Compute Nodes
The following table compares the compute nodes on Frontera to older nodes on Stampede3, another HPC resource at TACC. Both systems are supported by the National Science Foundation, and both are intended to enable large-scale computational research in a wide variety of fields. But as mentioned, Frontera is meant to meet the very biggest computational challenges; therefore, as one might expect, its Intel Xeon processors have overall better capabilities than prior generations. The main contrast, though, comes in the sheer number of processors: size is what really distinguishes Frontera from Stampede3. The table below emphasizes the numbers and key properties of certain Intel Xeon-based nodes found in each system (note, more recent Stampede3 node types are not shown).
| Frontera1 | Stampede32 | |
|---|---|---|
| Compute Node Type | Intel Xeon Platinum 8280 ("Cascade Lake") |
Intel Xeon Platinum 8160 ("Skylake")3 |
| Total Nodes | 8,368 | 1,060 |
| Cores/Node | 56 (28 cores/socket with 2 sockets) |
48 (24 cores/socket with 2 sockets) |
| Hardware Threads/core4 | 1 | 1 |
| Clock Rate5 | 2.7 GHz Max Turbo: 4.0 GHz |
2.1 GHz Max Turbo: 3.7 GHz |
| Memory | 192 GB DDR4-2933 | 192 GB DDR4-2666 |
| Peak Memory Bandwidth6 | 282 GB/s | 256 GB/s |
| L1 Cache | 32 KB per core | 32 KB per core |
| L2 Cache | 1 MB per core | 1 MB per core |
| L3 Cache | 38.5 MB per socket | 33 MB per socket |
| Local Storage (/tmp) | 144 GB partition on 240 GB SSD |
144 GB partition on 200 GB SSD |
1 Frontera User Guide: System Architecture
2 Stampede3 User Guide: System Architecture
3 At one time, 1,736 Skylake nodes were a major component of TACC's Stampede2, along with 4,200 Intel Xeon Phi 7250 ("Knights Landing") nodes. In 2024, 1,060 of the Skylake nodes were incorporated into Stampede3.
4 As of this writing, hyperthreading is not enabled on Frontera, nor on Stampede3. In the past, it was enabled on Stampede2.
5 Max Turbo applies when only 1 core is active; the Turbo Boost rate declines if more cores are active and/or AVX instructions are used.
6 For the DDR4-2933 memory on Frontera, (2 sockets) x (6 channels) x (23.5 GB/s) =
282 GB/s; a similar calculation for Stampede3 Skylake nodes appears in the Code Optimization topic.