RTX 50 Series Model Guide: What Each Card From 5050 to 5090 Is Actually For

Published October 27, 2026 · Independent guide by RTX 50 Series

How the RTX 50 desktop line-up is built

The RTX 50 family is easier to understand as a ladder of five silicon dies, each used for one or two products. Seeing the ladder first makes every spec sheet easier to read.

Five Blackwell dies, seven desktop products

NVIDIA's desktop cards use the GB202, GB203, GB205, GB206 and GB207 dies. GB202 powers the RTX 5090. GB203 powers the RTX 5080 and RTX 5070 Ti. GB205 powers the RTX 5070. GB206 powers the RTX 5060 Ti and RTX 5060. GB207 powers the RTX 5050. The companion article on GB202 to GB207 chips explains how enabled streaming multiprocessors (SMs) turn one die into two products.

SM count and CUDA cores

Each SM in these designs holds 128 CUDA cores, so you can verify any card by dividing its CUDA core total by 128. The RTX 5070 Ti, for example, has 8,960 CUDA cores, which is 70 SMs. Each SM also carries Tensor cores and an RT core, which is why AI and ray tracing throughput scales with SM count.

Why dies matter more than names

Two cards on the same die share their memory controller layout, cache hierarchy and media engines, so the difference between them is mostly enabled SM count, clocks and power. The page on RTX 50 series model names explains the suffixes.

The reference specification table

The figures below are NVIDIA reference specifications, with bandwidth calculated from memory data rate and bus width. Partner cards may differ in clocks and size.

Reading the table like an engineer

Three patterns stand out. First, bandwidth tracks bus width and memory type, not CUDA cores. Second, the three lowest cards share a 128-bit bus, so VRAM capacity and clocks differentiate them. Third, the x8 PCIe link on the lower cards matters on older PCIe 3.0 boards, as the guide to PCIe 5.0 and RTX 50 cards explains.

Derived metrics that explain the tiers

Raw specs hide ratios that explain how the cards behave. Two simple derived metrics help.

Memory bandwidth per CUDA core

Dividing bandwidth by CUDA cores shows how much memory throughput each unit of compute can draw on. The figures below are arithmetic from the table above, expressed as gigabytes per second for every 1,000 CUDA cores.

What the ratio tells you

A lower number means the card leans harder on its on-die cache and its compute, and a higher number means more memory headroom per core. The flagship has the lowest ratio because its compute is so large, and large caches and a very wide bus compensate. The RTX 5050 has the highest ratio because its compute is small, even though its absolute bandwidth is the lowest.

What it does not tell you

It does not predict frame rates. Games differ in how bandwidth-bound they are, and cache hit rates change with resolution.

Rated power per relative performance point

The site's RTX 5090 specifications page publishes relative performance with the RTX 5090 set to 100 percent: the RTX 5080 at 66, the RTX 5070 Ti at 57, the RTX 5070 at 45, the RTX 5060 Ti at 35, the RTX 5060 at 30 and the RTX 5050 at 24. These are aggregate figures from that page, and the resolution is not stated. Dividing rated TGP by each score gives a rough efficiency indicator.

Caveats

Rated power is a ceiling, not measured draw, and relative performance is an average across games. Treat the table as an indicator that tiers sit within about 20 percent of each other on this crude measure, and use independent power measurements for real efficiency work. The flagship's higher figure is the price of its extra headroom.

Which workload each card serves

The tiers map to workloads more cleanly than to prices.

RTX 5050 and RTX 5060: 1080p and esports

These cards target 1080p at high refresh rates in esports and mainstream titles. With 8 GB of VRAM and a 128-bit bus, they reward sensible texture settings and DLSS. The site's guides to the RTX 5050 build and RTX 5060 versus RTX 4060 show real behaviour.

RTX 5060 Ti: the memory-configuration card

The RTX 5060 Ti ships with 8 GB or 16 GB on the same 128-bit bus, using clamshell mode for the larger version. That makes it the card where capacity, not compute, decides the experience. See 8 GB versus 16 GB explained.

RTX 5070: 1440p

With 12 GB on a 192-bit bus, the RTX 5070 is a 1440p card that benefits from DLSS in heavy titles. The guide to 12 GB of VRAM on the RTX 5070 examines the limits.

RTX 5070 Ti and RTX 5080: high-refresh 1440p and 4K

Both use GB203 with 16 GB on a 256-bit bus. They differ by 14 SMs and 60 W of rated power. Read the RTX 5070 Ti versus RTX 5080 comparison to decide.

RTX 5090: 4K at maximum settings and heavy compute

With 32 GB and a 512-bit bus, the flagship serves 4K path tracing, large local AI models and creative work. It also demands a strong supply and a large case.

Choosing by your constraints

A decision framework avoids buying on name alone.

