Flopper.io Flopper.io
Data quality

Methodology

Flopper.io is the independent database for comparing AI accelerators, rack-scale systems, cloud GPU pricing and AI datacenters. This page states, in the same words the code uses, how every figure is sourced, normalised and labelled.

Database totals as of 6 September 2026: 290 accelerators, 121 systems, 786 datacenters across 37 countries, 46 providers tracked, 191 live on-demand price listings.

What we count

  • Accelerator is the umbrella term for every part in the AI accelerator database: GPUs, TPUs, NPUs, LPUs, wafer-scale engines, custom ASICs and APUs. Every row carries an accelerator type; "GPU" in a headline count means the GPU subset. Entity pages stay under /gpu/ for historical reasons.
  • Accelerator vendor designs the silicon (NVIDIA, AMD, Google, Huawei). System vendor builds the server or rack (Dell, HPE, Supermicro). Cloud provider rents it (CoreWeave, Lambda, AWS). Datacenter operator runs the facility. One company can be several of these; each page says which role it describes.
  • A server system is one node (DGX H100, XE9680). A rack-scale system is a rack sold as one unit (GB200 NVL72, Helios). An AI datacenter is a facility.
  • Counts exclude tombstoned and redirected entries and near-duplicate rows that defer to a canonical page. Countries are the distinct countries of datacenters with a known country. Providers tracked are active providers; providers with live pricing are those with at least one live listing.
  • Launch year is the year the part became available. Announcement and shipping dates are recorded separately where known, with a lifecycle status (announced, available, end of sale, end of life) that is left blank rather than guessed.

Performance figures

  • Every FLOPS figure on a spec page is peak theoretical performance: the vendor datasheet number (cores x clock x operations per cycle). Nothing was run. The word benchmark is reserved for a measured workload (MLPerf, LINPACK, tokens per second) and is labelled as measured wherever it appears.
  • Dense and sparse figures are stored in separate columns and never mixed in one ranking. Rankings, headline figures and per-watt numbers use dense. A sparse figure is shown only where the vendor publishes one; on hardware that supports 2:4 structured sparsity it is typically double the dense figure, and a ratio that is not 2:1 is flagged for review rather than corrected.
  • Scalar/vector FP32 and FP64 are kept apart from Tensor Core TF32, FP16 and FP8. Floating point is reported in FLOPS; integer formats (INT8, INT4) in OPS. Display units scale from GFLOPS through TFLOPS, PFLOPS and EFLOPS with the unit always printed.
  • Where a vendor documents that a part cannot run a datatype, the cell says unsupported. Where the vendor is silent, the cell is empty with the reason stated. A blank is never converted to zero.

Power

  • TDP (or TBP/TGP as the vendor names it) is the vendor's thermal design power for the accelerator board or module. Max power is the vendor's published maximum board power. When no maximum is published the page shows a Flopper estimate of TDP x 1.15, labelled as such, rendered in muted text and never in the same style as an official figure.
  • Efficiency is always labelled with its precision, basis and denominator, for example "FP32 dense TFLOPS per W of TDP". It is board power only and excludes host CPUs, networking, cooling and facility overhead (PUE).
  • Accelerator power, server input power, rack IT load and facility power are separate quantities and are never summed across those levels.

System and rack aggregates

  • A system FLOPS figure is either the vendor's published aggregate or, where none exists, the per-accelerator peak multiplied by the accelerator count. The latter is an aggregate theoretical peak, is marked with a dagger, and does not imply linear scaling. Where a vendor explicitly publishes no total for a precision, the cell stays empty rather than being derived.
  • Every stored aggregate carries a basis (listed under provenance) and, in its notes, the inputs and the scale it describes (for example 1,024 cards versus 8,192).
  • Aggregate memory across many accelerators is labelled total installed accelerator memory. System power below the sum of its accelerators' TDP is flagged unless the system is documented as power limited. Accelerator counts refer to accelerator packages unless the page says chips, modules, trays or nodes.

Interconnect and memory bandwidth

  • Memory bandwidth (HBM or GDDR to the accelerator) and interconnect bandwidth (accelerator to accelerator or to the switch fabric) are separate columns and separate rows on every page.
  • Interconnect figures state whether they are bidirectional or unidirectional and whether they are per accelerator or per link. Where the vendor does not say, the page prints "direction and scope not stated by vendor" rather than assuming a convention. NVIDIA NVLink figures are recorded as total bidirectional per GPU, which is how NVIDIA publishes them.
  • Unified-memory designs are described as accelerator memory, because they have no dedicated graphics memory of their own.

