GPU Instances
On-demand GPU instances for training and inference. H100, A100, L40S, A10G. Per-second billing, no minimums, dedicated hardware.
At a Glance
Available GPUs
From H100 clusters to single A10G
| Configuration | GPUs | VRAM | NVLink | CPU | RAM | Ideal For |
|---|---|---|---|---|---|---|
| H100 SXM5 8x | 8 | 640 GB HBM3 | 900 GB/s | 2x EPYC 9654 | 1.5 TB | Large-scale training |
| H100 SXM5 4x | 4 | 320 GB HBM3 | 900 GB/s | 2x EPYC 9654 | 1 TB | Medium training |
| H100 PCIe 2x | 2 | 160 GB HBM3 | — | 2x EPYC 9654 | 512 GB | Training & inference |
| A100 80G 4x | 4 | 320 GB HBM2e | 600 GB/s | 2x EPYC 9654 | 1 TB | Training & fine-tuning |
| A100 80G 2x | 2 | 160 GB HBM2e | — | 2x EPYC 9654 | 512 GB | Fine-tuning, inference |
| A100 40G | 1 | 40 GB HBM2e | — | EPYC 9654 | 256 GB | Development, small training |
| L40S 48G 4x | 4 | 192 GB GDDR6 | — | 2x EPYC 9654 | 1 TB | Inference, Omniverse |
| A10G 24G | 1 | 24 GB GDDR6 | — | EPYC 9654 | 128 GB | Inference, rendering |
Why Fugoku GPUs?
Dedicated hardware. Per-second billing. No sharing.
Dedicated Hardware
Your GPUs, your VRAM, your NVLink fabric. No GPU slicing, no time-slicing, no noisy neighbors. Consistent performance.
Per-Second Billing
Pay only for what you use. Minimum 1 minute, then per-second. No hourly minimums, no reserved instance lock-in.
Full Stack Access
Root access, IPMI, BIOS settings, driver versions, container runtime. Install what you need, how you need it.
How It Works
From selection to training in minutes.
Your GPUs, Your Control
Root access. IPMI. BIOS settings. Driver versions. Container runtime. Install what you need, how you need it. No platform restrictions.
And when you need help, our GPU engineers are a Slack message away.
Start Training on Dedicated GPUs
Per-second billing. H100, A100, L40S. Expert support. Deploy today.