NVIDIA Vera CPU: Great Silicon, Misleading Whitepaper

NVIDIA built a genuinely strong server CPU, then wrapped it in a whitepaper that a respected independent hardware outlet says doesn't hold up.

On August 6, 2026, Chips and Cheese published "NVIDIA's Vera Whitepaper Has a Thread Loose," a detailed teardown by George Cozma and Chester Lam of NVIDIA's 45-page marketing document for Vera — NVIDIA's first server CPU, and the CPU half of the Vera Rubin platform now headed to SpaceX's Starmind satellites. Their conclusion, in short: the silicon is impressive, and the whitepaper doesn't need the exaggeration it leans on to prove it.

TL;DR

Is NVIDIA's Vera CPU actually good?

Almost certainly yes. Strip away the marketing and Vera is built around Olympus, NVIDIA's first Arm v9.2 server core — 10-wide decode, up to two taken branches per cycle, six 128-bit SVE pipes, four load pipes and two store pipes, a 96KB L1 data cache, and a private 2MB L2 with roughly 10-cycle access latency. Olympus also ships with features that show real architectural ambition: value prediction, a neural branch predictor, and a "graph prefetcher" purpose-built for producer-consumer and pointer-chasing access patterns — the kind of irregular memory traffic that trips up simpler prefetchers in database and graph workloads.

The full compute die packs 88 cores monolithically, sharing a 164MB last-level cache across a 3.4 TB/s coherency fabric. The SoC pairs that with eight LPDDR5X (SOCAMM2) memory interfaces, up to 1.5TB of capacity, up to 1.2 TB/s of memory bandwidth, and a memory subsystem drawing roughly 50W when fully populated — competitive numbers by any measure.

Early independent testing backs this up, with caveats. Phoronix's Michael Larabel ran pre-production Vera hardware in a one-day test window under NVIDIA's chosen workload scope, no frequency or power monitoring permitted. Within those constraints, Vera's geomean landed 10% above a 5GHz EPYC 9575F, 1.55x a Xeon 6980P, and 1.63x Google's own Grace CPU — making it, per that test, the most performant Arm server CPU seen in public benchmarking so far. Chips and Cheese doesn't dispute any of this. Their critique starts exactly where NVIDIA's own document tries to translate that real hardware into competitive claims against AMD.

Is NVIDIA's whitepaper trustworthy?

No — not as published, and not without independent numbers to check it against. Chips and Cheese's piece is structured as a point-by-point fact-check of NVIDIA's 45-page document, and it finds problems in nearly every major comparison chart. Here's the breakdown.

1. "Spatial Multithreading" mischaracterizes how SMT actually works

NVIDIA's Figure 5 draws a contrast: "Traditional SMT (x86)" is illustrated as threads alternating through opportunistic time-sharing, wasting execution resources, while NVIDIA's own "Spatial Multithreading" is drawn as cleanly, statically partitioned resources — implying x86 SMT is inherently wasteful by comparison.

Chips and Cheese calls this a misrepresentation. Most modern SMT implementations, including on x86, already share thread-agnostic pipeline stages — execute and memory access — per-cycle across both threads rather than time-slicing sequentially. Per-cycle thread selection at fetch and decode can actually hide stalls more efficiently than NVIDIA's static partitioning approach, not less.

More notably, the analysis surfaces a real cost of Vera's own approach that the diagram omits entirely: an Olympus core reportedly takes roughly 10,000 cycles to transition back to single-thread mode once a sibling thread finishes — a real, software-visible penalty. NVIDIA's actual tradeoff is better quality-of-service and determinism, not necessarily better throughput. The diagram sells the second claim while hiding the cost of the first.

2. The "32 NUMA nodes" comparison is a straw man

NVIDIA's whitepaper claims large x86 systems can expose "as many as 32 NUMA domains," contrasted against Vera's clean one-per-socket design — framing NUMA sprawl as an unavoidable consequence of x86 architecture. Chips and Cheese points out this is an optional AMD EPYC configuration choice (NPS4, NPS2, NPS1, NPS0, or treating each LLC as a NUMA node), not the default and not the only option. NVIDIA presents an edge-case, maximum-granularity configuration as though it were baseline reality.

3. SPEC CPU 2026 compiler benchmarks are not "agentic benchmarks"

This is the piece's sharpest catch. NVIDIA selected four SPEC CPU 2026 integer components — CPython, GCC, LLVM, and Cppcheck — and labeled them "agentic benchmarks." Chips and Cheese's rebuttal is blunt: these are legitimate, CPU-heavy programs (a Python interpreter, two compilers, a static analyzer), but none of them are agents. No model serving tokens, no agent runtime, no sandbox or tool-calling loop — the term "agentic" doesn't describe what's being measured.

