How to Read a GPU Spec Sheet Without Getting Fooled: Cores, Clocks, Memory Bandwidth & Power Figures Explained

｜ HK GPU Fix

重點摘要

A graphics card spec sheet lists core counts, clock speeds, memory figures and power ratings, but none of those numbers means much on its own. This guide explains what each figure actually measures, where manufacturers define terms differently, and how to compare cards without being misled by a single headline number.

Why Two Cards With Similar Numbers Can Behave Differently

A reader comparing two graphics cards will often find that the specification sheets look reassuringly similar. The core counts are in the same region. The clock speeds are close. The memory capacities match. The power figures differ by only a small margin. On paper, the cards look interchangeable, and the cheaper one looks like the obvious choice. In use, the two cards may behave quite differently, and the reader is left wondering whether the specification sheet lied.

It did not lie, but it did not tell the whole story either, because a specification sheet is not a performance measurement. It is a list of design parameters, each measured under its own definition, each meaningful only alongside the others, and some defined differently by different manufacturers. A core count from one manufacturer does not count the same kind of unit as a core count from another. A boost clock is a conditional maximum, not a sustained speed. A memory capacity says nothing about how fast that memory can be fed to the processor. A power figure from one manufacturer does not measure the same thing as a power figure from the other. None of these facts makes the specification sheet useless. Together, they make it a document that must be read as a system, not as a league table.

This guide teaches that reading. It takes each major section of a typical specification sheet — processing cores, clock speeds, memory, and power — explains what the figure actually measures, where the traps lie, and what to check alongside it. It then covers the smaller sections that are easy to skip but expensive to get wrong: outputs, power connectors, physical size, and manufacturing process. It closes with a short pre-purchase checklist that turns the whole sheet into a sequence of checks, and points to the companion articles that cover what a specification sheet cannot: real-time monitoring after purchase, in article #108, and benchmark interpretation, in article #203.

One boundary matters throughout. This guide is about reading the sheet before purchase. It does not rank cards, it does not recommend a card for a particular game, and it does not repeat the detailed treatments that belong elsewhere: laptop graphics power classes belong in article #206, power-supply budgeting belongs in article #169, ray-tracing cost belongs in article #212, and frame-generation cost belongs in article #175. Where those subjects touch this guide, it points to them and moves on.

Processing Cores: A Count That Only Compares Within Its Own Family

The first figure most readers look at is the core count, and it is also the first figure that misleads across brands. NVIDIA describes its processing units as CUDA cores. AMD describes its units as stream processors. The two terms sound equivalent, and the numbers look comparable, but they count different things. The internal organisation of a processing unit, how much work it completes per clock cycle, and how it shares resources with its neighbours all differ between architectures. A card with a larger count of one manufacturer's units is not automatically faster than a card with a smaller count of the other's units, and ranking cards by core count across brands is not a rough approximation. It is a category error.

Within a single brand and a single generation, the core count has a more modest and genuine use. Cards that share the same architecture, differing mainly in how many processing units are enabled, can be roughly ordered by core count, because the units being counted are the same kind of unit doing the same kind of work. Even there, the ordering is rough rather than exact: memory configuration, clock behaviour, and power limits also differ between models, and any of them can outweigh a core-count advantage in a particular workload.

Across generations, even within one brand, the comparison weakens again. A new architecture may redesign the processing unit itself, so that each unit completes more work per clock cycle than its predecessor, or handles a particular kind of operation more efficiently. A newer card with fewer cores than an older card can therefore outperform the older card, and a reader who insists on comparing core counts across generations will conclude the opposite. The honest use of a core count is narrow: same brand, same generation, same architecture family, as one input among several. Outside that boundary, the question "which card is faster" must be answered by measured performance, as explained in article #203, not by counting units.

