Ohio Supercomputer Heart Debuts ‘Ascend’ GPU Cluster

Lower than 10 months after it was once introduced, the Columbus-based Ohio Supercomputer Heart (OSC) has debuted its Dell-built GPU cluster, “Ascend.” Designed to lend a hand meet the ever-growing want for GPUs within the OSC consumer neighborhood, Ascend is a small machine provided with round 100 Nvidia GPUs and weighs in round 2 top petaflops, complementing OSC’s extra conventional HPC techniques.

Ascend is a Dell PowerEdge machine with 24 nodes. Each and every node has twin AMD Epyc “Milan” 7643 CPUs (for a complete of 48); quadruple Nvidia A100 (80GB) GPUs (for a complete of 96); and 921GB of “usable” reminiscence (for a complete of ~22TB). The machine makes use of Nvidia InfiniBand (200Gb/s) networking and is estimated at 1.95 top petaflops; OSC hasn’t submitted Linpack benchmarks for the reason that release of its Owens machine in 2016. Talking of Owens, each it and the more moderen Pitzer machine (2018) are nonetheless in play at OSC: each are Dell-built and powered via Intel and Nvidia {hardware}, every include loads of nodes, and OSC says that they overall round 5.5 top petaflops.

OSU’s extra conventional HPC techniques, Owens (left) and Pitzer (proper). Pictures courtesy of OSC.

OSC is presenting Ascend as a complementary, instant resolution for its consumer neighborhood. Whilst Ascend would possibly constitute a ~36% building up in relation to top petaflops, the middle says that the brand new cluster (the middle stops wanting calling it a supercomputer) triples the middle’s capability for AI, modeling and simulation.

Ascend in context, flanked via OSC’s different HPC assets. Symbol courtesy of OSC.

“OSC advanced Ascend based on discussions with our consumer neighborhood, stakeholders and distributors, who recognized a right away want for better GPU assets to procedure analysis and simulations that depend on AI, large knowledge and device studying,” stated David Hudak, OSC’s government director. “We’re happy so that you can be offering this main new useful resource to the HPC neighborhood and give a boost to consumer developments in educational analysis and business applied sciences.”

One early get entry to consumer, OSU assistant professor Yu Su, reported sure effects, operating one of the most biggest neural community fashions – BLOOM-176B – on OSU {hardware} for the primary time and highlighting scholar stories of two× to three× quicker processing in comparison to the older {hardware}. One graduate scholar, Bargeen Turzo, reported drastic variations in protein prediction: “For some huge proteins I used to be now not even ready to get a unmarried prediction on Pitzer after operating the calculation for more than one weeks,” Turzo stated. “Whilst on Ascend the similar calculation completed in 12 hours.”

Ascend was once put in closing fall, with OSC checking out the machine from October thru December with different early get entry to customers like Su. “A part of the function of the early-user duration was once to get a greater working out of the way the consumer packages employ the GPUs that we’re supporting within the machine,” defined Doug Johnson, affiliate director of OSC. “We can proceed to fortify the instrument and control of the machine as we be informed extra from what we stumble upon supporting the early customers and running the machine for an extended time period.”

“OSC’s consumer products and services and medical packages groups can be to be had to lend a hand our purchasers resolve if their packages could make excellent use of the Ascend GPUs,” Johnson added. “For some packages there’s a huge efficiency receive advantages for the usage of the GPUs and Ascend will make it imaginable for our purchasers to take on some issues that may’t be solved on our present techniques.”

OSC additionally shared a peek at its long term plans with a observe that it intends to switch its Owens cluster (at seven years outdated, an old-timer via HPC requirements) this 12 months, operating the 2 techniques at the same time as for a duration earlier than phasing out Owens.

Closing month, OSC celebrated its 35-year anniversary. “It’s our process to repeatedly be at the slicing fringe of generation, comparing and deploying it, and making it to be had right here in Ohio,” Hudak stated on that instance. “Past merely offering the generation, despite the fact that, we additionally make it uniquely versatile, inexpensive and simple to get entry to due to 35 years of enjoy pushing Ohio’s functions ahead.”

Supply By way of https://www.hpcwire.com/2023/01/19/ohio-supercomputer-center-debuts-ascend-gpu-cluster/