Nvidia Vera CPU Review: High Memory Bandwidth for AI and Server Workloads

| 2 Min Read
Nvidia Vera pairs strong Python performance and high memory bandwidth with competitive results across general server workloads.

Vera’s case rests on more than AI marketing

Nvidia’s Vera is a serious attempt to make the CPU a first-class part of the AI data center. That matters because agentic systems can keep CPUs busy with tool calls, code execution and data handling even when GPUs are doing the model inference. Nvidia says Vera was built for this shift. The timing invites skepticism: demand for agentic workloads has surged only recently, and the company’s framing can make a fast-moving trend sound like a design target it anticipated all along.

The performance evidence is more persuasive than that story. There still aren’t standardized benchmarks for the full agentic loop, so no one can yet measure Vera’s suitability with a single definitive test. But the workloads we can test point to strengths that fit the job. Vera performs especially well in Python, a foundation of much AI software, and delivers high memory bandwidth at both the socket and system level. That bandwidth could matter when many agents are querying large databases at once.

Vera also has to earn a place in a crowded server market. Custom Arm processors are gaining ground, while AMD and Intel bring established data center platforms and extensive customer relationships. The surprise in our results is that Vera isn’t limited to agentic-adjacent tasks: it often matches or beats comparable AMD Turin and Intel Granite Rapids-AP systems on general server workloads. X86’s greater core density can still help in frontend tasks. Vera’s advantage is more apparent in backend work that moves and processes large amounts of data.

Nvidia Vera specifications
Cores / Threads88 / 176
L2 cache2 MB per core
L3 cache164 MB per CPU
MemoryUp to 1.5 TB SOCAMM2 LPDDR5X
Memory speedUp to 9600 MT/s
Memory bandwidthUp to 1.2 TB/s aggregate; 14 GB/s per core
PCIe88 PCIe 6.4 lanes for the CPU alone; 96 PCIe 6.4 lanes in Vera Rubin; bifurcation down to x2; CXL 3.1
Configurable TDP250W–450W

Vera is Nvidia’s first processor built around its own custom Olympus core, and it improves on Grace, which used an off-the-shelf Arm core design. The 88-core chip follows the Armv9.2-A architecture and supports two threads per core through Nvidia’s “spatial multithreading.” Its 10-wide decode is similar to Arm’s AGI, and it pairs the core with Nvidia’s second-generation Scalable Coherency Fabric (SCF).

SCF is central to the design: its mesh distributes L3 cache across the die and provides up to 3.4 TB/s of on-die bandwidth. Our Vera deep dive covers the Olympus core and SCF in more detail. The memory system is another standout. Replaceable SOCAMM2 LPDDR5X runs at up to 9600 MT/s, for as much as 1.2 TB/s of total bandwidth—about 14 GB/s per core, more than twice the figure for AMD’s dense Turin designs.

Those specifications set expectations. The more useful question is how Vera behaves in real workloads, and where it fits against today’s x86 servers before AMD Venice and Intel Diamond Rapids begin arriving.

Test platforms
Test platformMemoryTested processors
2P Nvidia MGX Vera reference platform16 × 96 GB (1.5 TB) Samsung LPDDR5-9600 ECC SOCAMM2Nvidia Vera
2P Supermicro SYS-222HA24 × 64 GB (1.5 TB) Micron DDR5-8800 ECC MRDIMMIntel Xeon 6980P
2P AMD Volcano reference platform24 × 64 GB (1.5 TB) SK hynix DDR5-6000 ECC RDIMMAMD EPYC 9965, EPYC 9755

What an agentic workload asks of a CPU

“Agentic CPU” still means different things to different people. Nvidia and AMD have helped define the term around concurrency and the varied work agents perform. Those are useful ideas, but the clearest definition comes from looking at how agent platforms are being deployed.

Systems such as Meta Muse and OpenAI Dots use persistent agents that receive a slice of cloud compute. CPU-only racks coordinate with inference servers, where the models run. In that setup, the important question isn’t how quickly one sandbox completes a task; it’s how well thousands of sandboxes keep working when the CPU is heavily loaded.

Some agent tools, including code compilation and large database queries, can demand substantial CPU time. Others—reading or writing files, running tests, or browsing the web in a headless browser—are modest in isolation. At scale, though, small per-call differences accumulate. CPU topology also affects whether performance holds up as more sessions run concurrently.

Concurrency is only part of the problem. Agent workloads vary, and the bursts can be separated by long idle periods. Mark Lohmeyer, Google Cloud’s VP and GM of AI and computing infrastructure, described the pattern to Tom’s Hardware with an example: an agent might help plan a vacation over a month, receive a burst of instructions, then sit idle until the user returns days later.