Step one: resolution and refresh rate

Work out the pixels per second you need. A 1440p, 165 Hz target is roughly 608 million pixels per second, while 4K at 120 Hz is roughly 995 million. Games vary, but the ratio shows why 4K needs a tier or two more than 1440p.

Step two: VRAM

Estimate your texture and ray tracing budget, as the guide to how much VRAM you need explains.

Step three: power and cooling

Check the power supply wattage, connectors and case length against the card's rating. A card that does not fit your case is the wrong card.

Step four: price and warranty

Compare current UK prices, retailer reputation and warranty. Prices change too fast to print here.

Where to go next in Cluster C

The card-by-card guides continue with the RTX 5070 Ti complete guide, the RTX 5070 complete guide, the RTX 5060 Ti 8 GB versus 16 GB guide and the RTX 5050 versus RTX 3060 comparison. For tuning, see the guides to RTX 5070 Ti overclocking, RTX 5070 Ti undervolting, RTX 5080 undervolting and the RTX 5080 fan curve. For heat and power, read about normal RTX 5060 temperatures, normal RTX 5080 temperatures and RTX 5090 power consumption.

Memory subsystem in more detail

Capacity, bus width and cache define how each card feeds its SMs, and they explain several real-world behaviours.

Bus width, data rate and bandwidth

Bandwidth equals data rate multiplied by bus width, divided by eight. The GDDR7 cards in this family run at 28 Gbps in the figures above, with the RTX 5080 listed at a faster rate to reach 960 GB/s. The guide to GDDR7 memory walks through the arithmetic and shows how clamshell mode doubles capacity without widening the bus.

On-die cache

Every Blackwell die includes a large L2 cache. The RTX 5090 lists 96 MB of L2 on the site's specification page. A large cache raises the hit rate for frequently read data, which reduces demand on external memory and helps narrow-bus cards punch above their raw bandwidth. Cache effectiveness falls as resolution and working set grow, which is one reason lower-tier cards lose ground at 4K.

Why this matters for 4K

At 4K the framebuffer, render targets and ray tracing structures are larger, so a given cache covers a smaller fraction of the working set. Cards with wider buses and more VRAM hold up better, which is why the upper tiers suit 4K.

VRAM capacity and failure modes

When a game exceeds the card's VRAM, assets spill to system memory across the PCIe link. A card with a PCIe 5.0 x8 link and a slow PCIe 3.0 board has far less spill bandwidth than an x16 card, so the same overflow hurts more. Read the guide to PCIe 5.0 and older motherboards for the numbers.

Compute resources and what they do

Each SM bundles several unit types, and knowing them helps you read reviews.

CUDA cores and clock speed

CUDA cores perform FP32 and INT32 arithmetic for shaders. Peak FP32 throughput equals CUDA cores multiplied by two operations per clock multiplied by clock speed. For the RTX 5070 Ti, with 8,960 cores and a rated boost near 2.45 GHz, that gives roughly 44 teraflops, which is a theoretical ceiling, not a game result.

Tensor cores

Fifth-generation Tensor cores run DLSS, Ray Reconstruction and frame generation. They also drive local AI workloads. Tensor throughput scales with SM count, so higher tiers generate frames faster. The DLSS 4 overview explains the features that use them.

RT cores

Fourth-generation RT cores accelerate ray-triangle intersection and bounding volume traversal. Ray tracing and path tracing cost scales with RT core count and cache efficiency, as the explainer on ray tracing versus path tracing discusses.

ROPs, TMUs and the display engine

ROPs write finished pixels, TMUs sample textures, and the display engine drives DisplayPort 2.1b and HDMI 2.1b outputs. These rarely decide purchases, but ROP counts matter for pixel fill at very high resolutions.

Media engines

Ninth-generation NVENC encoders and sixth-generation NVDEC decoders handle capture, streaming and video editing. The engines are shared design elements, so streaming quality is similar across the family, with the higher tiers adding more engine instances on some models.

Power, connectors and cooling by tier

Electrical and thermal requirements climb with each tier.

Power connectors

The RTX 5090 uses a single 16-pin 12V-2x6 connector rated for up to 600 W. Cards in the middle of the range may use one 16-pin connector or conventional 8-pin connectors, depending on the partner design, so check the product page. The article on dedicated GPU cables covers connector seating and cable quality.

Cooler size

Higher rated power means larger heatsinks. Many RTX 5090 partner cards occupy three or four slots and exceed 330 mm in length, while compact RTX 5060 cards can be dual-slot and short. Measure your case before ordering.

Thermal behaviour

Cards in the 130 W to 180 W range run cool and quiet on modest coolers. Cards above 300 W need strong airflow. The site's guide to thermal and power limits explains how clock throttling works at the top end.

