the next request wakes it byte-exact
720 GB of VRAM. 70B+ dense checkpoints fit with room for KV cache (rule of thumb: about half a byte per parameter at NVFP4).
$10/hr — $40/hr for four B200 180 GB cards, the lowest B200 hourly in the catalog (the 8× B200 row is $80/hr).
A 70B checkpoint at FP8 is ~70 GB of weights, so most of the 720 GB is KV cache and batch headroom — the catalog note targets ≤70B dense models here (rule of thumb).
No — the catalog lists b200-4x-fp4 at NVFP4. The catalog's 8-card FP8 row is b300-8x-fp8 (8× B300 288GB, $96/hr); there is no 4-card B200 FP8 preset.
RTX PRO 6000 slices
from $0.30/hr · 12-48GB
Blackwell slices for light jobs
1× NVIDIA T4 16GB
$0.44/hr · FP16
Starter
1× NVIDIA L4 24GB
$0.87/hr · FP8
Efficient
2× NVIDIA L4 24GB
$1.8/hr · FP8
Efficient ×2
4× NVIDIA L4 24GB
$3.6/hr · FP8
Efficient ×4
8× NVIDIA L4 24GB
$7.2/hr · FP8
Efficient ×8
1× NVIDIA RTX PRO 6000 Blackwell 96GB
$2.4/hr · FP8
Workstation
2× NVIDIA RTX PRO 6000 Blackwell 96GB
$4.8/hr · FP8
Workstation ×2
4× NVIDIA RTX PRO 6000 Blackwell 96GB
$9.6/hr · FP8
Workstation ×4
8× NVIDIA RTX PRO 6000 Blackwell 96GB
$19.2/hr · FP8
Workstation ×8
8× NVIDIA B300 288GB
$96/hr · FP8
Full Precision
8× NVIDIA B200 180GB
$80/hr · NVFP4
Throughput
4× NVIDIA B300 288GB
$48/hr · NVFP4
Minimal
16× NVIDIA GB300
$184/hr · FP8
Frontier Max