A GPU spec sheet looks like it should settle arguments, and it settles fewer than most people assume. The numbers are real, but several of the ones buyers lean on hardest, CUDA core counts especially, mean different things depending on context, and reading them literally leads straight to bad comparisons and bad purchases. This is a field guide to the specs that matter, what each one actually tells you, and where each one lies by omission.
CUDA cores: useful within a generation, meaningless across them
CUDA cores are the scalar arithmetic lanes inside each Streaming Multiprocessor. The advertised count is just the number of SMs multiplied by the lanes per SM, and within a single architecture it scales roughly with performance: a card with more cores of the same generation is generally faster.
Across generations the number falls apart, because the work one core does per clock changes every architecture. Clocks, cache behavior, and datapath design all move, so a core in one generation is not the same unit of work as a core in the next. The canonical example: the RTX 4070 (Ada) has 5,888 CUDA cores, while the RTX 3080 (Ampere) has 8,704. Despite roughly a third fewer cores, the 4070 matches or slightly edges the 3080 in rasterization while drawing far less power, 200 W against 320 W, because Ada's higher clocks, larger cache, and per-SM improvements raise the effective throughput of each core. The exact result varies by title and resolution, and the 3080's wider bus helps it at 4K, so the honest claim is "comparable raster, much better efficiency," not a blanket win. The lesson holds either way: comparing raw core counts between generations tells you almost nothing.
The failure this prevents is the most common spec mistake there is, dismissing a newer card because an older one has a bigger core number. That number is only comparable inside the same architecture.
Memory bus width and bandwidth
Bandwidth is where a lot of intuition goes wrong, usually because people read bus width in isolation. The relationship is simple:
bandwidth = bus width (bits) x memory data rate (per pin) / 8
Bus width is set by how many memory chips are wired up and how wide each interface is; the data rate is the effective per-pin transfer figure, the "Gbps" number on the spec sheet. Worked through for the RTX 4070, a 192-bit bus with GDDR6X at 21 Gbps gives 192 x 21,000,000,000 / 8 = 504 GB/s, which matches NVIDIA's published figure. Neither number alone tells you the bandwidth; you need both.
A narrow bus also is not the red flag it once was. Ada Lovelace enlarged the on-die L2 cache enormously, up to 96 MB on the AD102 die against roughly 6 MB on the previous generation's GA102. A large L2 serves more memory requests on-chip and cuts traffic to DRAM, which let NVIDIA pair 128-bit buses with the 4060-class cards without the bandwidth collapse that width would once have implied. The catch appears at 4K, where cache hit rates fall and the narrow bus starts to limit, which is why the 128-bit Ada cards are best read as 1080p-to-1440p parts. Bus width matters, but only in combination with data rate and cache, never on its own.
TDP, TGP, and total board power
These three terms get used interchangeably and they are not the same thing:
- TDP (Thermal Design Power) is, strictly, a thermal figure: the heat the cooler must dissipate. It is often quoted as if it were power draw, but that is loose usage.
- TGP (Total Graphics Power) or TBP (Total Board Power) is the power budget for the whole board, GPU plus memory plus VRM losses plus fans. This is the number that sizes your power supply. NVIDIA's desktop spec sheets list it; the RTX 4090, for instance, is 450 W, with board partners able to set a higher configurable ceiling.
For desktops, take the spec-sheet board-power figure as authoritative and ignore the terminology arguments. For laptops the distinction turns into a real buying trap. NVIDIA defines a configurable TGP range for each mobile GPU and lets the manufacturer pick a point within it. An RTX 4070 Laptop GPU is specified anywhere from 35 W to 115 W, and boost clocks scale accordingly, so two laptops with the identical GPU name can differ by tens of percent in performance. The marketing name says nothing about the power tier; the shipped TGP does, and you have to look it up per machine. The RTX 40 series guide works through how far that gap goes across the mobile lineup.
Die names: the number that cuts through the marketing
Every NVIDIA GPU has a die code, and it tells you more than the product name does. The scheme is a two-letter architecture prefix plus a three-digit tier number: GP for Pascal, TU for Turing, GA for Ampere, AD for Ada Lovelace, GB for Blackwell. The tier number runs backwards from intuition, a lower number is a bigger, higher-tier die. So AD102 outranks AD103 outranks AD104, and GB202 sits above GB203 and below nothing.
