Alps is a Frankenstein Cray EX supercomputer at CSCS composed of five different types of blades spread over two physical locations: CSCS in Lugano and EPFL in Lausanne.
System overview
Its most notable partition are its GH200 blades, but it has the following blades:1
| Num nodes | Node type | CPUs | GPUs |
|---|---|---|---|
| 2,688 | Cray EX254n | 4x Nvidia Grace | 4x Nvidia H100 |
| 1,024 | Cray EX425 | 2x AMD Rome | - |
| 144 | Cray EX235n | 1x AMD Milan | 4x NVIDIA custom A100 |
| 128 | Cray EX255a | - | 4x AMD MI300A |
| 24 | Cray EX235a | 1x AMD Trento | 4x AMD MI250X |
Also of note:
- It uses custom A100 GPUs are also wonky in that some of them have 96 GB of HBM2e.
- Its GH200 nodes have only 128 GB of LPDDR5X RAM, far less than the 512 GB supported by Grace.
Performance
As of August 2024, Alps is still in preproduction so its performance has not stabilized. However there are a few papers that talk about it already:
- Understanding Data Movement in Tightly Coupled Heterogeneous Systems: A Case Study with the Grace Hopper Superchip contains more architectural details of the system.
- Exploring GPU-to-GPU Communication: Insights into Supercomputer Interconnects (arxiv.org) discusses the performance of collectives on Alps.2