Pricing

  • Pricing classes are on-demand, spot, reserved and community, stored as a constrained vocabulary. Only on-demand listings feed a lowest price, a median or a headline count; the other classes are shown separately and labelled.
  • Live price. A live price is an active on-demand listing above $0 from a cross-listable provider, not flagged as an outlier, and confirmed by our ingest within the last 30 days (or published directly by the provider as a curated rate card).
  • Every price is per accelerator per hour; the configuration behind it (accelerator count, vCPUs, memory, region) is shown on the listing. Monthly cost is the hourly price x 730 and is not a quoted monthly contract. Tax, storage, networking, CPU and memory add-ons and availability may be excluded by the provider.
  • Median. Median of each provider's cheapest on-demand listing, so a provider with many configurations counts once
  • Listings not confirmed within 30 days are stale and drop out of live figures; curated partner rate cards with no sync timestamp are the exception and say so. Each listing stores when it was observed, when it was last checked and the source URL. Historical series carry a note where early points were reconstructed.
  • "From $X per GPU-hour" on a provider card is that provider's cheapest live on-demand listing for any accelerator; on a GPU page it is the cheapest for that accelerator.

Provenance and confidence

  • Every throughput figure and every system aggregate carries a basis:
Vendor published vendor_published
The figure appears in the vendor datasheet, product page or keynote material.
Vendor figure, basis unstated vendor_unstated
The vendor publishes the number but does not say whether it is dense or sparse, or at what scale.
Flopper derived (count x per-accelerator peak) derived_count_x_gpu
Aggregate theoretical peak: the per-accelerator peak multiplied by the accelerator count.
Flopper derived (clock x cores) derived_clock_x_cores
Computed from the published clock, core count and operations per cycle.
Flopper estimate flopper_estimate
An estimate with a stated formula, shown only where the vendor publishes nothing.
Third party third_party
Reported by a source other than the vendor; the source is cited on the page.
  • A part carries a spec-confidence badge (official, vendor claimed, undisclosed) for the record as a whole, plus per-field qualifiers where they matter: clock basis for peak figures, datatype support, sparsity support, price basis. A record is marked official only when its headline figures come from vendor documentation.
  • Parts and systems can cite several sources, each with a role (primary, corroborating, errata, pricing). Source publication dates and Flopper verification dates are stored separately.
  • Calculated fields are displayed as calculated. Values that disagree across sources go to an internal review queue and are resolved on this page's corrections list.

Validation rules

These rules run against the catalog before every deploy and file findings into a review queue. An error blocks a deploy; a warning is reviewed. Per-accelerator power is expected between 100 W and 3500 W.

  • system_power_below_gpu_tdp_sum error
    A system may not be rated for less power than its accelerators alone draw at their vendor TDP, unless the vendor states it is power limited.
  • system_watts_per_accelerator_out_of_range error
    A system's rated power divided by its accelerator count must land between 100 W and 3500 W per accelerator; wafer-scale parts are exempt because one wafer is the whole machine.
  • dense_exceeds_sparse error
    A dense figure can never exceed the sparse figure on the same row; when it does the two columns were swapped or a sparse number was halved into the dense column.
  • sparse_not_double_dense_tensor warn
    Structured 2:4 sparsity doubles dense tensor throughput, so a sparse-to-dense ratio other than 2.0 on a tensor precision means one of the two figures needs checking against the datasheet.
  • derived_value_conflicts_with_refusal error
    A deliberately empty system figure must be explained, and neither its note nor a GPU status note may call a per-chip figure unsourced while the catalog records that same figure as vendor published.
  • gpu_count_mismatch_notes warn
    A card, GPU, accelerator or chip count written in a system note must equal the stored accelerator count, or eight times it where the note counts chips per card.
  • datacenter_totals_all_null warn
    A datacenter with no power capacity, no calculated capacity, no H100 equivalents and no chip count reports nothing, and is counted so the site can say how many sites do report.
  • listing_stale_but_active warn
    An on-demand listing still marked active more than 30 days after its last sync is being shown as a current price that nobody has confirmed.
  • pricing_type_unknown error
    Every listing must carry one of the pricing types on-demand, spot, reserved, community; anything else cannot be placed in a price comparison.
  • price_basis_unknown warn
    A listing whose price basis is unknown cannot be normalised to a per-GPU hourly rate, so it is held out of every comparison until the basis is confirmed.
  • max_power_below_tdp error
    The maximum configurable power of an accelerator cannot be lower than its rated TDP.
  • suspicious_peak_dense warn
    A vendor-published dense FP16 or BF16 peak above 4000 TFLOPS on one accelerator row is usually a sparse figure or a multi-chip package figure, and is checked before it is trusted.
  • country_name_variant warn
    Datacenter countries are stored under one short common name each, so an ISO long form or a "(the)" suffix splits one country across two filter values.

Corrections

Every published correction is listed at /corrections with what the page said, what it says now and why. To report an error, use the contact form with the page URL and the source you are comparing against.