Power supply sizing method

Add the card's rated TGP, the processor's peak package power, around 50 to 100 W for the rest of the system, then add headroom of roughly 20 to 30 percent for transients. Compare the total with a quality supply's continuous rating, and see the guide to 1000 W to 1300 W supplies for high-end builds.

Matching card to platform

A balanced system avoids wasted money.

Processor pairing

Lower-tier cards rarely need a flagship processor. At 1080p with high refresh rates, however, processor limits appear sooner because the GPU finishes frames quickly. Read the guide to CPU and system bottlenecks for diagnostic steps.

Memory and storage

Sixteen gigabytes of system RAM is the practical minimum for modern gaming, and 32 GB gives headroom for streaming and background tasks. A fast NVMe SSD reduces asset-streaming stutter.

Monitor pairing

A card should match the monitor. A 4K 144 Hz display wastes a low-tier card's strengths, and a 1080p 60 Hz display wastes a high-tier card. The monitor and resolution pairing guidance in the site's ranked list shows common pairings.

Software features shared across the family

All desktop RTX 50 cards share the same software feature set, which simplifies comparison.

DLSS family

Super Resolution, Ray Reconstruction, frame generation and Multi Frame Generation work across the range, though higher tiers have more headroom to use them at higher settings. See the guides to Multi Frame Generation and DLSS presets.

Reflex and the NVIDIA App

Reflex lowers system latency on every card, and the NVIDIA App provides driver updates, overlays and overrides.

Video features

RTX HDR and RTX Video Super Resolution run on all RTX 50 cards, as the article on RTX HDR and Video Super Resolution explains.

Upgrade arithmetic: what the relative performance chart implies

The relative performance chart on the site's RTX 5090 specifications page also lists earlier generations, which lets you compute implied upgrade ratios. The scores are aggregate figures with the RTX 5090 at 100, and the page does not state a resolution, so treat the ratios as rough guides.

How to read the uplift column

Divide the new score by the old score. A ratio near 1.1 means a small raw improvement that new features, such as Multi Frame Generation, must justify. A ratio above 1.4 means a clear generational step. The table shows that owners of RTX 30 series cards gain the most, and that owners of an RTX 4080 Super gain the least.

Why features change the calculation

Raw ratios ignore DLSS 4 features and GDDR7 bandwidth. If your games support Multi Frame Generation and you own a high-refresh monitor, the effective experience can improve more than the table suggests. If your games do not, the raw ratio is closer to the truth.

Where to cross-check

Compare the table with tests of your own games at your resolution, and see the upgrade comparison in the site's ranked list before you spend.

How to read a relative performance chart critically

Charts are useful, but each has assumptions.

What the numbers aggregate

A single score averages many games, and games differ in how much they stress compute, bandwidth, VRAM and the processor. A card can lead the average and still lose in a specific title.

Resolution and settings

If a chart does not state its resolution, assume a mix. Cards with narrow buses lose ground at 4K, and cards with strong compute gain ground at 1080p when processors are not the limit.

Driver and game versions

Performance changes with drivers, patches and DLSS versions. A chart from the launch period may understate mature drivers.

How to use charts well

Use them to shortlist cards, then confirm with tests of the games you play. Look for 1% low frame rates, not only averages, as the glossary entry on frame time and 1% lows explains.

Future-proofing: what actually extends a card's useful life

People buy a tier up to avoid replacing a card early, and a few factors really do extend its life.

VRAM headroom

Game texture budgets and ray tracing structures have grown with each console generation, and VRAM shortage is the most common reason a card ages badly. Capacity above your current needs is the strongest future-proofing factor.

Memory bandwidth and cache

Bandwidth matters at high resolutions and with heavy ray tracing. A card with a wide bus holds its performance when settings rise.

Feature support

All RTX 50 cards share DLSS 4 and its successors, so a lower tier can still use the newest upscaling models. This narrows the gap between tiers in practice, because upscaling lets a card render fewer pixels, as the DLSS 4 guide explains.

Driver and software support

NVIDIA supports each generation with driver updates for years. A recent architecture such as Blackwell sits at the start of that period, which is a quiet advantage over buying an older card.

What does not future-proof

Extra factory clocks, lighting and expensive coolers add little to longevity. A well-cooled mid-range partner card ages as well as a premium one.

Deciding in ten minutes

If you want a quick decision, follow this short routine.

1. Write down your monitor's resolution and refresh rate.
2. Pick a target: for example 1440p at 120 frames per second in your main games.
3. Choose the lowest tier that reaches the target with DLSS Quality in a demanding game.
4. Check that it has at least the VRAM your settings need.
5. Check your power supply, connectors and case.
6. Compare two or three partner models on price, noise reviews and warranty.

