Phoronix benchmarks NVIDIA Vera 1P vs. 2P scaling on Linux 6.17

Phoronix published a standalone benchmark study comparing single-socket and dual-socket NVIDIA Vera CPU deployments across multi-threaded, NUMA-friendly workloads. Each Vera processor contains 88 Olympus cores with SMT for 176 threads per socket, yielding 352 threads total in a 2P configuration. The testing was performed remotely on Ubuntu 24.04 LTS running the Linux 6.17 kernel compiled with 64K page sizes.

The benchmark suite spans HPC, compression, cryptography, financial modeling, and rendering workloads including OSPRay, RocksDB, Speedb, QuantLib, OpenSSL, and Stockfish. Michael Larabel notes these results are isolated to Vera hardware without direct 1P/2P comparisons to AMD EPYC or Intel Xeon equivalents due to lack of matching 88-core products from those vendors. The data was collected during broader Vera testing that also examined SMT behavior and HPC competitiveness against rival server platforms.

Vera architecture and benchmark methodology

• Core configuration: Each NVIDIA Vera CPU provides 88 Olympus cores with simultaneous multithreading, delivering 176 threads per socket and 352 threads in a dual-socket deployment.
• Kernel environment: All benchmarks ran on Ubuntu 24.04 LTS with the Linux 6.17 kernel configured for 64K page sizes, a non-default setting relevant for large-memory server workloads.
• Workload selection: Tests were chosen specifically for high core/thread scaling and multi-socket NUMA awareness, spanning OSPRay rendering, Coremark, Helsing, Primesieve, Sysbench, Stream memory bandwidth, 7-Zip, Zstd compression, RocksDB/Speedb storage engines, QuantLib financial computing, DSP kernels, HPC benchmarks, OpenSSL cryptography, Stockfish chess engine, and ASTC texture encoding.
• Scope limitation: The study deliberately excludes cross-vendor 1P/2P comparisons because AMD EPYC and Intel Xeon lack exact 88-core product equivalents, making socket-to-socket scaling analysis vendor-specific to Vera.

Why it matters

These results provide early public data on how NVIDIA's Grace successor architecture scales beyond single-socket boundaries under Linux, informing data center procurement and software tuning decisions for ARM64 server deployments. For the Linux kernel and distribution ecosystem, the 64K kernel page size configuration used here highlights ongoing platform-specific optimization work for high-core-count ARM server processors entering the market.