Aguasto Research · Revision 1Editorial record: Review and correction policy ↗
- SourceIdentify the document, version and supported assertion.
- SeparateDistinguish a specification, observation, analysis and scenario.
- EvaluateKeep units, operating scope and contrary evidence visible.
- ReviewCorrect and preserve the publication record.
Start from an identifiable primary source
Hardware references begin with the manufacturer’s exact device or platform documentation. Model references use publisher checkpoints, model cards and licenses. Runtime records use the project’s own documentation or repository. Regulatory and operator case analyses identify the issuing authority or disclosing company and the dated document. A news summary can help locate a document, but does not silently replace it as the authority for a technical field.
Each source record retains a title, URL, publication date when stated, verification date and the claims it supports. A null publication date means the reviewed page did not state one; the verification date is not substituted as its release date. Web documentation is mutable. An unavailable source or unresolved discrepancy stays visible as an unknown instead of being filled from a similar product.
Record first publication and subsequent revision separately when the source exposes both. A paper version, a model license effective date, a checkpoint commit and a verification date answer different questions. Retain the version actually consulted; a family name or changing main branch is not proof of the latest release or an executable configuration.
Keep evidence classes visible
A manufacturer specification is an attributed product claim with a variant and unit boundary. An operator disclosure reports what that operator said at a particular date. An independent measurement needs an identifiable method, system, workload and test conditions. Aguasto analysis explains an inference drawn from those materials; it does not become an independent measurement simply because the site presents it in a diagram.
A conceptual scenario explains a mechanism or explores assumptions. It has no claim to be an Aguasto operating result. Approved founder statements describe the supplied professional background without inventing completed campuses, owned megawatts, installed inventory or customer agreements. Unknown capacity, licensing support or environmental performance is described as unknown. No manufacturer or operator reference implies an Aguasto partnership.
Compare compatible units and categories
The hardware comparison separates individual accelerators from complete rack systems. A rack aggregate and an individual-device figure may both be accurate while answering different questions. Memory, bandwidth, power and throughput require an explicit scope; a peak mathematical-operation figure is not model tokens per second. A manufacturer roadmap or shipping statement does not establish Aguasto procurement or commissioned capacity.
Model comparison retains checkpoint identity, architecture, license, modality and documented context limits. Total parameters and active parameters answer different questions for mixture-of-experts models. Theoretical weight payload is shown separately from runtime memory. The comparison makes no fabricated quality ranking, measured latency claim or guarantee that a reference works under a particular hardware/software combination.
Keep precision, dense or structured-sparse treatment, device count and memory scope attached to each manufacturer peak. Preserve conflicting configurations as separate variants instead of constructing a synthetic specification. A field described as a maximum must point to evidence of that maximum, even when another page publishes the same number under a different power label.
Expose formulas, inputs and exclusions
The weight tool uses parameter count × bits per weight ÷ 8 for a theoretical payload. Decimal GB divides bytes by 1,000,000,000; GiB divides bytes by 1,073,741,824. Cache, activations, workspace, quantization metadata and runtime allocations remain outside that lower bound. Any estimate should be replaced by observed allocation for the exact accepted configuration before a procurement decision.
The electricity tool estimates IT power × PUE × operating hours × electricity rate within its declared energy-cost boundary. The token-cost tool divides a supplied total cost by supplied completed tokens, then multiplies by one million. Those tools are hypothetical scenarios, not Aguasto prices, invoices, margins or capacity offers. Inspect units, exclusions and sensitivity; do not add incompatible costs or count failed and accepted output interchangeably.
Measure the complete accepted workload
A reproducible test records hardware and topology, software and model versions, precision, context limits, cache policy, prompt/output distributions, arrival pattern, concurrency and test duration. It defines warm-up and steady-state treatment. Quality and service acceptance are agreed before optimizing throughput so that a faster response is not accepted merely because it skips required work.
