GPU
- NVIDIA GPU with at least 16 GB of VRAM (Ampere or newer recommended)
- Driver: NVIDIA 555 or later, CUDA 12.4+
- AMD MI300X / MI325 supported via ROCm 6.2+ (beta)
- Apple Silicon supported for selected models on M3 Max / M4 Pro and above (preview)
Hardware guide
What you need to host on Perchy.
The host runtime is happy on a single GPU and will not co-tenant with other workloads at the GPU level. That keeps both customers and you safe — and avoids the long-tail latency that comes with shared VRAM.
Host machine
- 16 GB system RAM minimum, 32 GB recommended
- 100 GB free SSD for the runtime cache
- x86_64 Linux (Ubuntu 22.04+, Debian 12, Fedora 39+) or Windows with WSL2
- macOS 14+ on Apple Silicon
Network
- Stable internet, ≥ 50 Mbps up / 100 Mbps down recommended
- Outbound TLS 443 to api.perchy.ai (no inbound ports needed)
- IPv4 or IPv6; CGNAT is fine — connections are outbound-only
Compliance
- Verified Stripe Connect identity (KYC handled inside the console)
- Bank account in a Stripe-supported region
- You agree to the host terms and acceptable use policy
Sanity-check command
Once you've installed the CLI, run a five-minute self-test. It exercises outbound connectivity, GPU detection, and a tiny inference. No traffic is routed to you until you opt in.
shell — perchy host check
npx perchy host check --verbose