Hardware
Instances on KVM@TACC are virtualized, but the underlying hardware may be of interest. For general compute flavors (flavors that start with “m1”), the underlying hardware is:
Component |
Details |
|---|---|
Processor |
Intel Xeon E5-2670 v3 @ 2.30GHz |
Chassis |
Dell PowerEdge R630 |
Network Adapters |
2 x 1 Gbps |
For GPU flavors, (those starting with “g1.h100”), instances will be scheduled on the following hardware:
Component |
Details |
|---|---|
Processor |
Intel Xeon Platinum 8468 @ 2.1GHz, 48C/96T, 16GT/s, 105M Cache |
Memory |
1 TB DDR5-4800 RAM |
Chassis |
Dell PowerEdge XE9640 |
GPU |
4 x NVIDIA HGX H100 4-GPU SXM 94GB HBM2e 700W GPU |
Network Adapters |
1 x 25 Gbps, |
Two of the H100 nodes on KVM@TACC support NVIDIA Multi-Instance GPU (MIG) partitioning, which splits a single H100 into up to 7 smaller, isolated GPU instances instead of allocating full-GPU passthrough. This raises the effective per-node capacity from 4 to up to 28 instances, making it easier to get GPU access for smaller jobs that do not need a full H100.
Attention
Standard KVM@TACC instance leases can run for up to 6 months, but leases
for GPU-attached flavors (g1.h100.*) are capped at 7 days (1
week) — the same duration limit that applies to bare metal host leases.
This timeboxing keeps GPU capacity available to more users over time. See
the Chameleon FAQ
for the authoritative policy.