Methodology

The Comparator is a value calculator for PC hardware. It does not try to declare one product “the best” for everyone. Instead, it estimates how much value a product offers at the current observed price, under the selected category, preset, and user weights.

What Value Score Means

Value Score is an estimate based on our current dataset, selected weights, market prices, and scoring methodology. It is not an official product rating, a guarantee of quality, a forecast of future prices, or professional purchasing advice.

A score can change when:

• the product price changes;
• competing products become cheaper or more expensive;
• a product is added to or removed from the active dataset;
• benchmark indexes, presets, or formula versions are updated;
• you change weights, condition filters, region assumptions, or the TCO toggle.

This means a higher Value Score does not always mean the product itself got cheaper. It can also mean the market around it changed.

Formula Summary

The current engine uses a relative weighted benefit/cost product:

1. The comparison frame is built from the alternatives currently on screen: active filters, market segment, condition and grouping all apply before the boundaries are measured.
2. Every active factor becomes a unitless ratio to the best observed value. For benefits such as performance or capacity, r = value / frameMax. For costs such as price, TDP or latency, r = frameMin / value.
3. The slider controls the ratio’s strength: factor = r^(weight/100). Thus 0 ignores the factor, 50 applies its square root and 100 applies the full ratio.
4. The raw result is RawValue = product(factor). Benefits multiply the numerator; cost criteria act as inverse multipliers. With only performance=100 and price=100, the formula is exactly proportional to performance / price.
5. The frame leader is 100: Value = 100 × RawValue / max(RawValue in frame). Other products show their share of that leader. The worst product is not forced to 0, because doing so would exaggerate small or low-weight differences. The interface rounds the final score to a whole number for scanability; comparisons and sorting retain the unrounded internal result until the shared display boundary.

Equal and missing parameters: a factor with the same known value for the whole frame has ratio 1 and is neutral; it cannot add a block of identical points or compress the result. A factor with no frame evidence is excluded. If one product lacks a fact known for its alternatives, the engine conservatively uses the worst observed ratio for that factor rather than inventing an average or rewarding the missing value. True zero capability uses a small documented floor so a low non-zero slider does not annihilate the whole product. That floor applies only to a genuinely absent or zero capability: a measured positive value is always used as observed, however far behind the frame leader it sits. Clamping real measurements would make different products score identically — for example every SATA drive in the same read-speed band, or every CPU with a small L3 cache — and a slider that cannot separate them is a control that does nothing.

Lower-is-better criteria are never described as benefits and are never subtracted in their original units. They are explicit inverse ratios in the denominator side of the product. Freshness remains an explicit positive factor: at full weight, decade-old or undated hardware can fall close to 5% of the current-product freshness ratio instead of receiving a protected 25% floor.

This is a practical adaptation of the Weighted Product Model (WPM) used in multi-criteria decision analysis, not a claim that a formula can discover a universally correct purchase. The WPM compares dimensionless ratios under weights, while benefit/cost aggregation is a recognized MCDA problem. See Triantaphyllou and Mann’s comparison of weighted decision methods in Decision Support Systems (DOI 10.1016/0167-9236(89)90037-7) and Triantaphyllou and Baig’s study of benefit-versus-cost aggregation in IEEE Transactions on Engineering Management (DOI 10.1109/TEM.2005.845221). Our 0–100 sliders are exposed elasticities rather than a fitted consumer-preference model, so presets remain documented assumptions and can be overridden.

About warranty and condition: the scoring factor is deliberately categorical, not a guessed month count: full/new warranty ratio = 1.00; short or warranted Refurbished/Renewed/Open Box = 0.75; Used/no warranty = 0.50; genuinely unknown receives the worst observed warranty ratio in the frame. Exact duration above the full/partial threshold may still be shown, but estimated brand defaults do not earn extra Value.

Key properties:

• Weight = 0% means the parameter is completely ignored.
• If all weights = 0%, the fallback is the same as setting every configured slider to 100.
• Factor boundaries and the 100-point leader are relative to the current filtered view.
• The price input is a dated snapshot, not a live checkout price.
• Affiliate status is not a scoring factor and does not affect the formula.