Report latency distributions, time to first token, inter-token behavior, successful throughput, errors and retries at a named client or server boundary. State the denominator used for tokens and cost. Retain configuration and result artifacts with enough detail to repeat the experiment. Aguasto has not published an operated-service benchmark; the methodology and calculator examples are not substitutes for those measurements.
Respect physical, environmental and status boundaries
Track electricity authorization, facility integration, commissioned IT load and customer-service acceptance separately. A drawing of a module explains an interface; it does not prove a location, permit, installed system or commissioning outcome. Public case-study milestones are attributed and dated. Announced, financed, contracted and delivered scopes may overlap and must not be summed without an explicit common boundary.
Resource analysis identifies location, reporting period, IT/facility scope, water definitions and useful-output conditions. PUE, local water and embodied impacts are different quantities. Reuse and prefabrication can change selected activities while leaving transport, safety, maintenance and replacement requirements. No universal modular schedule, uptime result or environmental advantage is asserted without project-specific evidence.
For operator cases, distinguish disclosure date, observation or acceptance date, project scope, unit and reported state. A quarterly report can describe an after-quarter event, while a financing transaction can cover construction and debt repayment together. Preserve conflicting milestone wording and leave its reconciliation unknown rather than inventing a cause or selecting the most favorable status.
Review, correct and preserve the publication record
Typed records pass schema, chronology, related-link and source checks before entering the canonical registry. Draft records are excluded from publication routes and feeds. Substantive changes receive an actual modified date and revision; a build or deployment does not automatically refresh a source verification date. A dated disclosure remains a historical statement unless later evidence supports a changed status.
Corrections should identify the affected assertion, the replacement evidence and the resulting scope. The editorial standards distinguish re-review from a cosmetic edit and preserve original publication dates. Superseded resources should be merged or retired with an appropriate canonical destination rather than left as thin duplicates. To flag a factual issue, email contact@aguasto.com with the page, assertion and primary source; online inquiry delivery remains disabled.
Sources & verification
The register preserves each source's publication and verification date. Manufacturer claims and documentation are attributed; Aguasto has not run the cited products in production.
- Metrics — NVIDIA NIM Benchmarking 2.0.0 ↗
Client TTFT and end-to-end latency boundaries; ITL excludes first-token latency; denominator output tokens minus one; Different tools use different definitions. Publication: date not stated in the reviewed source. Verified: 2026-10-12.
- Supported Models — vLLM official documentation ↗
Architecture and feature support must be checked for the intended model path; Native and Transformers model implementations are distinct documented paths. Publication: date not stated in the reviewed source. Verified: 2026-10-13.
- Supported Hardware — NVIDIA TensorRT-LLM ↗
Documented NVIDIA hardware families; Hardware support is distinct from a model/precision qualification. Last modified: 2026-10-09. Verified: 2026-10-13.
- ROCm 10.1.0 compatibility matrix — AMD ↗
Firmware, driver and user-space version alignment; Configuration-specific operating-system and GPU requirements. Publication: 2026-08-26. Verified: 2026-10-13.
- Optimization and Tuning — vLLM ↗
Eligible chunked prefill and decode prioritization in V1; Cache pressure, preemption and recomputation. Publication: date not stated in the reviewed source. Verified: 2026-10-13.
- Parallelism and Scaling — vLLM official documentation ↗
Single-replica GPU and node boundaries; Tensor/pipeline parallelism, execution environment and communication dependencies; Documented internal-network isolation requirement. Publication: date not stated in the reviewed source. Verified: 2026-10-13.
- Data Centers and Telecommunication Facilities — ASHRAE 2023 Handbook, chapter 20 ↗
Air/liquid cooling and facility heat-rejection interfaces; Thermal envelopes and site-specific operating conditions. Publication: date not stated in the reviewed source. Verified: 2026-10-13.
- DCGM Diagnostics — NVIDIA DCGM documentation ↗
Deployment, software, topology, power/thermal and memory diagnostic scope; Diagnostics are not comprehensive hardware diagnosis, repair or RMA authorization. Publication: date not stated in the reviewed source. Verified: 2026-10-13.