A related trap concerns model names. Manufacturers sometimes sell cards under the same family name with different core counts in different markets or form factors. The most important case is the laptop version of a desktop card name, where the laptop variant may carry a substantially different specification under a familiar name. The rule is simple and non-negotiable: compare the full model name and the specification page for that exact model, not the family name. Laptop graphics power classes and their consequences are covered in full in article #206, and readers comparing a laptop card should read that article before trusting any desktop specification sheet.

Clock Speeds: Three Different Numbers Wearing One Name

Clock speed is quoted as a frequency, typically in megahertz or gigahertz, and readers naturally treat a higher number as a faster card. The difficulty is that a specification sheet quotes more than one clock figure, and the figure that matters in sustained use is not necessarily the one printed largest.

The base clock is the manufacturer's baseline: the clock speed the card is designed to sustain under its rated power and thermal conditions while running a typical demanding workload. It is a reference point, not a promise that the card will sit at that speed at all times, and not a floor below which the card never drops. Light workloads may run below it, and the card's power management may idle far below it when there is nothing to do.

The boost clock is the conditional maximum: the highest speed the card may reach when power headroom and thermal headroom allow. The word "may" is doing real work in that sentence. Boost is available when the card is cool enough, when the power limit has not been reached, and when the workload permits it. A card in a poorly ventilated case, a card whose power limit is reached quickly, or a card running a sustained heavy workload will spend much of its time below its boost figure. Boost is a ceiling under favourable conditions, not a cruising speed.

The actual clock speed, the speed the card runs at moment by moment during a real workload, is determined by the interaction of three limits: the power limit, the temperature limit, and the voltage-frequency behaviour of the silicon. As a workload continues, heat builds, the power limit is approached, and the card settles at whatever speed those limits jointly permit. That sustained speed may be closer to the base figure than to the boost figure, and it is the speed that actually determines sustained performance. The mechanism of power and temperature limits, and the related subject of undervolting, is covered in article #134, and fan behaviour in article #146; this guide confines itself to the reading lesson, which is that the boost figure must never be treated as the running speed.

The practical misunderstanding follows directly: a card with a higher boost clock is not automatically faster than a card with a lower boost clock. If the higher-boost card has a lower power limit or poorer cooling, its sustained clock may fall below the sustained clock of the lower-boost card with more power headroom and better cooling. The specification sheet alone cannot settle that comparison, because sustained clock depends on the card's cooling design and power design as built, not only on the silicon's rated ceiling. That is why the boost figure must be read as "how high this card may briefly go" and never as "how fast this card runs".

Memory: Three Numbers That Must Be Read Together

Memory is where single-figure shopping does the most damage, because the figure most readers look at first — capacity in gigabytes — is the least informative about speed, and the figures that determine speed are the ones most often ignored.

Capacity answers one question: how much data can the card hold in its own memory at once? A larger capacity allows larger textures, larger working sets, and more data resident without spilling to system memory. Capacity matters most at the boundary where it is insufficient; below that boundary, a shortfall produces stutter and late-arriving detail, as described in article #156. Above that boundary, extra capacity does not make the card faster. A workload that fits comfortably in a smaller capacity gains nothing from a larger one, because the extra memory simply sits unused. Capacity is a threshold, not a speedometer.

Bandwidth answers a different question: how fast can data move between memory and the graphics processor? It is measured in gigabytes per second, and it matters most when large amounts of data must move continuously — high resolutions, high-quality textures, and complex scenes all demand sustained bandwidth. A card with generous capacity but narrow bandwidth can hold a great deal of data and still starve the processor while waiting for it. A card with modest capacity but wide bandwidth can feed its processor briskly, provided the workload fits.

Bus width, measured in bits, is the third figure, and it is the one most often mistaken for a ranking. Bus width describes how many bits move in each transfer cycle. A wider bus moves more bits per cycle; a narrower bus moves fewer. But bandwidth is the product of bus width and transfer rate, so a narrow bus running at a high transfer rate can deliver the same bandwidth as a wide bus running at a lower rate. Judging a card by bus width alone, without the transfer rate, is like judging a road by its number of lanes without asking the speed limit. Common widths such as 128-bit, 192-bit, 256-bit, and 384-bit appear across many cards; those figures describe the design's shape, not a quality ranking, and this guide deliberately does not rank them.