Scenario Tiers Are Not Value Scores

The S–F scenario tier answers a different question: how capable is this product for a named job if price is ignored? A Value Score answers: how much of the weighted capability do I get for the current price? Keeping those answers separate prevents an inexpensive but unsuitable product from being presented as the strongest choice.

• Scenario tiers use a fixed category rubric and never include price, retailer, affiliate status, or stock count.
• N/A means the verified fields required by that scenario are missing. It is not an F grade and no substitute value is guessed for the tier.
• The rank printed beside a tier is relative to the products that had enough verified data to be graded for that scenario.
• Tier lists, product pages, versus pages, and best-for guides use the same shared scenario definitions.
• A preset never silently applies a tier floor. The calculator’s Minimum Tier control remains an independent user filter; generated pages and widgets do not receive an invisible equivalent.
• Best-for pages rank the complete declared page scope with the same preset vector used elsewhere. Any hard scope such as DDR5, NVMe, capacity or a published budget must be stated by the page itself rather than hidden inside the preset.

Presets and User Weights

Presets are saved groups of weights for common use cases. A gaming preset, a productivity preset, and an efficiency preset can rank the same products differently because they reward different strengths.

You can also adjust weights manually. Higher weights apply more of that factor’s relative ratio. Equal weights give equal multiplicative elasticity regardless of whether the source fields are dollars, watts, gigabytes or benchmark points. Setting a weight to zero removes the factor. Price remains a primary factor only in cost-sensitive presets such as Budget, Student, 1080p Gaming and Balanced. In demanding 4K, ray tracing, AI, professional, workstation, server and longevity presets, the primary workload factors have the larger elasticities and price is secondary. Preset QA runs every vector against the complete grouped US market and rejects a scenario whose unfiltered top three contains an F/N/A fit; it also rejects non-price-led presets where price is at least as strong as the main workload factor. There are no brand/model demotions or hidden tier shortlists.

Value Score does not measure brand reputation, long-term reliability, driver quality, merchant service quality, compatibility with your exact PC case or motherboard, or your personal tolerance for used-market risk unless those factors are explicitly represented in the current dataset.

GPU Scoring

For graphics cards, the main good factors are:

• raster performance index (score);
• ray-tracing performance index (rt_score);
• AI/compute index (ai_score);
• VRAM capacity;
• warranty/condition status;
• release-year freshness.

The main bad factors are:

• total price, including shipping when available;
• power draw (tdp);
• optional three-year electricity cost when Total Cost of Ownership is enabled.

GPU presets change the balance. For example, a 4K preset gives more weight to performance and VRAM, a ray-tracing preset gives more weight to rt_score, and an AI/LLM preset gives more weight to ai_score and VRAM.

CPU Scoring

Processors use their own category-specific fields. CPU scoring does not reuse the GPU raster score as its main input.

For CPUs, the main good factors are:

• multi-thread performance (multi_score);
• single-thread performance (single_score);
• gaming-oriented performance (gaming_score);
• core and thread count;
• L3 cache;
• warranty/condition status;
• release-year freshness.

The main bad factors are:

• total price, including shipping when available;
• power draw (tdp);
• optional three-year electricity cost when Total Cost of Ownership is enabled.

CPU presets are workload-oriented. Gaming, productivity, streaming, home-server, developer, budget, efficiency, and longevity presets each emphasize different fields.

RAM Scoring

RAM Value uses kit capacity, data rate, first-word latency, memory generation, warranty/condition, and total price. Speed and latency remain separate, explicit linear factors: more transfer rate is better, while less first-word latency is better. Presets reduce their combined budget to avoid double-counting correlated performance.

Gaming, productivity, server, budget, and balanced presets change the capacity-versus-performance-versus-price trade-off. Product type, capacity, module count, form factor, and motherboard support remain compatibility facts: a DDR5 kit is not a substitute for DDR4 just because its Value Score is higher.

RAM capacity is intentionally not capped or manually penalized. The active generation filter is part of the comparison frame. A DDR5 view therefore recalculates factor min/max from DDR5 alternatives only; revealing DDR4 or DDR3 recalculates those factor boundaries again.

SSD Scoring

SSD Value uses capacity, sequential read speed, sequential write speed, write endurance (TBW), warranty, and total price. System-drive, game/file storage, professional, budget, and balanced presets weight those factors differently.

