Compute Nodes
Work In Progress
Karolina NG is still WIP and subject to change.
Karolina NG is a cluster of x86-64 AMD based nodes built with the Eviden technology. The cluster contains only accelerated compute nodes:
- 20 nodes
- 2560 cores in total
- 2x AMD EPYC Turin 9555, 64-core, 3.2GHz processors per node
- 1520GB DDR5 RDIMM per node
- 8x GPU accelerator AMD Instinct™ MI355X, 288GB HBM3E memory per node
- Rpeak (FP64) 628.8TFLOP/s per compute node
- Acn[01-20]
Processor Architecture
The AMD EPYC™ 9555 is a server/workstation processor with 64 cores, launched in October 2024, using the Zen 5 (Turin) architecture with Socket SP5. The 9555 has a TDP of 360W with a base frequency of 3.2GHz and a boost frequency of up to 4.4GHz.
- Family: EPYC™
- Architecture: Zen5
- Cores: 64
- Threads: 128
- L1 Cache: 80KB (per core)
- L2 Cache: 1MB (per core)
- L3 Cache: 256MB (shared)
- Instructions/Features: MMX, SSE, SSE2, SSE3, SSSE3, SSE4A, SSE4.1, SSE4.2, AES, AVX, AVX2, AVX-512, AVX-VNNI, VNNI/VEX, BMI1, BMI2, BFloat16, SHA, F16C, FMA3, AMD64, EVP, AMD-V, SMAP, SMEP, X2AVIC, AIBRS, SEV-SNP, 5-level NPT, SMT, Precision Boost 2
- Frequency: 3.2GHz
- Max turbo: 4.4GHz
- Process: 4nm, 6nm (I/O)
- TDP: 360W
GPU Accelerator
Karolina NG is equipped with an AMD Instinct MI355X GPU accelerator. The Instinct MI355X is a data center GPU, launched on June 12th, 2025. It is built on the 3 nm process and features 288 GB HBM3e memory, 16384 shading units, 1024 texture mapping units, and 0 ROPs. It also includes 1024 Matrix Cores which help improve the speed of machine learning applications with support for the FP4, FP6, INT8 and FP8 data formats.
| AMD Instinct MI355X | |
|---|---|
| Architecture | CDNA 4 |
| Shading Units | 16384 |
| TMUs | 1024 |
| Compute Units | 256 |
| Matrix Cores | 1024 |
| L3 Cache | 256 |
| GPU Memory | 288GB HBM3e |
| Memory Bandwidth | 8.19TB/sec |
| ECC Support | Yes |
| Rpeak FP64 | 78,64TFLOPS |
| TDP | 1400W |