The relationship between the figures is exact enough to compute, and learning the computation is the best defence against being misled by any one of them. Theoretical memory bandwidth is calculated as:

Bandwidth = per-pin transfer rate × bus width ÷ 8

The division by eight converts bits to bytes. Worked through with figures chosen only to demonstrate the arithmetic: a bus 384 bits wide, with a per-pin transfer rate of 21 gigabits per second, gives 384 × 21 ÷ 8 = 1008 gigabytes per second. That is the theoretical peak bandwidth of such a configuration. The word "theoretical" must be kept attached to that result, because sustained bandwidth in real workloads falls below the peak. Access patterns, cache behaviour, and memory compression all affect how close a workload comes to the peak, and no workload sustains the peak continuously. The formula is a tool for comparing the design's ceiling and for checking whether a quoted bandwidth figure is consistent with the quoted bus width and transfer rate. It is not a prediction of delivered performance, and a card with a higher theoretical bandwidth is not automatically faster in every workload, because processing power, cache design, and power limits also govern the result.

Memory type enters the formula through the per-pin transfer rate. Successive memory generations — GDDR6, GDDR6X, and GDDR7, alongside the very different high-bandwidth memory used on some specialised cards — achieve higher per-pin rates through different signalling methods. GDDR6X, for example, uses a four-level signalling scheme to pack more data into each cycle, while GDDR7 uses a three-level scheme with its own trade-offs, and high-bandwidth memory achieves its bandwidth through an extremely wide interface running at a lower rate rather than through a high per-pin rate. The detail that matters for reading a specification sheet is not the alphabet of memory types but the rate each type enables, because the rate is what enters the bandwidth formula. The thermal behaviour of one particular memory type is covered in article #145, and readers concerned about memory temperatures should consult that article rather than treating memory type as a pure speed label.

Memory at a Glance

The single-figure fallacy deserves a plain example, stated hypothetically. Imagine two cards where one carries more memory capacity than the other but a narrower effective bandwidth, while the other carries less capacity but substantially higher bandwidth. A reader who buys on capacity alone chooses the first card and discovers that demanding scenes, which move large amounts of data continuously, run worse than expected, because the processor waits for data. A reader who buys on bandwidth alone chooses the second card and discovers that a workload with a very large working set spills over, because the capacity is insufficient. Both readers made the same mistake in opposite directions: they treated one figure as the whole memory story. The memory story is always the three figures together, checked against the workload's actual demands.

Power: Three Terms That Must Not Be Confused

Power figures cause a special kind of confusion, because three similar-sounding terms circulate in specification sheets and reviews, and they do not measure the same thing. Mixing them up produces comparisons that look precise and mean nothing.

TDP, thermal design power, traditionally describes the heat that a chip's cooling solution is designed to dissipate. It is a thermal figure, not an electrical consumption figure for the whole card, and different manufacturers and different eras have used the term with different definitions. Treating a TDP figure as "how much electricity this card consumes" is therefore unsafe: the figure may describe the chip's thermal envelope rather than the board's power draw, and its definition may have shifted between product generations.

TBP, typical board power, is the term used by AMD for the typical power consumption of the whole graphics board. It is a board-level figure, and as such it is closer to what a reader wants when asking "how much power does this card draw". But "typical" carries its own caution: it describes a representative workload, not a maximum, and short peaks above the typical figure occur.

TGP, total graphics power, is the term used by NVIDIA for the total power of the whole graphics card. Like TBP, it is a board-level figure rather than a chip-only figure. The two board-level terms sound equivalent, and they serve a similar purpose, but they are defined and measured by different manufacturers under different methods. Comparing a TBP figure directly against a TGP figure as if they were measured identically is not an apples-to-apples comparison, even though both describe whole boards. Within one brand, board-power figures have comparative value. Across brands, they are indicative at best, and a conclusion of the form "this card draws more watts, therefore it is faster" or "therefore it is less efficient" cannot be drawn from the two figures alone.

