Bede is based on @IBMPowerSystems Power 9 CPUS and @NVIDIA GPUs; the same architecture as the US government's Summit and Sierra machines.
Bede’s training system is based on the Power Systems AC922; IBM’s accelerated compute flagship.
- 32 nodes, each with a dual Power 9 processor
- 512GB DDR4 System RAM per node
- 4 NVIDIA Tesla v100 GPUs per node, each with 32GB of memory
- NVIDIA NV Link interconnects enabling up to 150Gb/s transfer
The most exciting and novel part of Bede is IBM’s IC922 system, which is designed for inference workloads.
- 6 nodes, each with a 20 core 2.9GHz Power 9 processor
- 4 nodes will feature 4 NVIDIA T4 GPUs each with 16 GB of RAM
The inclusion of the IC922 system on Bede is one of the first uses of this new hardware anywhere in the world. The use on NVIDIA’s high-bandwidth NV Link enables tensor outputs to be moved to the larger system memory block. When loss calculations occur and the algorithm is updated this data can be efficiently moved back to the GPU. This unique architecture ultimately enables deeper model to be trained using much higher resolution imagery than has previously been possible.