Computational Resources
|
Node
|
Number
|
Logical cores
|
Memory
|
GPU
|
GPU Architecture
|
Temp Disk
|
Networking
|
FP64 Performance 2
|
Partitions
|
|---|---|---|---|---|---|---|---|---|---|
|
EPYC Genoa CPU
|
16
|
384
|
768GB -1.5 TB
|
-
|
|
Yes
|
50 Gb
|
9.2 TFLOPS
|
fairshare ,unsafe
|
|
EPYC Rome CPU 3
|
24
|
256
|
1 TB
|
-
|
|
Yes
|
50 Gb
|
5 TFLOPS
|
priority ,unsafe
|
|
H100 Quad GPU
|
4
|
128
|
1 TB
|
4
|
NVIDIA HGX
|
Yes
|
100 Gb
|
136 TFLOPS
|
gpufairshare ,gpuunsafe
|
|
A100 Quad GPU
|
2
|
96
|
500 GB
|
4
|
NVIDIA HGX
|
Yes
|
100 Gb
|
39 TFLOPS
|
gpupriority ,gpuunsafe
|
|
B6000 Dual GPU1
|
8
|
64
|
550 GB
|
2
|
PCI-E
|
Yes
|
100 Gb
|
3.9 TFLOPS
|
gpupriority ,gpuunsafe
|
|
A100 Dual GPU
|
3
|
32
|
250 GB
|
2
|
PCI-E
|
Yes
|
50 Gb
|
20 TFLOPS
|
gpupriority ,gpuunsafe
|
|
A40 Dual GPU
|
25
|
32-192
|
250 GB-1.5 TB
|
2
|
PCI-E
|
Yes
|
50 Gb
|
1 TFLOPS
|
gpupriority ,gpuunsafe
|
|
T4 Quad GPU
|
4
|
48
|
512 GB
|
4
|
PCI-E
|
No
|
25 Gb
|
1 TFLOPS
|
gpupriority ,gpuunsafe
|
|
T4 Dual GPU
|
20
|
28
|
384 GB
|
2
|
PCI-E
|
No
|
25 Gb
|
0.5 TFLOPS
|
gpupriority ,gpuunsafe
|
|
Total
|
106
|
15360
|
~65 TB
|
152
|
|
|
|
|
|
- When referring to B6000, we are referring to the NVIDIA GPU named "NVIDIA RTX PRO 6000 Blackwell Server Edition 450w".
- These are rough calculations to determine order of magnitude. B6000 and H100 have significant architecture advantages for AI/ML, depending on workload.
- The EPYC Rome systems lack AVX512 support, so workloads that do media processing (such as image processing) may want to move to Genoa.