Power Terms at a Glance

Two further cautions complete the power picture. First, a specification sheet's power figure is only one input to power-supply planning. The full machine's power budget, including the processor, drives, and other components, plus the need for headroom above typical draw for short peaks, determines the power supply requirement. That budgeting is covered in article #169, and this guide deliberately does not compute a power-supply size from a card's power figure alone, because doing so from one number would mislead. Second, laptop cards obey a different power logic entirely: the same laptop card name can carry very different power ratings depending on the laptop model it is fitted to, and desktop power figures must never be applied to a laptop card. That subject belongs to article #206, and it is the single most expensive specification-sheet error a laptop buyer can make.

The Smaller Sections: Easy to Skip, Expensive to Get Wrong

Beyond cores, clocks, memory, and power, a specification sheet carries several smaller sections that readers often skim. Each one can invalidate an otherwise careful choice if it is ignored.

Dedicated ray-tracing and acceleration units deserve a brief note, because their presence is often treated as a performance promise. A card that lists ray-tracing units or tensor-style acceleration units supports those kinds of computation, but the listing says nothing about how fast those computations run on that card. Ray-tracing performance varies enormously between cards that all "support" ray tracing, and its cost in a real game is covered in article #212. Frame generation, which depends on related acceleration hardware, has its own cost profile in article #175, and upscaling in article #162. The specification sheet's tick in a support column is a statement of capability, not of speed.

Output interfaces are the next section to check, and the check is about versions, not counts. A card may offer several display outputs, but the version of each output — which HDMI revision, which DisplayPort revision — sets the maximum bandwidth that output can carry, and therefore the resolutions and refresh rates it can sustain. A card with many outputs is not automatically a card that can drive many high-refresh displays simultaneously; multi-display bandwidth limits are covered in articles #199 and #138. Readers should match the output version to the display they intend to use, using the guidance in articles #110 and #190, rather than counting ports.

Power connectors are a physical compatibility check that no power figure can replace. The specification sheet lists how many power connectors the card needs and what type they are, and the reader's power supply must provide them. A card whose power figure looks comfortable can still be unusable if the required connector type is absent, and connector varieties and their histories are covered in articles #181, #195, and #142. The check is mechanical: list the connectors the card requires, open the case or the power supply's specification, and confirm each one exists before purchase.

Physical dimensions are the most prosaically decisive section. Length, thickness in slots, and weight determine whether the card fits the case, whether it blocks neighbouring slots, and whether it needs additional support against sagging. Measuring the case before purchase is covered in article #178, and card sag in article #123. A card that cannot physically fit is not a bargain at any price, and this check costs nothing but a tape measure.

Finally, the manufacturing process node, quoted in nanometres, should be treated with the greatest scepticism of any figure on the sheet. Different manufacturers measure and label process nodes differently, and the same nominal node size can conceal very different transistor densities and power characteristics. Architecture, power design, and cooling matter far more than the node label, and a smaller node number must never be read as proof of a faster or more efficient card. The node figure is, at best, a hint about the generation of the manufacturing technology. It is not a performance ranking.

A Worked Reading: How the Sections Combine

The value of reading the sheet as a system becomes clear when the sections are combined on a single hypothetical card. Imagine a specification sheet that shows a high core count, a high boost clock, a large memory capacity, and a moderate power figure. Read section by section, each figure flatters the card. Read as a system, questions appear. The high core count is meaningful only if the architecture matches the comparison set. The high boost clock is a ceiling, and the sustained clock will depend on whether the moderate power figure permits it under sustained load. The large capacity is reassuring only if the bandwidth, computed from bus width and transfer rate, can feed the cores. The moderate power figure, if it is a chip-level thermal figure rather than a board-level figure, may understate the board's actual draw. None of these questions accuses the sheet of dishonesty. Each one reflects the fact that a design is a set of trade-offs, and the sheet lists the trade-offs' parameters rather than their outcome.

