Requirements
Becoming a TensorDock host requires an excellent setup and high-end hardware.
Non-negotiables:
Keep your hardware running.
Customers are actively using your servers. As such, excessive downtime will be penalized. Be ready to keep your servers running 24/7, and be responsive in investigating issues when they arise.
Customer data respect
Before you join, you'll be required to sign a data protection policy, which affirms that you will not attempt to look at customer data. Additionally, you're not allowed to cirvument TensorDock and contact customers directly.
Each location requires...
1 Gigabit Internet Connection
This can be shared amongst multiple servers, up to 40 GPUs.
A static IP address
This is required for customers to SSH into their machines.
Power and Internet Redundancy
For security and reliability, all new hardware added to our platform must be hosted in a data center.
Server level:
PCIe Bandwidth: 16 GB/s (4.0 x8 or 3.0 x16)
Customers need to be able to fully utilize their GPU. PCIE 3. Anything below PCIE 3.0 x16 or 4.0 x8 per GPU is not enough for modern workloads.
RAM: 16 GB+ of RAM per GPU
DDR4 required, DDR5 preferred. For GPUs with larger amounts of memory (e.g. A100 80GB), we'd recommend having slightly more RAM than the VRAM. For example, we'd recommend at least 64 GB of RAM for an A6000 48 GB.
CPU: 8+ vCPUs with 3+ GHz boost per GPU
Each GPU should have access to at least 8 CPU threads (4 physical cores when hyperthreading is enabled) that can boost to a clock speed of at least 3.0 GHz.
Storage: 1 TB NVMe SSD per GPU
DRAM cache and TLC NAND required, data center-grade preferred. Our customers use a lot of storage! If you're unsure whether your NVMe meets our requirements, please contact us.
Network: 100 forwarded ports or dedicated IP
We can lend out /24 to /22 ARIN IPv4 address subnets if you would prefer to use our IP addresses instead of yours.
GPU Passthrough
A CPU's integrated graphics, an old GPU not otherwise suitable to be rented to customers, or an onboard motherboard graphics device is suitable. This is necessary for the NVIDIA GPU to be fully passed through to the custuomer virtual machines.