Interface and form factor are compatibility constraints, not editorial bonuses. Random I/O, controller, NAND, DRAM cache, and heatsink fields are shown when verified, but they do not silently enter the current Value formula. Best-for guides also apply an explicit capacity and interface eligibility frame so an 8TB archive drive cannot win an OS-drive page merely because capacity has a positive weight.

Capacity is both a product strength and a requirement the buyer should set explicitly. In the previous SSD preset calibration, a 3:1 capacity-to-price weight ratio said that doubling capacity could justify paying eight times as much; TBW then often rewarded the same larger NAND pool again. The current presets keep price meaningful and account for that correlation. They do not contain model-specific demotions or hidden price ceilings: use the capacity filter for a hard 1TB, 2TB, 4TB, or 8TB requirement, then compare the eligible drives.

An SSD with one missing actively weighted specification receives no points for that factor; it does not lose every other verified contribution. The calculator does not substitute plausible-looking read, write, endurance, or warranty values for unknown data.

Motherboard Scoring

A motherboard is not faster than another motherboard. A graphics card has a measured performance number; a board does not. It either suits the machine you are building or it does not — and it suits different machines differently. So motherboards are scored by a different engine from the rest of the site, and the single most important thing to know about it is this:

There is no such thing as “the best motherboard”. There is only fit for a stated purpose.

What is measured, in physical units

The engine first records what a board can do in real units — gigabytes, gigabits, port counts, DIMM sockets, form factor — and never in points. Points only appear later, when there is something to compare against: a stated requirement. Mixing “how much there is” with “how good that is” is what used to make a board with many connectors look like a gaming board.

Enough is a property of your task, not of our engine

Each scenario states two numbers per axis: the minimum below which the capability is not useful, and the level at which it is sufficient. Past the sufficient level the Fit score stops rising.

That is deliberate: four M.2 sockets are not twice as good as two when you need two. But — and this is the part that matters — the sufficient level belongs to the scenario, not to the engine. A workstation scenario sets a higher memory target than a gaming one, and the same board legitimately scores differently under each. An engine that carried its own fixed idea of “enough” would be making the decision for you.

“We don’t know” is not “it doesn’t have it”

Every fact carries a state, and these are not interchangeable:

• known — the value is recorded;
• absent — the board genuinely does not have it, and it scores as zero, which is fair;
• not collected — we have not gathered it. This axis is dropped from the score entirely and lowers the confidence figure instead;
• not measured — the quantity needs an independent bench test, not a spec sheet (sustained VRM behaviour is the main one);
• conflicting — trusted sources disagree, so the value is withheld rather than guessed.

If we know less than half of what a scenario asks about, no Fit score is published at all. A number computed from a third of the evidence looks like a rating without being one.

Hard requirements are a gate, not a low grade

If a board cannot physically host the build — no full-speed graphics slot, or an E-ATX board for a small-case scenario — it is marked not suitable for that scenario, not given a poor score. And the reverse also holds: a missing data point never triggers that gate. Not knowing how many slots a board has does not mean it has none.

Value: what you pay for what you actually use

Fit is deliberately capped, so a flagship and a mid-range board frequently reach the same Fit for the same purpose. What separates them is price. Value compares each board only against boards that deliver at least the same Fit for the same scenario, and reports:

• paid overkill — how much more you pay than the cheapest board achieving the same result;
• Pareto-efficient — true when no cheaper board reaches the same Fit. This one depends on no tuning constant at all: it is either true or it is not.

Capability you get for free is never penalised. A board whose memory ceiling exceeds what your scenario needs is not worse for it — that ceiling usually belongs to the socket and costs nothing.

What we do not yet measure, stated plainly

Three things a serious buyer cares about are not in our data, and we would rather say so than approximate them:

• sustained VRM behaviour under load — requires independent bench testing; phase counts and advertised amperage are marketing figures and are not used as a substitute;
• memory tuning behaviour — headline “DDR5-9000” figures depend on how many modules and ranks you populate, which we do not yet track;
• resource topology — on many boards, populating one M.2 socket drops the graphics slot from x16 to x8, and some CPU generations disable sockets entirely. A flat “4 M.2 sockets” does not express this.

Where these are missing, the affected axes report “not known” and lower the confidence figure. They are listed rather than removed, because an axis that quietly disappears is one nobody ever notices is absent.