- NVIDIA NVLink and NVLink Switch — official product reference ↗
Manufacturer describes a scale-up GPU-to-GPU fabric; Specifications are marked preliminary; generation/system boundaries apply. Publication: date not stated in the reviewed source. Verified: 2026-10-13.
- Collective Operations — NVIDIA NCCL documentation ↗
AllReduce, AllGather and ReduceScatter semantics; Rank order and device mapping matter. Publication: date not stated in the reviewed source. Verified: 2026-10-13.
- GPU troubleshooting — NVIDIA NCCL documentation ↗
GPU-to-GPU and GPU-to-NIC access depend on topology and software; PCIe peer access, driver and virtual/container configuration limitations. Publication: date not stated in the reviewed source. Verified: 2026-10-13.
- RCCL usage tips — AMD RCCL official documentation ↗
Eight-MI300X example has dedicated pairwise Infinity Fabric links; Peer access, topology, partition and channel choices are configuration-specific. Publication: date not stated in the reviewed source. Verified: 2026-10-13.
- GPUDirect RDMA and GPUDirect Storage — NVIDIA GPU Operator documentation ↗
RDMA GPU-to-peer-device path differs from direct storage path; Kernel/driver and DMA-BUF or nvidia-peermem prerequisites. Publication: date not stated in the reviewed source. Verified: 2026-10-13.
- Disaggregated Serving — NVIDIA Dynamo architecture documentation ↗
Prefill, KV transfer and decode are distinct phases; NIXL transfer and backend-specific metadata; Pool allocation and communication costs remain workload-dependent. Publication: date not stated in the reviewed source. Verified: 2026-10-13.
- PD Disaggregation — SGLang official documentation ↗
Cache reservation, transfer handshake and backend/layout restrictions; Policy support must be checked per communicating-worker revision. Publication: date not stated in the reviewed source. Verified: 2026-10-13.
- Networking Troubleshooting — NVIDIA NCCL documentation ↗
Interface selection and low-level fabric checks; Link state/rate, bandwidth and latency diagnose different failures; InfiniBand and RoCE diagnostics are distinct. Publication: date not stated in the reviewed source. Verified: 2026-10-13.
- Parameters and Best Practices — NVIDIA NIM Benchmarking ↗
Concurrency and request-rate load modes; Arrival rate can exceed capacity and grow queues; Prompt/output lengths and batch size matter. Publication: date not stated in the reviewed source. Verified: 2026-10-12.
- Revisiting Service Level Objectives and System Level Metrics in Large Language Model Serving ↗
Primary authors discuss incentive problems with SLO/system metrics; Smooth-goodput framework; no Aguasto replication. Last modified: 2025-10-29. Verified: 2026-10-12.
- MLPerf Inference: Datacenter — MLCommons ↗
Results include submitter, software, accelerator count and metadata; Power measured at full-system AC boundary. Publication: date not stated in the reviewed source. Verified: 2026-10-12.
- MLPerf Results Messaging Guidelines — MLCommons ↗
Identify benchmark version, division, category, verification status, scenario and chip count; Do not compare MLPerf with arbitrary non-MLPerf results. Publication: date not stated in the reviewed source. Verified: 2026-10-12.
- Core Scientific Form 8-K Exhibit 99.2 — project disclosure ↗
Approximately 590 MW contracted net critical IT versus 900 MW secured grid capacity; Approximately 350 MW energized and over 185 MW billed as of March 13 2026; Phased delivery and completion expectations are forward-looking. Publication: 2026-04-21. Verified: 2026-10-12.
- Core Scientific Announces Second Quarter 2026 Results ↗
395 billing MW at Q2 end and 437 as of mid-July; Approximately 1.1 GW total leased customer power capacity includes expanded portfolio; Potential contracted revenue distinct from quarter revenue. Publication: 2026-07-28. Verified: 2026-10-12.