“The actual use of the underlying compute, it's happening in basically bursts, right? There's like a burst of use at the start, maybe for an hour, and then it goes away, and then a couple days later, [it] bursts again for some other purpose and goes away,” Lohmeyer said. “They're very long-running workloads, effectively long-running, but often idle, basically.”

That burstiness complicates scheduling. When several agents’ bursts overlap on a server already under contention, sessions can slow to a crawl. Their work may include compilation, graph traversal and large database queries, mixed with lighter tasks. Each agent runs its own loop, and those loops don’t line up neatly.

Agentic AI loop.

(Image credit: The University of Texas at Austin)

An August 2026 study, Architectural Implications of Agentic AI Workflows, by researchers at The University of Texas at Austin and Microsoft Azure, describes three CPU roles in an agentic system. The CPU schedules GPU inference, as it does for a one-shot LLM query; orchestrates agent state and communication; and runs tools and agent logic, often in Python.

Agent execution time.

(Image credit: The University of Texas at Austin)

The study illustrates contention as four agents send requests to a GPU, receive results and continue their work over less than 30 seconds. Even brief CPU tasks can collide, creating gaps in GPU activity while agents wait. That’s a reminder that faster inference hardware alone won’t keep an agent system busy.

Nvidia calls Vera an agentic CPU, but the industry still lacks a settled way to score an agentic loop. “The industry built a pretty good way to benchmark traditional workloads,” Lohmeyer said. “I think this is happening now for GPUs and TPUs for inference, for example. I think we're going to need the same type of industry-level benchmarks for agentic workloads that are probably still emerging.”

There are no standardized end-to-end agent benchmarks comparable to SPEC or TPC yet. We’re developing internal tests, and companies such as Modal have published vendor-level results, but the industry has not settled on a common full-loop method. For now, we can test components that matter to agents: code compilation for coding systems, database performance for retrieval, JSON parsing with simdjson, and Python performance across a large part of the AI software stack.

That’s the basis for comparing Vera with AMD and Intel in established benchmarks, while leaving room to test full agent loops once a sound procedure is available. Nvidia provided remote access to a dual-socket Vera system and did not restrict our tests or coverage. We checked the platform configuration, then ran our own workload choices—including agent-relevant tests and general server benchmarks for Java, cryptography, compression and rendering.

A server result depends on the job

Server CPUs don’t map neatly to client-PC recommendations. Operators have deeper access to the hardware and often tune software for a specific workload. A gaming recommendation can point to an AMD X3D chip, or a single-threaded performance lead can identify Intel’s latest flagship. Server choices depend much more on what the systems need to do.

That makes broad winner-and-loser verdicts less useful here. Hyperscalers run large fleets with different hardware for different tasks, while enterprises must weigh workload fit, total cost of ownership and many other business-specific constraints. Performance figures matter, but they don’t make the purchasing decision by themselves.

SPEC CPU 2026 estimates

SPEC CPU 2026 intrate has been the main published benchmark for Vera so far. We ran the full SPEC CPU 2026 suite to check Nvidia’s claims and examine floating-point performance. The results below are estimates because we aren’t submitting them to SPEC, whose reporting rules are strict. “Estimated” does not mean guessed: we followed the standard procedure and used the same compiler and flags across the CPUs.

We compiled with GCC 15.2, using common options including link-time optimization and -march/-mcpu settings to expose each processor’s supported instructions. We used the general-purpose jemalloc allocator throughout.

Nvidia Vera SPEC intrate.

(Image credit: Tom's Hardware)

SPEC intrate runs copies of integer-heavy applications across available threads to measure throughput under simultaneous workloads. Vera’s overall result is close to Nvidia’s published figure: Nvidia reported a 3% lead over EPYC 9755, while our tests put the lead at 1.5%. That comparison needs context. The two-socket EPYC 9755 system has 512 threads, against 352 for Vera, and Vera’s per-core performance is markedly stronger.

Intel’s two-socket Xeon 6980P also has 512 threads, but its core count doesn’t put it in contention here. Switching to Intel’s oneAPI compiler improved its score by 10%. Even then, Vera running generic GCC remained 8.6% ahead. The results vary by test: Vera leads some workloads and trails in others, as the charts show. GCC and Python stand out as strengths, with Vera even beating AMD’s top-end EPYC 9965 in those areas.