Socket, chipset, form factor, BIOS support, case clearance and the manufacturer’s CPU support list still decide compatibility. A higher score never means two boards are interchangeable.

Power Fields: What “TDP (Max)” Actually Means

Power values on The Comparator are intentionally the maximum sustained package or board power, not always the manufacturer’s marketing TDP. This is a deliberate choice: for value scoring and electricity-cost estimates, the number that matters is what the part can actually draw under load.

• CPUs. The TDP (Max) column is the maximum package power (for modern parts this corresponds to PPT / PL2-class limits — e.g. a CPU marketed as “170W TDP” may show 230W here). The manufacturer’s base TDP is stored separately (tdp_base) and is not used in scoring.
• GPUs. The power value is the board power for the card (TDP/TBP as published in the specification database), i.e. the whole card under load, not just the chip.
• SSDs and RAM. Power draw is negligible for value purposes and is not part of their formulas.

Both the Value formula (power is inverted to a higher-is-better efficiency score when you give it weight) and the optional Total Cost of Ownership electricity estimate use this maximum figure. That makes TCO slightly conservative — it assumes load-level draw during your configured usage hours — which we prefer over understating running costs with an idle-flattering number.

If a power value looks wrong for a specific model, it is more likely a data issue than a methodology one — check the model page and tell us.

Prices and Offers

Prices on The Comparator are offer snapshots with visible freshness information. They can become outdated quickly, and the final price, shipping, stock status, tax, return policy, and seller terms are always the retailer’s responsibility.

For Amazon offers, we may link users to Amazon instead of showing a live price on The Comparator. The value calculation may still use the last recorded price snapshot so the ranking remains useful, but users should confirm the current Amazon price on Amazon before making a decision.

Used and refurbished offers are treated as separate conditions when available. They can be good value, but they also carry extra uncertainty around warranty, seller quality, condition, returns, and remaining lifetime.

Target Price

Some widgets and future product pages may show a target or fair-value price. This is the estimated price at which the product would reach a selected Value Score threshold under the current dataset, formula, and preset.

Target price is not a prediction that the market price will fall to that level. It is a calculation from the same value model, useful for deciding whether to wait, compare alternatives, or check a used/refurbished listing more carefully.

Affiliate Independence

The Comparator may earn commissions from qualifying purchases through affiliate links. Affiliate relationships do not change the formula, weights, benchmark inputs, or ranking order. We do not manually boost a product because a link pays a higher commission.

Sponsored or partner relationships, if any, must be disclosed separately. Product names, brand names, and logos belong to their respective owners and are used for identification only.

Confidence and Freshness

Where confidence is shown, it describes how much trust to place in the current data match and offer snapshot. Confidence can depend on source quality, offer count, product-name matching, condition clarity, and how recently the price was checked.

Data refresh frequency varies by source and category. The site should be read as a dated snapshot system: always use the visible price date, freshness marker, and retailer page as the final check before purchase.

Data Freshness and Corrections

The dataset combines product specifications, benchmark indexes, retailer offer snapshots, and category-specific normalization logic. We aim to keep the data useful and explainable, but mistakes and stale offers can happen.

If you see an incorrect price, specification, link, condition, or product match, please report it at contact@thecomparator.tech. Factual corrections can change the resulting score because the formula is recalculated from the updated data.

Publication Scope

Product and versus pages are generated only for an explicit current-generation wave with per-category and per-generation caps. Existing published URLs remain stable, while new pages prioritize products with current offers and complete specifications. Automatic RAM comparisons stay within the same memory generation and kit capacity; SSD comparisons stay within the same capacity and form factor; motherboard comparisons stay within the same CPU socket. This prevents a routine data refresh from creating an uncontrolled Cartesian product of pages.

Current Formula Versions

• GPU value formula: gpu-value-v1.4-benefit-cost-product
• CPU value formula: cpu-value-v1.4-benefit-cost-product
• RAM value formula: ram-value-v1.4-benefit-cost-product
• SSD value formula: ssd-value-v1.5-benefit-cost-product
• Motherboard value formula: mobo-value-v1.4-benefit-cost-product
• Scenario tier rubric: scenario-tiers-v1

Formula versions are listed here so future changes can be documented clearly. Any scoring change that materially affects rankings should receive a new formula version and a public explanation.