Key takeaways
Vera targets agentic AI workloads in data centers.
Memory bandwidth reaches up to 1.2 TB/s, with up to 1.8 TB/s between sockets.
SPEC CPU 2026 scores are estimates, not submitted results.
AMD and Intel retain advantages in several server workloads.
A data-center CPU built for agentic AI
Nvidia positions Vera specifically for agentic AI data-center workloads . Its strongest benchmark areas—Python execution and memory throughput—align with software orchestration, data preparation and large-scale data handling in those workflows.
Python underpins agentic frameworks including LangGraph, CrewAI and AutoGen. Vera led the PyBench tests, ahead of the EPYC 9755, which in turn outperformed Intel’s Xeons and AMD’s larger EPYC 9965.
The comparisons used two-socket platforms with 1.5 TB of memory each : an Nvidia MGX Vera reference system with Samsung LPDDR5-9600 ECC SOCAMM2, a Supermicro Xeon 6980P system with Micron DDR5-8800 ECC MRDIMM, and an AMD Volcano platform with SK hynix DDR5-6000 ECC RDIMM.
Nvidia’s Vera CPU performance: bandwidth and SPEC results
Vera’s memory interface supports up to 1.2 TB/s through LPDDR5X , alongside up to 1.8 TB/s between sockets . STREAM and MBW testing showed strong throughput, although CacheBench placed Vera behind competitors in its combined read/modify/write test.
In SPEC CPU 2026 integer throughput testing, Vera finished 1.5% ahead of the EPYC 9755 system, compared with the 3% advantage in Nvidia’s white paper. The comparison paired Vera’s 352 threads against 512 threads on the AMD system, with Vera showing a particularly strong per-core advantage.
Intel’s Xeon 6980P gained 10% when retested with oneAPI-optimized binaries. Vera, using generic GCC, still finished 8.6% ahead in that comparison.
These SPEC results are estimates, not reported to SPEC . The tests followed standard procedures and used matching compilers and flags for the initial comparisons. Vera led many floating-point tests too, but not every workload.
HPC and databases shine; Java and crypto vary
Vera led discrete Fourier transform, Octave and CloverLeaf benchmarks, while database testing also showed strengths. It placed mid-pack in m-queens and behind processors with larger core arrays in Primesieve.
In SQLite, Vera led the 64-thread and 128-thread tests; only a single-socket Xeon 6980P beat it at maximum threads. ClickHouse throughput approached that of a dual EPYC 9755 system on later runs, while Vera’s first cold-cache pass was nearly 40% faster . It also remained competitive in simdjson parsing.
Java SciMark showed solid results in FFTs, sparse matrix multiplication and Monte Carlo simulations, but weaker dense LU decomposition and Jacobi SOR performance. DaCapo and Renaissance results were competitive, with AMD and Intel retaining advantages in particular workloads.
Vera trailed in some OpenSSL tests and John the Ripper password cracking. A dual Xeon 6980P system also beat it in both 7-Zip compression and decompression.
Compilation produced a sharp split: Vera completed a GCC build about 12% faster than dual EPYC 9755 processors, but finished last in LLVM compilation using both Makefiles and Ninja. These tests measured building the compilers themselves.
Rendering gains do not extend to every workload
Encoding and rendering results were mixed, reinforcing Vera’s position as a credible data-center competitor rather than an across-the-board replacement for x86 processors. Its Python, bandwidth and HPC results provide the strongest evidence of that challenge.
The EPYC 9755 led AV1 encoding, while Vera remained competitive. Vera also performed well in x265 without consistently beating the 9755. In rendering, it topped OSPRay real-time path tracing but trailed the leading competitors in some scientific visualization tests.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.