Concretely, AD102 is the Ada flagship die (around 608 mm², powering the RTX 4090), and GB202 is the Blackwell flagship (around 750 mm², powering the RTX 5090), roughly 24% larger. GP102 was the Pascal flagship behind the GTX 1080 Ti and Titan Xp.
The die code matters because marketing names get reshuffled between generations while the die tells you the actual silicon tier, transistor budget, and bus-width ceiling. A "70-class" card can shift dies from one generation to the next, and a laptop "4090" is built on AD103, not the desktop 4090's AD102. When two products with similar names are built on different dies, they are different GPUs with different ceilings, and the die code is what exposes that.
Compute capability: the spec that decides whether your software runs
For anyone doing CUDA or AI work, compute capability is the spec that actually gates whether code runs, and it is invisible on a gaming spec sheet. It is a version number, written like 8.9, that identifies the hardware feature set a GPU exposes to CUDA: supported instructions, warp and shared-memory limits, tensor-core generation. It is tied to the architecture, not the core count or the SKU, so every RTX 40 card from the 4060 to the 4090 is compute capability 8.9.
Current values, verified against NVIDIA's CUDA GPU list:
| Architecture | Compute capability |
|---|---|
| Ampere consumer (RTX 30) | 8.6 |
| Ampere data center (A100) | 8.0 |
| Ada Lovelace (RTX 40) | 8.9 |
| Hopper (H100/H200) | 9.0 |
| Blackwell data center (B200, GB200) | 10.0 |
| Blackwell consumer (RTX 50) | 12.0 |
The Blackwell split is real and worth noting: the data-center and GeForce Blackwell parts carry different compute capabilities, 10.0 and 12.0, because the same architecture family exposes different feature sets in different products. This matters because CUDA toolkits and AI frameworks such as PyTorch and TensorRT gate prebuilt kernels on compute capability. If a framework build does not include kernels for your card's capability, you get a "no kernel image is available" error or a slow fallback, not a graceful degrade. The failure this prevents is subtle and expensive: buying or deploying a card whose compute capability your software stack does not yet support, then losing days to build errors that a spec sheet never warned you about.
Same name, different silicon across segments
The final trap is that one marketing name maps to different physical GPUs depending on the segment it ships in:
- Desktop: full board power, GDDR6 or GDDR6X, display outputs, active axial-fan cooling. The name maps to a specific die (desktop RTX 4090 to AD102).
- Laptop: frequently a smaller die at a configurable low TGP. The laptop RTX 4090 is AD103, materially less silicon and power than the desktop card of the same name.
- Workstation (RTX PRO, formerly Quadro): the same architecture but usually more VRAM with ECC, certified professional drivers, and blower or passive cooling for multi-card chassis. Marketed under RTX PRO names rather than GeForce.
- Data center (A100, H100, B200 and similar): built for compute, not display. No display outputs, HBM memory instead of GDDR, ECC throughout, passive cooling driven by chassis airflow, high-bandwidth NVLink, and features like MIG partitioning. A different die entirely, with a different compute capability from the same-generation GeForce part.
The takeaway is that the name is a market label. The die code, the memory type (GDDR versus HBM), the presence of ECC, the power spec, and whether the card even has display outputs are what tell you what the chip actually is. The recent RTX PRO 6000 pricing story is a good illustration of how far the professional silicon diverges from the GeForce card that shares its architecture.
A source checklist
Specs are only as good as where they came from. When you need to trust a number:
- Start with NVIDIA's own product page and spec sheet for launch figures, MSRP, and official memory and power specs.
- Cross-check against a maintained database such as TechPowerUp for die-level detail (transistor count, die size, SM and ROP counts) that the marketing page omits.
- For CUDA and AI work, confirm compute capability against NVIDIA's developer CUDA GPU list, not a forum post.
- Never infer a missing spec from a similar card or the same architecture. A cut-down die, an OEM variant, or a laptop part can differ in exactly the field you are guessing at. If a source does not state it, mark it unknown rather than filling the gap.
Bottom line
Read specs for what each one actually measures. CUDA cores compare only within a generation, bandwidth needs bus width and data rate together, board power is the number that sizes a PSU and varies wildly on laptops, the die code cuts through renamed products, and compute capability quietly decides whether your CUDA software runs at all. Get those five right and a spec sheet becomes genuinely useful. Read them literally and it will talk you into the wrong card. For worked examples of the same reasoning applied to real hardware, see the RTX 40 series guide and the GTX 1080 Ti retrospective.
Sources: