Lighthouse AdvisorySLED AI Adoption Intelligence
← Back to results

From the NVIDIA edition of September 22, 2026

Independent researchMixedNew this fortnight

Signal65: Vera 1.64× per-core on agentic CPU tasks vs EPYC 9755 (not whole-server)

Signal65 · Independent AI infrastructure evaluation · Lab comparison (NVIDIA Vera vs AMD EPYC 9755)

NVIDIA
Publisher
Signal65
Original publication
September 15, 2026
Source retrieved
2026-09-23
Event date
2026-09-15
Read original source

What happened

Signal65 validated NVIDIA Vera on agentic host CPU work using Terminal-Bench 2 oracle (reference-solution) runs that exclude the LLM. Across 69 matched tasks under full load (one sandbox per physical core), Vera’s geometric-mean per-core task lifecycle was 1.64× faster than AMD EPYC 9755; 61/69 tasks favored Vera. Signal65 reproduced NVIDIA’s Vera pass profile hands-on, but the 1.64× x86 comparison remains based on NVIDIA’s paired measurements. Explicitly not measured: whole-server throughput, power, price, or memory capacity (88 Vera cores vs 128 EPYC cores).

Why it matters

Newly relevant as SLED teams evaluate Vera Rubin agentic stacks: GPU AgentX headlines omit CPU-bound sandbox time—this study bounds per-core CPU gains while warning against socket-level capacity claims.

Evidence and measured results

Date Sep 15, 2026; Terminal-Bench 2.0 oracle agent; 69 matched tasks geo-mean 1.64×; 61/69 favor Vera; 34 tasks ≥2×; 8 favor x86; hands-on Vera pass-profile match; 1.64× vs x86 still NVIDIA paired data; no whole-server/power/cost; 88 vs 128 cores.

Limitations and uncertainty

Oracle agent removes model variance; NVIDIA provided hardware and paired x86 data; subsetting to most agent-like 20 tasks (1.87×) is author judgment; PINNACLE CPU suite still developing.

Put this evidence to work

Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-09-23; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.

Sales

Role takeaway
Customer problem
Agent fleets stall on CPU sandbox work even when GPUs are fast.
Stakeholders
  • AI platform
  • CFO
  • infra architects
Discovery questions
  • Are agent loops CPU-bound on tools/tests?
  • Do you size by per-core or socket?
  • Will you accept paired-lab comparisons?
Value hypothesis
Per-core gains can matter if tool loops dominate and socket math still closes.
Potential engagement
Two-week host bake-off on a Terminal-Bench subset.
Unsupported claims
  • 1.64× whole-server
  • Lower $/agent guaranteed
  • Power win proven

Pre-sales engineering

Role takeaway
Fit
Strong for sandbox-heavy agent hosts; weak for inference-only nodes.
Architecture and integration
One sandbox per core pinning model used in study; separate from GPU inference path.
Prerequisites
  • Task corpus
  • Bare-metal or equivalent isolation
  • Core-count-aware capacity model
Constraints
  • Paired NVIDIA x86 data
  • No power/cost
  • 88 vs 128 cores
Security
Sandbox isolation assumed; CC not tested.
Proof of value
  • Subset geo-mean reproduced
  • Pass/fail tables
  • Socket math documented

Delivery

Role takeaway
Work
Run oracle Terminal-Bench subset on Vera and x86 candidates.
Dependencies
  • Hardware access
  • Harbor/Terminal-Bench tooling
Ownership
Platform performance engineering.
Skills and adoption
Perf engineers learn agent-host metrics beyond GPU.
Governance checkpoints
  • Disclose paired-data basis
  • No whole-server claim
  • PINNACLE follow-up planned
Acceptance criteria
  • Geo-mean reported
  • Core-count normalized view
  • Limitations published
Risks
  • Subset cherry-picking
  • Model variance excluded
  • Procurement overclaim

Implementation considerations

Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.

Architecture and integration

What must connect, and where does the AI sit in the workflow?

Size agent fleets on per-core sandbox latency and core count together; do not translate 1.64× into rack throughput without socket math and power.

Governance

Who approves, reviews and stays accountable for outcomes?

Require disclosure when vendor-paired measurements underpin independent write-ups; ask for whole-server PINNACLE follow-ons before procurement.

Security and privacy

What data, permissions and controls need testing?

Bare-metal sandbox methodology; no confidential-computing or multi-tenant isolation claims.

Accessibility and workforce

Who is affected, and what skills or accommodations follow?

Faster agent tool loops can improve developer assistant responsiveness; no a11y features claimed.

Procurement

What should contracts, pricing and exit terms secure?

Hold Vera CPU purchases pending whole-server/power/cost evidence and competitive PINNACLE runs.

Operating model

Which teams own the service once it runs?

Add CPU-host KPIs to agent platform SLOs alongside GPU tokens/s.

Publication history

  1. 2026-09-22NVIDIA · Issue 178 resources
Read preserved resource versions (JSON)

Stable resource ID: signal65-vera-agentic-cpu-1-64x-20260915