The numbers underneath compound the problem. NVIDIA's headline 1.7x-1.8x figures are normalized per physical core on a fully-loaded two-socket system. The full-system SPECrate2026 Integer totals tell a much narrower story: Vera scores 925 against the EPYC 9755 comparison system's 898 — a 3.0% system-throughput advantage, because Vera's system runs 176 physical cores against 256 for the EPYC system. Divided per core, Vera genuinely is about 50% faster across the full suite, and 70-80% faster on the cherry-picked four components — a real story, just far narrower than the headline chart implies.

4. Cross-ISA "IPC" comparisons without definitions aren't verifiable

NVIDIA claims up to 2.3x more branch predictions per cycle, 3.5x more taken branches per cycle, 2.4x more instruction-fetch ops per cycle, and 4.3x more "backend operations" per cycle versus EPYC — with no PMU event names, no raw counts, and no disclosed clock frequencies. Arm and x86 instructions aren't equivalent units of work, so a cross-ISA IPC comparison without retired-work and frequency context isn't independently checkable. It's a number, not a benchmark.

5. The memory bandwidth claim is overstated — and Chips and Cheese has the receipts

This is where the piece is most concrete. NVIDIA's whitepaper shows Vera reaching roughly 1.1 TB/s against an EPYC 9755 Turin baseline that levels off near 400 GB/s — implying roughly a 3x advantage.

Chips and Cheese had already independently tested the same Turin platform and measured ~570 GB/s, not 400. Recalculated against their own measured baseline, Vera's real advantage comes out to roughly 1.9x total bandwidth and ~2.8x per-core bandwidth — meaningfully lower than NVIDIA's claimed 3x and 4.1x. Both chips convert a similar 92-93% of their theoretical peak bandwidth into sustained real-world bandwidth, which means Vera's advantage is mostly a function of having roughly double the peak memory-interface bandwidth of a single Turin socket — not evidence of any inherent architectural superiority of a monolithic die over chiplet-based designs, as NVIDIA's prose implies.

The comparison also aged badly, fast. AMD launched its 6th-Gen EPYC lineup on July 23, 2026 — two days after NVIDIA published the Vera whitepaper — including a 96-core EPYC 9686F with 16 memory channels supporting DDR5-8000 or MRDIMM-12800, offering up to 1,024-1,638 GB/s per socket depending on SKU and memory type. Depending on configuration, that can exceed Vera's theoretical bandwidth outright. NVIDIA's chosen comparison generation was already the prior one by the time the ink was dry — a useful reminder for anyone tracking AMD's own AI infrastructure moves this year that the CPU/GPU bandwidth race resets every product cycle, not every product generation.

6. The PageRank chart is thin — and the RL chart isn't a benchmark at all

A "2.6x PageRank advantage" chart omits the benchmark parameters used, including a core count capped at 32 despite 88- and 128-core chips being available — making it effectively unreproducible. But the standout failure is Figure 24, a claimed "1.8x for RL training" advantage, illustrated only as a grid of small colored squares with zero labels: no model, no environment, no framework, no batch size, no power measurement, no repetition count. Chips and Cheese's verdict, quoted directly: "This is not a bad benchmark, it simply is not a benchmark at all."

What is "Spatial Multithreading"?

It's NVIDIA's marketing name for how each Olympus core divides its execution resources between two hardware threads — the Arm-world equivalent of what x86 calls simultaneous multithreading (SMT), or what Intel branded Hyper-Threading. NVIDIA's pitch is that static, spatial partitioning gives more predictable per-thread performance than opportunistic sharing. That's a real, legitimate design tradeoff — better determinism and quality-of-service guarantees can matter a lot for latency-sensitive server workloads. The problem Chips and Cheese identifies isn't the design choice itself; it's that NVIDIA's own diagram (Figure 5) misrepresents how competing x86 SMT implementations actually behave, and omits the ~10,000-cycle thread-mode-switch penalty that is Spatial Multithreading's own real cost. A fair comparison would show both designs' actual tradeoffs — determinism versus raw throughput-under-load — rather than drawing one as broken and the other as clean.

Why does this matter for AI infrastructure buyers?

Because Vera isn't a general-purpose server chip competing on spec-sheet bragging rights — NVIDIA is explicitly positioning it as the CPU host for AI and agentic-workload systems, the compute-adjacent half of racks built around Rubin GPUs. It's already the CPU NVIDIA and SpaceX picked for the Starmind AI1 orbital compute payload, and it's the CPU side of the Vera Rubin platform datacenter operators are budgeting against for 2027 deployments.

