Benchmarks
Oenerga evaluates systems against deployment reality: wall power, workload relevance, throughput per rack, energy per token, active-state efficiency, and communication cost.
Benchmark philosophy
Enterprise and sovereign buyers do not operate on isolated component marketing figures. They evaluate systems based on what matters in production: how much work is delivered per rack, how much energy is consumed per useful unit of output, how communication scales, and how active state affects real cost.
The AI hardware industry has a benchmarking problem. Processor-level TOPS numbers routinely characterize systems that spend more energy moving data than computing with it. Peak throughput numbers are measured on microbenchmarks that do not reflect production inference patterns. Memory bandwidth numbers reflect theoretical maximums, not useful bandwidth under real workloads.
Oenerga's benchmark approach is built for the buyers who have been through that cycle. We measure what procurement teams pay for, under conditions they can audit, with methodology they can review before making infrastructure decisions.
Wall-power accounting
Energy per token measured at the system level — compute, memory, interconnect, and control included. Not chip-level thermal design power.
Application-level throughput
Tokens per second and requests per second on workloads representative of real serving patterns — not synthetic kernel benchmarks.
Rack-level outcomes
Throughput per rack and cost per rack are the relevant units for procurement decisions — not chip-level peak figures.
Explicit baseline comparison
Every Oenerga benchmark names the baseline system and describes the test conditions. Numbers without baselines are not numbers.
Performance targets
Architecture-level advantages expressed in the terms that matter to procurement and infrastructure teams. Full methodology available on request.
[X] values represent target performance ranges under defined workload conditions. Specific figures available under NDA to qualified technical reviewers. Baselines, methodology, and test conditions disclosed in full.
Methodology
Oenerga benchmark methodology is designed to support serious technical review. Baselines, workload classes, context lengths, model families, concurrency levels, and power accounting boundaries should be visible and reviewable. Optical and system power are included where relevant.
Infrastructure buyers evaluating Oenerga systems should not need to trust marketing copy. They should be able to review the test conditions, replicate the environment under their own infrastructure assumptions, and verify the numbers with their own technical teams. That is the standard Oenerga is building toward.
Benchmark reviews are conducted under appropriate confidentiality arrangements for qualified partners. Full methodology documentation — including model families, context lengths, concurrency levels, baseline system specifications, and power measurement methodology — is available for review before any procurement decision.
Explicit baseline systems
Every comparison names the reference system, its configuration, and the software stack used. Comparisons without named baselines are not published.
Defined workload classes
Workloads are defined by model family, context length, concurrency level, and serving pattern — not by a single representative token count.
Wall-power accounting
Energy measurements are taken at the system boundary. Optical, memory, interconnect, and control subsystem power are included in the total.
Application-level metrics
Primary output metrics — tokens per second, energy per token, latency at load — are measured at the application layer, not at the compute plane in isolation.
Rack-level outcome focus
Final reported results are normalized to per-rack and per-watt terms. Chip-level performance claims are only published alongside system-level context.
Reproducible evaluation workflow
Benchmark environments and test scripts are documented to allow partner teams to replicate evaluation conditions independently, with Oenerga support where needed.
Pilot evaluations
These evaluation scenarios reflect the workload classes and decision criteria relevant to Oenerga's primary buyer categories. Partner and deployment details are disclosed under appropriate confidentiality arrangements.
[Pilot Partner Name] evaluated LUCID against a conventional serving baseline for long-context inference under high session concurrency. Oenerga demonstrated [X]% lower energy per token and [X]× improved throughput per rack in the target workload regime.
A memory-heavy transformer serving workflow was used to compare state-handling efficiency and communication burden between Oenerga and a GPU-cluster baseline. Oenerga showed [X]% lower movement-related overhead and improved active-state utilization across tested context lengths.
A sovereign AI deployment scenario assessed rack-level economics, scalability, and architecture-level differentiation under realistic serving profiles. Oenerga delivered [X] in effective state density and [X]% lower projected infrastructure cost under the modeled serving profile.
Resources
Available to qualified technical and procurement teams. Contact Oenerga to request access or discuss any of the materials below.
Architecture overview, market positioning, and strategic differentiation. Designed for CTO and infrastructure leadership audiences.
Complete architecture documentation, benchmark methodology, and system specifications. For technical and procurement review teams.
End-to-end pilot evaluation framework: environment requirements, workload definition, measurement protocol, and success criteria.
Five-pillar architecture summary for engineering and infrastructure planning teams evaluating integration requirements.
Oenerga conducts technical benchmark reviews with qualified infrastructure teams. Baselines, methodology, and full test conditions are disclosed in advance.