Daily Image
03-01-2024
Our fastest computer ever
| Submitter: | John Romein |
| Description: | Just before the Christmas break, Santa delivered two new, highly innovative Grace Hopper computer systems, in behalf of the RADIOBLOCKS project (GA-101093934). The innovation comes from the tight integration between the CPU and the GPU: they can access each other's memories coherently, 14 times faster than in our current GPU systems. This removes a common bottleneck for basically all our GPU applications. The CPU, GPU, and memory are prefabricated on a small module (approximately 20x15 cm2), which is fully hidden by the two heatsinks in the middle of the image. The carrier board with cable connectors is partially visible. The module contains the most powerful GPU ever built, consumes up to 1000 Watt, and provides roughly the same compute power as the 6-rack Blue Gene supercomputer that we had in the early LOFAR days. Those people that were still in the building when we powered on the Grace Hopper for the first time, may have noticed that the noise from the system fans is similarly powerful. NVIDIA designed these systems for heavy AI tasks like training large language models. However, the same tensor-core technology that accelerates such AI tasks can also accelerate correlator and beam former computations, at very high energy efficiency. The tight CPU-GPU integration may be beneficial for tasks like hi-res imaging and dedispersion too. In short, interesting times are ahead, but we first have to learn how to unleash the computational power and I/O capabilities of such an extremely powerful platform. |
| Copyright: | John Romein |
| Tweet |  |