cfsChip2600 Campaign · 10 real machines · 1 cluster
Chip Foundry ServicescfsChip2600 Cluster
Ten real lab computers, ten different GPUs, working together as one gigantic cluster. They are training a real neural network right now, each machine on its own slice of the data, and averaging what they learn every step.
Connecting…
0/60
Training step
–
Loss
–
Held-out test accuracy
0/10
Machines contributing
The cluster (live per-machine step time)
Learning curve
What the cluster learned (decision map)
Application
Name
CFS Cluster Trainer
Task
3-class spiral classification
Model
MLP 2-96-96-3 (tanh), 9,891 weights
Method
data-parallel SGD (Adam)
How it works
1
Each machine holds 12,000 training points
2
It computes gradients on its own CPU/GPU host
3
The gateway averages them (all-reduce) and updates the weights