Why the routine works

It starts from your needs, not from a model number, and it uses DLSS as part of the plan, which matches how the cards are designed to be used. If you land between two tiers, choose the higher one only when the extra VRAM or bandwidth solves a specific problem you have.

A note on how this guide was built

Every number in the tables above is either an NVIDIA reference figure, a value from the site's own specification page or a direct calculation from them, such as bandwidth, SM count and watts per relative performance point. No frame rates are quoted, because they depend on the game, driver, settings and test machine. When you need game-level numbers, look for recent tests that publish their methodology, including the processor, memory speed, driver version, resolution, upscaling mode and whether frame generation was on. If a chart omits those details, treat it as marketing, not measurement. This approach keeps the guide accurate as drivers change, and it gives you a repeatable way to check any claim you meet elsewhere.

Keeping the page honest

Specifications are stable, but prices, stock and driver behaviour change. When a figure here conflicts with NVIDIA's product page or a partner's data sheet, trust the manufacturer's current page and treat this guide as the explanation of how to interpret it.

A short glossary for this guide

SM is a streaming multiprocessor, the repeating compute block of each die. TGP is total graphics power, the rated board power. Bus width is the size of the memory interface in bits. Clamshell is a board layout that puts memory chips on both sides to double capacity on the same bus. Bandwidth is bus width multiplied by data rate, divided by eight. The graphics card glossary defines every other term used here.

Verifying what you bought

After purchase, confirm that the card matches the specification.

Step one: read the identity in Windows

Open Task Manager, choose Performance and select the GPU. Check the model name and dedicated memory.

Step two: read the die and shader count

A utility such as GPU-Z shows the die name, the shader count, the ROP count, the memory type and size, and the bus width. Compare each figure with the table above.

Step three: check the link

The bus interface line shows the PCIe link. Run the built-in render test to see the link under load. On the lower cards it should show x8, and on the upper cards x16.

Step four: record the baseline

Run a repeatable benchmark and save the result with the driver version. A baseline helps you spot regressions after later changes.

Common mistakes when choosing a card

Avoid these recurring errors.

Buying on VRAM alone

VRAM capacity prevents stutter, but compute and bandwidth set the frame rate. A card with plenty of VRAM and weak compute still runs slowly.

Ignoring the power and case constraints

A card that exceeds your supply or your case wastes time and money.

Forgetting resolution

A 1080p player gains little from a flagship, and a 4K player is held back by an entry card.

Trusting launch charts alone

Early reviews used early drivers and games. Look for recent tests that match your games and your resolution.

Summary

The RTX 50 desktop family is five Blackwell dies spread over seven products. Bus width, VRAM and rated power separate the tiers more than names do, and derived metrics such as bandwidth per CUDA core and watts per relative performance point show that the tiers trade capability and headroom, not simple efficiency. Choose by resolution and refresh rate, then VRAM, then power and platform, then price. Use the Cluster C guides for the details of each card, and revisit this page when you compare options.

Frequently asked questions about RTX 50 series cards

Which RTX 50 series card is best for 1440p gaming?

The RTX 5070 handles 1440p at high settings with DLSS in demanding games, while the RTX 5070 Ti and RTX 5080 add headroom for high refresh rates and ray tracing. Choose by your target frame rate and VRAM needs, because 12 GB is adequate for many games but limiting in heavy ones.

Are the RTX 5070 Ti and RTX 5080 the same chip?

Yes, both use the GB203 die with a 256-bit bus and 16 GB of GDDR7. The RTX 5080 enables 84 SMs and the RTX 5070 Ti enables 70, with rated power of 360 W and 300 W respectively. The gap is real but smaller than the difference in name suggests.

Why does the RTX 5060 Ti have two memory versions?

NVIDIA offers 8 GB and 16 GB versions on the same 128-bit bus. The larger version places memory chips on both sides of the board in clamshell mode. Bandwidth stays the same, so the extra capacity helps games that exceed 8 GB, not games already limited by speed.

Does the RTX 5050 use GDDR7?

No. The RTX 5050 uses 8 GB of GDDR6 on a 128-bit bus, giving about 320 GB/s of bandwidth. Most other desktop RTX 50 cards use GDDR7. The older memory type helps keep the entry card affordable, and it suits the 1080p workloads the card targets.

What power supply do I need for an RTX 50 card?

It depends on the card and the rest of your system. NVIDIA and partner brands list a recommended supply for each model, for example 750 W for the RTX 5070 Ti. Check the exact card's product page, count your processor's draw and leave headroom for transient spikes.

Which RTX 50 card is the best value?

Value depends on current UK prices and on your resolution. The metrics here show that tiers are close on rated power per relative performance, so the better question is which tier meets your needs without waste. Compare prices in the week you buy, and avoid paying for headroom you will not use.