Skip to main content

Supercomputer Private Nodes

Research Computing supports managing faculty-purchased vendor hardware within the Sol supercomputer. This includes supplying rack space, power, cooling, and networking. Faculty can purchase pre-approved compute node profiles to match their CPU, memory, and GPU needs. These nodes are created in a privately-accessible partition that provides prioritized access to the faculty member and supporting group.

Compute Node TypeProcessorCoresRAMGPUApproximate* Price
Standard ComputeAMD EPYC128512GBN/A$59,393.64
High MemoryAMD EPYC1282048GBN/A$169,205.36
GPU Option 1AMD EPYC96768GB48GB L40S × 4$152,583.64
GPU Option 2AMD EPYC128768GB94GB H100NVL × 4$261,386.58
info
  • Prices are approximate from recently provided quotes. These numbers are subject to bulk pricing discounts as well as supply-chain disruptions.

Included Services​

  • Systems are tailored to integrate into the supercomputers seamlessly
  • 5-year warranty on the server hardware and upkeep
  • Customized configuration options can be done on a case-by-case basis
  • Tax information

Community Use​

To offset the ongoing cost of managing privately owned hardware, owners make their nodes available to the ASU community whenever they sit idle. Two QoS options make this possible: the private QoS, and the public QoS in the htc partition. QoS stands for Quality of Service: the scheduling rules that set a job's time limit, priority and preemption.

Details

Owning group (grp_labname QoS, general partition). Members of the owning lab get a QoS created for their group. Their jobs may run for up to 30 days, have practically no effect on the group's fairshare, and always start ahead of every other job waiting for the lab's nodes.

Everyone else, long jobs (private QoS, general partition). Other users may run on idle privately owned nodes for up to 14 days on Sol or 7 days on Phoenix. These jobs are preemptible: if the owning group needs the node, the job is cancelled (not requeued) once it has run for at least 30 seconds.

Everyone else, short jobs (public QoS, htc partition). Jobs of up to four hours may run on privately owned nodes with no risk of preemption. When the owners need the node, their job starts as soon as the htc job finishes.