GPU nodes - LUMI-G¶
This page is about a hardware partition that has not yet reached general availability.
The LUMI-G hardware partition consists of 2560 GPU based nodes.
|2560||1x AMD EPYC 7A53||64||512 GiB||4x AMD Instinct MI250X,
220 compute units,
|none||4x 200 Gb/s|
Each LUMI-G compute node is equipped with four AMD MI250X GPUs. Based on the 2nd Gen AMD CDNA architecture, the MI250X GPUs each features 2 Graphics Compute Die (GCD) each with 110 compute units for a total of 220 compute units per GPU. Each compute unit has 64 stream processors for a total of 14080 stream processors per GPU. Each MI250X GPU provides 128GB of HBM2e delivering 3.2 TB/s of memory bandwidth. An in-package Infinity Fabric interface between the GCDs deliver a theoretical maximum bi-directional bandwidth of up to 400 GB/s.
The theoretical peak performance of one MI250X GPU is:
- FP64/FP32 Vector : 47.9 TFLOPS
- FP64/FP32 Matrix : 95.7 TFLOPS
- FP16/BF16 : 383.0 TFLOPS
- INT4/INT8 : 383.0 TOPS
More information can be found here:
- MI250X information on the AMD website
- Applications performance and benchmarks
- Frontier GPU node description
Each LUMI-G compute node is equipped with a single AMD EPYC 7A53 "Trento" CPU with 64 cores.
The LUMI-G compute nodes each have four 200 Gb/s interfaces to the Slingshot-11 interconnect with one interface directly connected to each GPU. Thus, all network communication (even for the CPU) must go via one of the GPU network interfaces.
There is no local storage on the compute nodes in LUMI-G. You have to use one of the network based storage options.