Skip to content
GIRIVON
Infrastructure+
Systems+
Components+
Computing Devices+
Solutions+
Compute Tools+
Knowledge+

AI Training

Training and pre-training neural networks across one or many accelerators.

What matters, in order

  1. Accelerator memory capacity
  2. BF16/FP8 tensor throughput
  3. GPU-to-GPU and node-to-node bandwidth
  4. Fast storage for data and checkpoints
  5. Sustained power and cooling

Engineering guidance

Training compute is roughly 6 × parameters × tokens floating-point operations. Mixed-precision Adam needs about 16 bytes of accelerator memory per parameter for weights, gradients and optimizer states before activations, so large models are sharded across GPUs with fast interconnects.

Size it

Categories

Products tagged for AI Training

27 products

Compare selected