Nvidia Vera SPEC intrate results.
Tom's Hardware
Nvidia Vera SPEC intrate results.
Tom's Hardware
Nvidia Vera SPEC intrate results.
Tom's Hardware
Nvidia Vera SPEC intrate results.
Tom's Hardware
Nvidia Vera SPEC intrate results.
Tom's Hardware
Nvidia Vera SPEC intrate results.
Tom's Hardware
Nvidia Vera SPEC intrate results.
Tom's Hardware
Nvidia Vera SPEC intrate results.
Tom's Hardware
Nvidia Vera SPEC intrate results.
Tom's Hardware
Nvidia Vera SPEC intrate results.
Tom's Hardware
Nvidia Vera SPEC intrate results.
Tom's Hardware
Nvidia Vera SPEC intrate results.
Tom's Hardware
Nvidia Vera SPEC intrate results.
Tom's Hardware
Nvidia Vera SPEC intrate results.
Tom's Hardware

Vera’s results depend on the workload

Vera’s strongest numbers come with a caveat: it doesn’t lead every test. SEAL’s crypto benchmark exposes weaker cryptography performance, consistent with our other results. Zstd tells a similar split story. Vera compresses well, but our separate tests found it slower at decompression, leaving it behind AMD and Intel overall.
Nvidia Vera SPEC fprate.
Tom's Hardware
Nvidia Vera SPEC fprate.
Tom's Hardware
Nvidia Vera SPEC fprate.
Tom's Hardware
Nvidia Vera SPEC fprate.
Tom's Hardware
Nvidia Vera SPEC fprate.
Tom's Hardware
Nvidia Vera SPEC fprate.
Tom's Hardware
Nvidia Vera SPEC fprate.
Tom's Hardware
Nvidia Vera SPEC fprate.
Tom's Hardware
Nvidia Vera SPEC fprate.
Tom's Hardware
Nvidia Vera SPEC fprate.
Tom's Hardware
Nvidia Vera SPEC fprate.
Tom's Hardware
Nvidia Vera SPEC fprate.
Tom's Hardware
Nvidia Vera SPEC fprate.
Tom's Hardware
Floating-point workloads are a different story. Nvidia’s initial benchmarks highlighted SPEC intrate, so we wanted to see how Vera handled floating-point work too. It leads the results overall, with particularly strong showings in several tests. Our estimated PALM figures point to Vera’s second-generation Scalable Coherency Fabric (SCF) as a strength; ROMS stresses memory bandwidth, and that advantage also appears in Marian and LBM. FemFlow and CACTUS are notable exceptions, but they don’t change the broader result.
Nvidia Vera SPEC intspeed.
Tom's Hardware
Nvidia Vera SPEC intspeed.
Tom's Hardware
Nvidia Vera SPEC intspeed.
Tom's Hardware
Nvidia Vera SPEC intspeed.
Tom's Hardware
Nvidia Vera SPEC intspeed.
Tom's Hardware
Nvidia Vera SPEC intspeed.
Tom's Hardware
Nvidia Vera SPEC intspeed.
Tom's Hardware
Nvidia Vera SPEC intspeed.
Tom's Hardware
Nvidia Vera SPEC intspeed.
Tom's Hardware
Nvidia Vera SPEC intspeed.
Tom's Hardware
Nvidia Vera SPEC intspeed.
Tom's Hardware
Nvidia Vera SPEC intspeed.
Tom's Hardware
Nvidia Vera SPEC intspeed.
Tom's Hardware
Nvidia Vera SPEC intspeed.
Tom's Hardware
We also ran SPEC CPU 2026’s intspeed suite and estimated results using every available thread. That gives a useful look at maximum throughput, but it’s an artificial setup: server users are unlikely to dedicate every thread to one application.

Nvidia Vera Python and AI benchmarks

Python performance is a meaningful signal for agentic AI workloads. There’s no single language for AI agents, but Python—and, to a lesser extent, TypeScript—underpins frameworks such as LangGraph, CrewAI, and AutoGen. Vera’s performance here is a real strength, though it shouldn’t be mistaken for a complete measure of AI capability.

Nvidia Vera PyBench.