- Optimizing LLMs for Speed and Memory — Hugging Face ↗
Parameter storage at float32 and BF16/FP16; Weight storage differs from long-context inference memory. Publication: date not stated in the reviewed source. Verified: 2026-10-12.
- Cache strategies — Hugging Face Transformers ↗
KV cache grows during generation; Static cache allocation, offloading and cache quantization trade-offs; Model-specific cache limitations. Publication: date not stated in the reviewed source. Verified: 2026-10-12.
- Bitsandbytes — Hugging Face Transformers ↗
Quantized linear modules and mixed compute/storage precision; Nonquantized modules and offloading constraints. Publication: date not stated in the reviewed source. Verified: 2026-10-12.
- CUDA semantics: Memory management — PyTorch 2.14 ↗
Allocated tensor memory versus allocator-reserved memory; Graph-private memory pools and caching allocator. Publication: date not stated in the reviewed source. Verified: 2026-10-12.
- Dedicated Model Inference — Together AI ↗
Provider distinguishes per-token shared service and GPU-hour dedicated resource billing; Dedicated and serverless are provider-specific service categories. Publication: date not stated in the reviewed source. Verified: 2026-10-12.
- Dedicated vs Serverless models — Together AI Knowledge Base ↗
Dedicated hardware has cost while idle; High utilization changes economic comparison. Publication: date not stated in the reviewed source. Verified: 2026-10-12.
- Autoscaling — Baseten official documentation ↗
Independent replicas; minimum/maximum capacity; Scale-to-zero and startup billing/cold-start latency; Running-replica billing boundary is provider-specific. Publication: date not stated in the reviewed source. Verified: 2026-10-12.
- Orca: A Distributed Serving System for Transformer-Based Generative Models — USENIX ↗
Historical iteration-level scheduling mechanism; Whole-request versus changing batch membership. Publication: date not stated in the reviewed source. Verified: 2026-10-13.
- Continuous batching — Hugging Face Transformers 5.11.0 ↗
Request handling, cache and scheduling controls; Cancellation and manager lifecycle. Publication: date not stated in the reviewed source. Verified: 2026-10-13.
- Paged Attention, IFB, and Request Scheduling — NVIDIA TensorRT-LLM ↗
Token budget, workspace and cache interactions; Context chunking; Utilization saturation and latency tradeoffs. Publication: date not stated in the reviewed source. Verified: 2026-10-13.
- Metrics — NVIDIA NIM Benchmarking ↗
Client-observed first-token and generation timing; Queueing and network measurement boundaries. Publication: date not stated in the reviewed source. Verified: 2026-10-13.
- Best Practices Guide for Energy-Efficient Data Center Design — DOE FEMP, July 2024 ↗
PUE measures supporting infrastructure rather than IT productivity; Annual PUE and site/source WUE distinctions; Heat-reuse host and backup heat rejection. Publication: 2024-07-26. Verified: 2026-10-13.
- Quantifying the impact on PUE and energy consumption when introducing liquid cooling — Vertiv/NVIDIA study summary ↗
Cooling changes can affect both sides of the PUE ratio; Vendor-modeled configurations are not universal measured savings. Publication: 2023-02-15. Verified: 2026-10-13.
- Recommendations for Measuring and Reporting Overall Data Center Efficiency, version 2 — DOE-hosted industry task force ↗
Dedicated and mixed-use measurement boundaries; Metering categories and disclosed assumptions; Comparisons need climate and reliability context. Publication: 2011-05-17. Verified: 2026-10-13.
- HPC Data Center Power Usage Effectiveness — National Laboratory of the Rockies ↗
Example of metered support loads and attributed shared loads; Site-specific ESIF reporting; not an Aguasto benchmark. Publication: date not stated in the reviewed source. Verified: 2026-10-13.
- Cooling Water Efficiency Opportunities for Federal Data Centers — DOE FEMP ↗
Closed equipment loop can reject to an evaporative tower; Annual site WUE definition; Treatment can add electricity and maintenance. Publication: 2019-01-09. Verified: 2026-10-13.