The reverse hypothetical is equally instructive. A sheet showing a modest core count, a modest boost clock, and a modest capacity may describe a card whose architecture does more work per core, whose sustained clock holds close to its boost figure because its power and cooling design is generous, and whose memory bandwidth amply feeds its cores for the workloads it targets. Such a card can outperform a superficially stronger sheet in sustained use. The sheet did not hide that outcome; it stated the parameters from which a careful reader could suspect it. The difference between the two readers is not access to information. It is the habit of reading the parameters together and asking what limits what.

The Pre-Purchase Checklist

A short checklist turns this guide into a procedure. It is designed to be worked through in order, with the specification sheet open and the manufacturer's specification page for the exact model — not the family page, not a retailer's summary — as the source.

First, confirm the exact model. Write down the full model name, including any suffix that distinguishes variants, and confirm that the specification page being read belongs to that exact model. If the purchase is a laptop, stop here and read article #206 first, because laptop power classes change the meaning of the rest of the sheet.

Second, read the cores within their family only. Note the core count, but use it only to order cards of the same brand, generation, and architecture. For any cross-brand or cross-generation comparison, set the core count aside and turn to measured performance in article #203.

Third, separate the three clocks. Note the base clock as the reference, the boost clock as the conditional ceiling, and remember that the sustained clock will be set by power and thermal limits. Do not rank cards by boost clock alone.

Fourth, work the memory figures together. Note capacity, bus width, and transfer rate, and compute the theoretical bandwidth with the formula above, keeping the word "theoretical" attached. Check that the quoted bandwidth, if one is given, is consistent with the computed figure. Judge capacity against the workload's needs, using article #156 for the consequences of a shortfall, and judge bandwidth as the ceiling on data delivery.

Fifth, identify the power term. Determine whether the quoted figure is a chip-level thermal figure, a typical board figure, or a total graphics power figure, and compare it only within the same manufacturer's usage. Do not use it alone to size a power supply; use article #169 for the full budget.

Sixth, run the physical and compatibility checks. Confirm output versions against the intended display, confirm power connectors against the power supply, and confirm length, thickness, and weight against the case. These checks take minutes and prevent the most expensive kind of mistake: a card that cannot be used at all.

Seventh, verify against measured performance. Use article #203's guidance on reading benchmarks to check whether the card's measured behaviour in workloads like the intended one matches what the sheet suggested. If the measurements contradict the sheet-based expectation, trust the measurements, and revisit which sheet parameter was misread. The sheet is the hypothesis; measurement is the test of it.

What a Specification Sheet Cannot Tell You

A specification sheet, read perfectly, still leaves important questions unanswered, and knowing those limits is part of reading it well. It cannot tell you how loud the card's cooler is under load, because acoustic performance depends on the cooler design and the case, not on any figure the sheet quotes. It cannot tell you how the card behaves in the specific game or application you care about, because that depends on how that software stresses each part of the design. It cannot tell you about driver maturity or software support over time. And it cannot tell you whether a particular card, as manufactured, sustains its rated behaviour, because manufacturing variation and cooler design vary between board partners selling cards based on the same chip.

Those gaps are not reasons to ignore the sheet. They are reasons to treat the sheet as the first stage of selection — the stage that narrows the field and eliminates incompatible or unsuitable candidates — and measured performance and post-purchase monitoring as the later stages that confirm the choice. Article #108 covers the monitoring stage, and article #203 covers the measurement stage. A buyer who uses all three stages — sheet, measurement, monitoring — has the full picture. A buyer who uses only the sheet has a hypothesis. A buyer who uses only a single headline figure has a guess.

Frequently Asked Questions

Can I compare core counts between NVIDIA and AMD cards?