(Image credit: Tom's Hardware)
On PyBench’s average test times, Vera finishes ahead of the field, 4% faster than AMD’s EPYC 9755. That chip leads Intel’s Xeons and the larger EPYC 9965. The pattern makes sense for these basic Python functions: much of the work is single-threaded, so the 9755 pair outpaces the 9965 pair across our Python tests.
Nvidia Vera PyPerformance.
Tom's Hardware
Nvidia Vera PyPerformance.
Tom's Hardware
Nvidia Vera PyPerformance.
Tom's Hardware
Nvidia Vera PyPerformance.
Tom's Hardware
Nvidia Vera PyPerformance.
Tom's Hardware
Nvidia Vera PyPerformance.
Tom's Hardware
Nvidia Vera PyPerformance.
Tom's Hardware
Nvidia Vera PyPerformance.
Tom's Hardware
Nvidia Vera PyPerformance.
Tom's Hardware
Nvidia Vera PyPerformance.
Tom's Hardware
Nvidia Vera PyPerformance.
Tom's Hardware
Nvidia Vera PyPerformance.
Tom's Hardware
Nvidia Vera PyPerformance.
Tom's Hardware
Nvidia Vera PyPerformance.
Tom's Hardware
Nvidia Vera PyPerformance.
Tom's Hardware
Nvidia Vera PyPerformance.
Tom's Hardware
Nvidia Vera PyPerformance.
Tom's Hardware
Nvidia Vera PyPerformance.
Tom's Hardware
Nvidia Vera PyPerformance.
Tom's Hardware
Nvidia Vera PyPerformance.
Tom's Hardware
Nvidia Vera PyPerformance.
Tom's Hardware
Nvidia Vera PyPerformance.
Tom's Hardware
Nvidia Vera PyPerformance.
Tom's Hardware
Nvidia Vera PyPerformance.
Tom's Hardware
Nvidia Vera PyPerformance.
Tom's Hardware
Nvidia Vera PyPerformance.
Tom's Hardware
PyPerformance widens the view, covering interpreter startup, regular expressions, and complete Python applications. Vera remains at or near the top in nearly every test, with a few weaker results in tests aimed at web servers. For teams building agent systems today, that’s a useful advantage. Python is central to the current stack, even as other languages gain ground, but these scores are only an early indicator of performance in a fast-changing field.
Llama.cpp Nvidia Vera.
Tom's Hardware
Llama.cpp Nvidia Vera.
Tom's Hardware
Llama.cpp Nvidia Vera.
Tom's Hardware
Llama.cpp Nvidia Vera.
Tom's Hardware
We also tested Llama.cpp with a CPU BLAS backend, using gpt-oss-20b-q8 to measure text generation and prompt processing at different token lengths. Treat the prompt-processing figures with care: a GPU handles much of that work in a typical setup, although the CPU still contributes, particularly to tokenization. For prompt processing, a single Xeon 6980P leads, followed by a two-socket 6980P system. Intel’s AI-specific instruction-set extensions may explain the result. AMD is expected to expand support for those instructions with Zen 6, which could narrow the gap.

Nvidia Vera memory and compression benchmarks

Memory bandwidth is one of Vera’s clearest advantages, and STREAM gives that claim a strong showing. Vera’s memory bus is more than twice as wide as Grace’s, and it uses LPDDR5X-9600. Nvidia rates the interface for up to 1.2 TB/s of bandwidth, with up to 1.8 TB/s between Vera sockets.
Nvidia Vera Stream performance.
Tom's Hardware
Nvidia Vera Stream performance.
Tom's Hardware
Nvidia Vera Stream performance.
Tom's Hardware
Nvidia Vera Stream performance.
Tom's Hardware
The comparison is striking against AMD’s DDR5-6000 support and Intel’s MRDIMMs, which reach up to 8,800 MT/s. Vera leads even without tuning. After our first STREAM run, Nvidia recommended the OpenMP “spread” policy, which spaces processes apart to limit cache and memory contention. That setting slightly reduced AMD and Intel’s results while giving Vera a large bandwidth boost. The size of that change makes the tuning choice relevant when comparing the scores. AMD’s Venice and Intel’s Diamond Rapids are both expected to raise memory bandwidth, so the current gap may shrink.
Nvidia Vera MBW.
Tom's Hardware
Nvidia Vera MBW.
Tom's Hardware
Nvidia Vera MBW.
Tom's Hardware
Nvidia Vera MBW.
Tom's Hardware
Vera also posts strong results in the straightforward MBW bandwidth test across copy sizes. CacheBench is less flattering: Vera falls behind in the combined read, modify, and write test. That weakness is real, but modest, and the much stronger memory-bandwidth results outweigh it in the broader picture.

