Nvidia targets datacentre memory bottleneck
Nvidia hopes to take graphics processing units (GPUs) within the datacentre to the subsequent stage by addressing what it sees as a bottleneck limiting knowledge processing in conventional architectures.
In common, the central processing unit (CPU) in a datacentre server would cross on sure knowledge processing calculations to a GPU, which is optimised to run such workloads.
But, based on Nvidia, memory bandwidth limits the extent of optimisation. A GPU will often be configured with a comparatively smaller quantity of quick memory in contrast with the CPU, which has a bigger quantity of slower memory.
Moving knowledge between the CPU and GPU to run a knowledge processing workload requires copying from the slower CPU memory to the GPU memory.
In an try to take away this memory bottleneck, Nvidia has unveiled its first datacentre processor, Grace, based mostly on an Arm microarchitecture. According to Nvidia, Grace will ship 10 occasions the efficiency of right this moment’s quickest servers on probably the most advanced AI and high-performance computing workloads. It helps the subsequent technology of Nvidia’s coherent NVLink interconnect know-how, which the corporate claims permits knowledge to maneuver extra rapidly between system memory, CPUs and GPUs.
Nvidia described Grace as a extremely specialised processor focusing on the biggest data-intensive HPC and AI functions because the coaching of next-generation natural-language processing fashions which have a couple of trillion parameters.
The Swiss National Supercomputing Center (CSCS) is the primary organisation publicly asserting it is going to be utilizing Nvidia’s Grace chip in a supercomputer known as Alps, due to go surfing in 2023.
CSCS designs and operates a devoted system for numerical climate predictions (NWP) on behalf of MeteoSwiss, the Swiss meteorological service. This system has been operating on GPUs since 2016.
The Alps supercomputer will probably be constructed by Hewlett Packard Enterprise utilizing the brand new HPE Cray EX supercomputer product line in addition to the Nvidia HGX supercomputing platform, which incorporates Nvidia GPUs, its high-performance computing software program developer’s package and the brand new Grace CPU. The Alps system will change CSCS’s present Piz Daint supercomputer.
According to Nvidia, making the most of the tight coupling between Nvidia CPUs and GPUs, Alps is predicted to have the ability to prepare GPT-3, the world’s largest pure language processing mannequin, in solely two days – 7x sooner than Nvidia’s 2.8-AI exaflops Selene supercomputer, presently recognised because the world’s main supercomputer for AI by MLPerf.
It stated that CSCS customers will have the ability to apply this AI efficiency to a variety of rising scientific analysis that may profit from pure language understanding. This consists of, for instance, analysing and understanding large quantities of information obtainable in scientific papers and producing new molecules for drug discovery.
“The scientists will not only be able to carry out simulations, but also pre-process or post-process their data. This makes the whole workflow more efficient for them,” stated CSCS director Thomas Schulthess.