GPU Research Methodology and Reproducibility
Requant · Published 15 November 2026 · Updated 19 November 2026.
The nine GPU and cryptography reports and fifteen Consumer Blackwell studies ask where exact GPU arithmetic, cryptographic systems and complete application costs meet. A fast kernel is useful evidence, but it is not automatically a faster service, solver, recovery operation or prover. This page explains how to read the measurements and reproduce their bounded conclusions.
Questions and decisions
Each report states a concrete question, the experiment that addresses it, the checks applied to its outputs and the resulting implementation decision. Negative outcomes are retained: an optimization can be exact and still be slower, too memory-intensive or too small to matter in an application. A resource gate is reported as a gate, not as a completed large-scale experiment.
These are organizational technical reports. Independent seeded cohorts are internal repetitions on the documented server; they are not independent external replication, peer review or a security audit. No DOI, academic affiliation or external endorsement is claimed.
Hardware and experimental conditions
The main GPU is an NVIDIA RTX 5060 Ti with 16 GB of memory, using the SM120 instruction path. The heterogeneous erasure experiment also uses a CMP 50HX with 20 GB, using its supported SM75 backend. The server has 64 GB of host memory. CUDA, driver, upstream revisions, parameters and build identities are recorded in the repositories rather than inferred from a product name.
GPU timing campaigns run serially. The original single-GPU studies retain their documented server-service conditions. Four GPU-consuming services were temporarily stopped for the exclusive two-GPU erasure experiment and restored after the last cohort. Clock and power-limit settings were unchanged. FHE energy observations precede that service change and are not pooled with the erasure campaign.
During the two-GPU campaign, read-only link metadata reported Gen3 x8 for the RTX and Gen2 x1 for the CMP. That metadata is not a measured bandwidth result. The separately timed recovery paths establish the actual transfer cost for that deployment. See erasure coding and two-GPU scaling.
Timing boundaries
- Resident device time: inputs are already on the GPU. Setup, allocation and validation may be outside this boundary and are identified separately. A result at this boundary does not include moving an application's host data.
- Resident host wall: host dispatch, synchronization and joining workers may be included, while validation downloads remain outside the timer. This is distinct from device-event time.
- Pinned offload: the caller already owns reusable pinned input/output buffers. Allocation or ordinary-memory staging is not silently included.
- Staged full path: the stated boundary includes ordinary-to-pinned copies, uploads, computation, downloads and output assembly where those operations are part of the workload.
- Complete process: startup, model/library loading and the native task are charged as specified. External independent verification is either included or explicitly reported separately.
- Setup: compilation, key generation, commitment preparation, matrix generation and planning are recorded separately where appropriate. Their exclusion from steady-state latency is an explicit application assumption.
The report tables label the boundary they measure. For example, the verifiable-inference study includes process/model setup in generation time and records CPU proof verification separately; it does not benchmark a persistent resident inference server. The Wagner comparison discloses that GPU verified time includes independent verification while the CPU process timer excludes subsequent Python proof replay.
Correctness before performance
Exact arithmetic is compared with independent CPU references. Erasure recovery checks every output byte. Finite-field matrix experiments check every output element or final dependent state. Binary-field experiments compare complete transcripts, rather than a small output checksum. Sparse solvers regenerate their characterized matrices and independently replay residual and vector-independence checks.
Proof experiments retain the original native verifier, declared public statements and acceptance coverage. Changed-proof controls must reject. Native acceptance does not by itself prove formal security, hiding, arbitrary-model compatibility or the correctness of every possible parameter choice.
The PoUW report retains the historical conformance failure and tests stricter share targets. The sumcheck report retains invalid first-campaign timing records as instrumentation controls, excludes them from authoritative timing conclusions and uses explicit streams, graph-completion fences and cleared output buffers for the corrected campaign.
Samples, cohorts and uncertainty
Most fixed-input latency experiments use five warm-ups and thirty measured observations, with independent seeded cohorts and alternating variant order where stated. Raw observations and median/p95 summaries remain available. These sample counts do not apply to every experiment:
- Sparse complete-solve comparisons use one warm-up and three measured fresh processes per configuration, with a retained process cap.
- Algebra search uses bounded trials and retains unsuccessful searches at their budgets. Its exact paired policy analysis is explicitly exploratory and post-hoc.
- Wagner stochastic cohorts use distinct nonces and retain zero-yield inputs; those are not thirty repeated latency observations of one header.
- The two native PoUW ZK certificates are feasibility observations, not a thirty-sample proving-throughput study.
- FHE energy has two complete-process observations per variant with reversed order, not a calibrated energy-efficiency distribution.
Small percentage differences can reflect order, system variation or an application bottleneck. We do not promote a sign-changing sub-percent observation into an established speedup, or extrapolate two internal cohorts into a universal hardware ranking.
Memory and energy
Allocated arrays, observed allocator reservation, sampled process allocation and a true instantaneous whole-device peak are different quantities. Their boundaries are named. NVML samples can miss a brief peak. A source-derived storage budget is not an implemented, correct kernel.
The full inference target requires at least 280 GB available host memory. The unchanged script stopped before any GPU phase on this 64 GB host. Summing two cards' device memory does not satisfy that requirement, and the guard was not bypassed.
The FHE energy study records both a GPU cumulative-energy counter and a sampled-power integral over complete processes. The methods disagree by about 6–9% under load; idle observations also disagree. No idle subtraction or calibrated wall-meter claim is made. No readable CPU-package energy counter was found for the account, so the report does not establish whole-system or CPU-versus-GPU efficiency.
Blackwell transfer and service measurements
The lossless transfer study reports logical wire bytes, requested buffers and host service separately. Its 20 GiB stream reuses a pair of pinned inputs. The board-energy study compares counter deltas over paired windows and measures latency from the offered request time. Its held-out selection outcome is distinct from the earlier full-process CKKS energy campaign.
Evidence, code and reproduction
Every web report identifies an immutable repository revision and its supporting reports. The Blackwell studies distinguish downloadable CPU replay data and compact receipts from native reproduction that depends on retained input or runtime archives. Source, generated code, binaries, artifacts and results are bound by hashes where the study records them. Historical raw observations and failed controls are preserved rather than rewritten to fit the conclusion.
Run the report's documented build and study commands in its required CUDA/library environment, then use its CPU verification scripts. The repositories' CI checks the arithmetic references, public artifacts, retained provenance and derived tables available to a hosted CPU runner. It does not independently rerun GPU timings. Downloading a retained JSON record is evidence replay, not a new hardware replication.
For citation, use the canonical report URL, visible publication date and its evidence revision. The earlier GPU and cryptography collection appeared on 15 November 2026, and the Blackwell collection appeared on 19 November 2026. Each date identifies the website publication. Later substantive changes should update the visible modification date and the corresponding structured metadata together.
Relationship to Requant
The GPU series extends Requant's open experimental work; it does not change the deployed TNet work function. A cryptographic or inference benchmark is not evidence that miners are performing an external customer's useful AI work. ASIC resistance, mainnet readiness, production security and financial returns do not follow from these studies.
For the separate protocol programme, read the TNet research programme, measurements and open questions. Return to all research to browse the GPU, protocol and network reports.