- Device Queries: Total Energy Consumption and Power Usage — NVIDIA NVML API Reference ↗
GPU energy counter in millijoules since driver reload; Power sampling behavior and unsupported-feature errors; Device telemetry does not meter all facility loads. Publication: date not stated in the reviewed source. Verified: 2026-10-13.
- Data Center Metering and Resource Guide — DOE Better Buildings ↗
Metering enables efficiency diagnosis and verification; Separate planning guidance from dated market statistics. Publication: 2016-07-28. Verified: 2026-10-13.
- United States Data Center Energy Usage Report: 2025 Update — LBNL, published June 2026 ↗
2030 reference case and sensitivity scenarios; Idle draw, utilization and lifetime assumptions; Consumption forecasts do not establish available grid capacity. Publication: 2026-06-18. Verified: 2026-10-13.
- Life Cycle Assessment of Dell PowerEdge R740 — Dell/thinkstep, June 2019 ↗
Cradle-to-grave product-study boundary; Configuration, geography and useful-life assumptions; Historical example; not an AI GPU-system inventory. Publication: date not stated in the reviewed source. Verified: 2026-10-13.
- Data center product decarbonization: Environmental Product Declarations — Schneider Electric ↗
Product declarations build on lifecycle assessments; Supplier embodied-impact information; Product declarations do not certify a whole facility. Publication: 2024-07-19. Verified: 2026-10-13.
- Sustainability Benefits of Prefabricated Data Centers: Core and Shell Carbon Reduction — Schneider Electric WP240, version 1 ↗
Representative electrical-room core-and-shell comparison; Clearing, freight, construction and materials boundary; Vendor scenario is not universal whole-facility savings. Publication: 2026-07-08. Verified: 2026-10-13.
- Emergence and Expansion of Liquid Cooling in Mainstream Data Centers — ASHRAE, 2021 ↗
Facility and equipment temperature boundaries; Liquid-to-air interim architecture; Wetted-material compatibility and maintenance. Publication: date not stated in the reviewed source. Verified: 2026-10-13.
- HPC Data Center Heat Capture and Reuse — National Laboratory of the Rockies ↗
ESIF exchanger and building/district heat interfaces; Heat supply and demand are distinct boundaries. Publication: date not stated in the reviewed source. Verified: 2026-10-13.
- Beyond the compressor-less hype: Matching data center cooling technology to climate reality — Vertiv ↗
Climate and coolant conditions constrain compressor-free cooling; Dry, adiabatic and mechanical operating modes; Site water and energy trade-offs. Publication: 2026-05-14. Verified: 2026-10-13.
- Facility Water Cycle and Emissions Crosswalk — DOE FEMP ↗
Water supply, distribution and treatment use energy; Illustrative energy-water factors require local context. Publication: date not stated in the reviewed source. Verified: 2026-10-13.
- Commissioning your data center for greater availability and safety — Vertiv ↗
Commissioning scope spans requirements/design/construction/operations; Vendor startup, integrated test and turnover are distinct. Publication: date not stated in the reviewed source. Verified: 2026-10-13.
- Prefabricated Data Center brochure — Vertiv ↗
Factory/site/service integration phases; Engineering, acceptance plans, installation and commissioning roles. Publication: date not stated in the reviewed source. Verified: 2026-10-13.
- Liquid Cooling Integration and Logistics White Paper, revision1.0 — Open Compute Project ↗
FWS/TCS/CDU terminology and liquid-to-liquid isolation; Manufacturing, transport, inspection and commissioning handoffs; Wetted materials and project-specific tests. Publication: date not stated in the reviewed source. Verified: 2026-10-13.
- Network edge enabled by prefabricated modular data centers — Schneider Electric ↗
Factory assembly and site work may proceed in parallel. Publication: 2022-05-17. Verified: 2026-10-13.
- Startup and Site Acceptance Test — Vertiv ↗
Electrical/mechanical startup checks; Functional testing in operational modes and integrated site acceptance. Publication: date not stated in the reviewed source. Verified: 2026-10-13.