Log inGet started

Thunder Compute
on your own GPUs

License Thunder Compute’s orchestration software for your VPC or on-prem cluster. Every team draws GPUs from one shared pool.

Talk to usHow it works
your-cluster · 4 nodes · 32 GPUs● healthy
node-01
  • train
  • train
  • train
  • train
  • train
  • train
  • eval
  • idle
node-02
  • infer
  • infer
  • infer
  • train
  • train
  • train
  • train
  • notebook
node-03
  • eval
  • eval
  • train
  • train
  • infer
  • infer
  • notebook
  • idle
node-04
  • notebook
  • notebook
  • notebook
  • eval
  • eval
  • infer
  • infer
  • train

ml-research · inference · evals · notebooks, one pool30/32 busy

Terminal
$ sudo thunder up
saved auth token in /etc/thunder/thunderd.env
started thunderd.service

$ thunder monitor
Zone Peers
  roster=1      status=self    latency=-
  roster=2      status=ok      latency=1ms
  roster=3      status=ok      latency=1ms
● runningthunderd · node-01 · 8 GPUs
How it works

One command per machine

Install the thunder CLI on each GPU server and run thunder up. The machine enrolls, finds its peers and joins the pool.

Drivers handled
thunder ensure-driver installs the NVIDIA driver Thunder needs
Updates itself
Nodes update thunderd when a new release is rolled out
Architecture

No control plane to run

thunderd runs on every machine. The nodes elect one manager and fail over to a standby if it goes down.

Peer to peer
Nodes track each other directly, with no scheduler to host on site
Enrolled with Thunder
Each node registers with Thunder Central to join its zone
thunder centralhosted by Thunder
your VPC or data center
  1. node-01● managerthunderd
    hostd
  2. node-02● standbythunderd
    hostd
  3. node-03● memberthunderd
    hostd
  4. node-04● memberthunderd
    hostd
gossip · peer to peer
Thunder Central, hosted by Thunder, enrolls four nodes in your VPC or data center. Each node runs thunderd and hostd, the nodes gossip peer to peer, and when the manager node goes down the standby takes over.
Deployment and security

Your cloud or your data center

Runs inside your VPC or on your own servers. Your workloads stay there, and here is exactly what Thunder sees.

Your VPC or data centeryou run
Runs on
cloud GPU instances or your servers
You manage
hardware or cloud account, network
CUDA calls
client process to GPU server, direct
Datasets and weights
on your storage
Gossip between nodes
encrypted with your zone’s key
Thunderwhat we do and see
Setup
installed with our team
→ Enrollment
hostname, IP, ports, GPU type, count
→ Check-ins
node health, on a regular cadence
→ Logs
thunderd and its components, for support
← Updates
new thunderd releases

Put every GPU in one pool

Tell us about your cluster and we’ll walk you through a deployment.

Talk to usBook a demo