DeepGreen is a massively parallel cluster with 80 GPUs, capable of over 8 petaflops of mixed-precision calculations using the NVIDIA Tesla V100 architecture. Its hybrid design can expedite high-throughput artificial intelligence and machine learning workflows, and its extreme parallelism will forge new and transformative research pipelines. It is well-suited to support training and inference using neural models.
Hardware
- 10 Penguin Relion XE4118GTS nodes, each with:
- 2 Intel(R) Xeon(R) Gold 6130 CPU @ 2.10GHz (2x 16 cores, 22M cache)
- 768GB RAM (256GB for GPFS pagepool)
- 8 NVIDIA Tesla V100s with 32GB RAM
- 4 2-lane HDR (100Gb/s per link, 400Gb/s per node) Infiniband links to a QM8700 switch
- 2 NVMe nodes, each with 88TB NVMe devices (12x8TB), replicated to provide dedicated 88TB NVMe-over-fabrics filesystem