That context is exactly why the whitepaper's accuracy matters beyond hardware-enthusiast trivia. Datacenter capacity planning — rack density, cooling budgets, power provisioning, and the CPU-to-GPU ratio in a given AI cluster — runs on vendor performance and bandwidth claims that buyers frequently can't independently verify before signing a purchase order. An overstated 3x memory bandwidth figure that's really 1.9x, or a "1.7x agentic" headline that's actually a 3% full-system gain, changes the math on whether Vera-based systems are worth their premium over AMD alternatives for a given workload. Buyers pricing out datacenter construction and staffing around these platforms are making multi-year, multi-million-dollar bets on numbers that, per this analysis, need independent re-verification before they're treated as ground truth.

Chips and Cheese's own recommendation is the practical takeaway: NVIDIA should give independent reviewers unrestricted production hardware with full frequency and power telemetry, rather than leaning on curated marketing framing. As they put it, the actual silicon doesn't need the exaggeration to look competitive — which, if anything, makes the misleading framing a stranger choice than if the hardware were mediocre to begin with.

What people are asking

Does this mean Vera is bad? No — this is the opposite of a "the chip is bad" story. Chips and Cheese explicitly calls Olympus "genuinely strong silicon" with real architectural innovation. The critique is entirely about NVIDIA's marketing document, not the transistors.

Should buyers wait for independent benchmarks before trusting NVIDIA's numbers? That's Chips and Cheese's implicit recommendation — treat whitepaper comparison charts as a starting point, not a procurement input, until reviewers get unrestricted hardware with telemetry access. Phoronix's early pre-production numbers (10% over a 5GHz EPYC 9575F) are a better reference point than NVIDIA's own charts, and even those came with a one-day window and NVIDIA-controlled test scope.

Is this normal for chipmaker whitepapers? Cherry-picked baselines and normalized-per-core framing aren't unique to NVIDIA — most vendor marketing decks lean toward the most favorable framing available. What sets this critique apart is the sheer number of distinct issues found in one document (six separate categories) and the fact that at least one chart (Figure 24) doesn't meet the bar of being a benchmark at all.

Will AMD's 6th-Gen EPYC change this comparison further? Likely yes. The EPYC 9686F's 16-channel DDR5-8000/MRDIMM-12800 support, landing just two days after NVIDIA's whitepaper, already threatens to exceed Vera's theoretical bandwidth on paper depending on SKU. Expect NVIDIA's next whitepaper revision, or Rubin-generation follow-up, to update its comparison baseline — worth re-checking before relying on any single-generation chart.

The takeaway

Vera is a legitimately impressive first server CPU from a company that had never shipped one before — an 88-core Arm design with real architectural bets (value prediction, graph prefetching, a wide out-of-order core) and a memory subsystem that beats current x86 competition on raw bandwidth. None of that is in dispute. What Chips and Cheese's George Cozma and Chester Lam document is that NVIDIA's own whitepaper doesn't trust that story enough to tell it straight — reaching instead for a strawman SMT diagram, a cherry-picked NUMA edge case, mislabeled "agentic" benchmarks, undefined cross-ISA multipliers, an understated bandwidth baseline, and one chart that isn't a benchmark at all. For anyone sizing AI infrastructure around Vera Rubin systems, the lesson is simple: read the whitepaper for the architecture, and wait for independent, telemetry-backed benchmarks before you build a capacity plan on its comparison charts.

Related on explainx.ai

• SpaceX Q2 2026 earnings and the Nvidia Starmind partnership
• AMD and Anthropic: $5B, 2GW of MI450, and Claude bring-up on ROCm
• NVIDIA GB300: TSMC Arizona wafers and Amkor packaging
• AI companies hiring electricians and carpenters for datacenter buildouts
• NVIDIA Computex 2026: Nemotron 3 Ultra and the GB300 recap
• Stanford memory prices: DRAM, HBM, and NAND history
• Wafer: fastest GLM-5.2 inference on AMD MI355X

Primary source: Chips and Cheese — "NVIDIA's Vera Whitepaper Has a Thread Loose" by George Cozma and Chester Lam (August 6, 2026)

Technical specifications, benchmark figures, and analysis quoted here reflect Chips and Cheese's August 6, 2026 article and NVIDIA's published Vera whitepaper as of that date. Both are subject to revision; verify current figures against primary sources before making infrastructure purchasing decisions.