No, not meaningfully. The two manufacturers count different kinds of processing units, organised differently and with different per-clock efficiency. Core counts are useful only for rough ordering within the same brand, generation, and architecture family. Across brands or generations, use measured performance instead.

Is a higher boost clock always faster?

No. Boost clock is a conditional maximum available when power and thermal headroom allow. Sustained clock speed under load is set by power limits, temperature limits, and cooling, and a card with a lower boost figure but more headroom can sustain a higher speed than a card with a higher boost figure that quickly reaches its limits.

How do I calculate memory bandwidth from a spec sheet?

Multiply the per-pin transfer rate in gigabits per second by the bus width in bits, then divide by eight to convert bits to bytes. For example, 384 bits at 21 gigabits per second gives 384 × 21 ÷ 8 = 1008 gigabytes per second. That result is the theoretical peak bandwidth, not the sustained bandwidth delivered in real workloads, which falls below the peak depending on access patterns, cache behaviour, and compression.

Does more VRAM always mean a faster card?

No. Capacity is a threshold: enough capacity avoids spillover, but capacity beyond what the workload needs does not raise speed. Bandwidth, processing power, and clock behaviour matter alongside capacity, and a large-capacity card with narrow bandwidth can be slower in data-hungry scenes than a smaller-capacity card with ample bandwidth.

What is the difference between TDP, TBP, and TGP?

TDP traditionally describes the chip's thermal design envelope and is not a whole-card electricity figure, with definitions varying by manufacturer and era. TBP is AMD's typical board power, the whole board's typical draw. TGP is NVIDIA's total graphics power, the whole card's power figure. Board-level figures are comparable within one brand's usage, but TBP and TGP are defined differently and should not be compared directly as if identically measured.

Can I use a desktop card's power figure for a laptop card of the same name?

No. Laptop variants of a desktop card name can carry very different power ratings depending on the laptop model, and desktop figures do not apply to them. Laptop graphics selection is covered in article #206 and should be consulted before any laptop purchase.

Does a newer memory type guarantee a faster card?

No. Memory type indicates the achievable per-pin transfer rate, which is only one input to bandwidth. A newer type on a narrow bus can deliver less bandwidth than an older type on a wider bus, and bandwidth itself is only one factor in overall performance. Always compute the bandwidth and read it alongside capacity and processing power.

Is a smaller manufacturing process node always better?

No. Process node labels are measured and named differently by different manufacturers, and architecture, power design, and cooling have larger effects on performance and efficiency than the node number alone. Never rank cards by node size.

What should I check on a spec sheet besides the headline numbers?

Output versions against the intended display, power connector types against the power supply, and physical dimensions against the case. Also confirm the exact model name against the manufacturer's specification page for that model, not a family summary. These checks prevent a card that cannot be used at all, which no headline figure can compensate for.

If the spec sheet and a benchmark disagree, which should I trust?

Trust the measurement, and revisit the sheet reading. The sheet lists design parameters; a benchmark measures delivered behaviour in a specific workload. A disagreement usually means a parameter was misread — a boost figure taken as sustained speed, a board figure compared across brands, or a single figure treated as the whole story — and the measurement shows which reading was wrong.

A Final Note

A specification sheet rewards the reader who treats it as a system of interlocking parameters and punishes the reader who treats it as a ranking. Cores count only within their family. Clocks describe a range, not a speed. Memory is capacity, bandwidth, and bus width together, with bandwidth computed and labelled as theoretical. Power terms describe different things and must not be swapped. The smaller sections — outputs, connectors, dimensions — decide whether the card can be used at all. Read that way, the sheet stops being a trap and becomes what it was meant to be: a precise, if incomplete, description of a design, to be checked against measurement before money changes hands.

For help reading a particular card's specification sheet, or for a check of whether a card suits a particular machine and workload, readers are welcome to contact the shop at Room 1209, Dali Plaza, Shek Mun, by phone on 2186 8010, or by WhatsApp on 6558 6806.