Compression results depend on the workload

Vera’s compression results are a reminder that one benchmark rarely tells the whole story. With Zstd, it leads at both level 3 and level 19, across standard and Long Mode runs. Decompression is less favorable: AMD’s Turin processors pull ahead, leaving Vera around the level of—or behind—a dual 6980P system. And 7-Zip, which has traditionally favored Zen, puts both Turin configurations at the top; Vera trails the dual 6980P in both compression and decompression.
Nvidia Vera compression benchmarks.
Tom's Hardware
Nvidia Vera compression benchmarks.
Tom's Hardware
Nvidia Vera compression benchmarks.
Tom's Hardware
Nvidia Vera compression benchmarks.
Tom's Hardware
Nvidia Vera compression benchmarks.
Tom's Hardware
Nvidia Vera compression benchmarks.
Tom's Hardware
Nvidia Vera compression benchmarks.
Tom's Hardware
Nvidia Vera compression benchmarks.
Tom's Hardware
Nvidia Vera compression benchmarks.
Tom's Hardware
Nvidia Vera compression benchmarks.
Tom's Hardware

Java and database workloads show a more mixed picture

For teams building around Java, the relevant point is that Vera can be strong in specific kinds of computation without winning the overall score. We used Java SciMark, DaCapo, and Renaissance. In SciMark’s composite result, AMD’s Zen 5 based 9755 comes out ahead. Vera, though, holds its own in FFTs, sparse matrix multiplication, and Monte Carlo simulation, while trailing in dense LU factorization and Jacobi SOR. DaCapo covers a range of real Java workloads with a focus on native performance; Renaissance also spans varied Java workloads, with an emphasis on virtual machines.
Nvidia Vera java benchmarks.
Tom's Hardware
Nvidia Vera java benchmarks.
Tom's Hardware
Nvidia Vera java benchmarks.
Tom's Hardware
Nvidia Vera java benchmarks.
Tom's Hardware
Nvidia Vera java benchmarks.
Tom's Hardware
Nvidia Vera java benchmarks.
Tom's Hardware
Nvidia Vera Java benchmarks
Tom's Hardware
Nvidia Vera Java benchmarks
Tom's Hardware
Nvidia Vera Java benchmarks
Tom's Hardware
Nvidia Vera Java benchmarks
Tom's Hardware
Nvidia Vera Java benchmarks
Tom's Hardware
Nvidia Vera Java benchmarks
Tom's Hardware
Nvidia Vera Java benchmarks
Tom's Hardware
Nvidia Vera Java benchmarks
Tom's Hardware
Nvidia Vera Java benchmarks
Tom's Hardware
Nvidia Vera Java benchmarks
Tom's Hardware
Nvidia Vera Java benchmarks
Tom's Hardware
Nvidia Vera Java benchmarks
Tom's Hardware
Nvidia Vera Java benchmarks
Tom's Hardware
Nvidia Vera Java benchmarks
Tom's Hardware
Nvidia Vera Java benchmarks
Tom's Hardware
Nvidia Vera Java benchmarks
Tom's Hardware
Nvidia Vera Java benchmarks
Tom's Hardware
Nvidia Vera Java benchmarks
Tom's Hardware
Nvidia Vera Java benchmarks
Tom's Hardware
Nvidia Vera Java benchmarks
Tom's Hardware
Nvidia Vera Java benchmarks
Tom's Hardware
Nvidia Vera Java benchmarks
Tom's Hardware
Nvidia Vera Java benchmarks
Tom's Hardware
Database tests are a stronger showing. Vera performs well in both ClickHouse and SQLite: it trails only a single socket 6980P in the max thread SQLite test, then leads the equivalent 64 thread and 128 thread comparisons. ClickHouse adds a useful wrinkle. Vera is competitive with a dual EPYC 9755 on the second and third runs, but its first, cold cache run is nearly 40% faster. That gap suggests cache state has a substantial effect on the comparison; the results don’t explain the cause in full.
Nvidia Vera database benchmarks.
Tom's Hardware
Nvidia Vera database benchmarks.
Tom's Hardware
Nvidia Vera database benchmarks.
Tom's Hardware
Nvidia Vera database benchmarks.
Tom's Hardware
Nvidia Vera database benchmarks.
Tom's Hardware
Nvidia Vera database benchmarks.
Tom's Hardware
Nvidia Vera database benchmarks.
Tom's Hardware
Nvidia Vera database benchmarks.
Tom's Hardware
Nvidia Vera database benchmarks.
Tom's Hardware
Nvidia Vera database benchmarks.
Tom's Hardware
Nvidia Vera database benchmarks.
Tom's Hardware
The final test looks at ingesting data: simdjson measures JSON parsing and depends heavily on SIMD instructions. Turin has the built in advantage here, thanks to native AVX 512 execution. If your workloads span compression, Java, and databases, Vera’s results are uneven by task; the practical read is to match the processor to the operations you actually run, rather than infer a general winner from any one chart.
Source: Jake Roach · www.tomshardware.com

Comments

Please sign in to comment.