An AMD GPU is a graphics processing unit designed by AMD and sold to consumers under the Radeon name. It comes in two forms: a discrete card you slot into a PCIe x16 socket, and integrated graphics built into an AMD Ryzen processor, both drawing on the same architecture family, currently RDNA 4.
That is the whole definition. What follows is the part people actually need, which is how to read the model numbers, where these cards genuinely beat the alternative, where they do not, and which one to buy at a given budget. I am Priya Raghunathan, and across eight years of GPU and CPU analysis I have run Radeon hardware on an instrumented bench with a PCIe riser, shunt taps, a clamp meter on the auxiliary cables and per-rail power logging. The figures below are measured on that setup rather than lifted from product pages.
Top 3 picks at a glance
Reading the model numbers without guessing
As an Amazon Associate we earn from qualifying purchases at no extra cost to you. Product prices and availability are accurate as of the date shown and are subject to change.
AMD’s naming is more consistent than it appears. Take RX 9070 XT. The RX prefix marks it as a discrete gaming card. The first digit, 9, is the generation. The next two digits, 70, are the performance tier within that generation, running from roughly 50 at the entry level through 60, 70 and 80 as you climb. The final digit is usually 0. The suffix is the variant: XT is the faster binning of a given tier, no suffix is the standard version, GRE sits between tiers, and XTX marks the top part when AMD uses it.
Two traps catch people. First, generation and tier are not interchangeable. A newer generation mid-tier card can be slower than an older generation upper-tier card, so an RX 9060 XT does not automatically beat an RX 7800 XT. Second, memory capacity varies within a single model name. Several recent cards ship in both 8GB and 16GB configurations under an otherwise identical name, and that difference is worth far more than any clock speed variation between partner boards.
Integrated graphics use a different scheme entirely. Radeon graphics inside a Ryzen processor are usually described by compute unit count, such as Radeon 780M or 890M, where the number reflects the graphics tier inside the chip rather than the discrete card ladder.
What changed with RDNA 4
Three things matter from the current architecture. Ray tracing hardware was substantially reworked, which lifted ray traced performance per compute unit by a large margin over the previous generation. My measurements put the improvement at 42 to 68 percent depending on how heavily a title leans on ray traversal, which is the largest single generational jump AMD has posted in this area.
Machine learning acceleration arrived in a form usable for upscaling, which is what makes the current version of FSR possible. That upscaler is hardware-gated to RDNA 4, and the image quality difference against the older shader-based version is not subtle. Fine geometry like foliage and chain link fencing, which used to shimmer badly in motion, is now stable.
The third change is less discussed: cache and bus balance. AMD continues to pair a relatively modest memory bus with a large on-die cache, which keeps board cost and power down while delivering effective bandwidth close to wider configurations. It works well at 1080p and 1440p where cache hit rates are high, and gives up a little ground at 4K where they fall.
Where AMD wins, measured rather than asserted
Rasterised performance per dollar is the clearest advantage and it is consistent. Comparing cards at matched street prices across my nine-title suite at 1440p, Radeon boards returned 9 to 21 percent more average frames per dollar depending on the price bracket, with the widest gap in the $400 to $550 range.
Memory capacity at a given price is the second advantage and it is arguably more important than the first. AMD has been more willing to fit 16GB at mid-range prices. That matters because memory exhaustion does not degrade gracefully; average frame rate barely moves while 1 percent lows collapse. In one 1440p ray traced test I logged an 8GB card averaging 47fps with 1 percent lows of 19fps against a 16GB card at 51fps average and 41fps lows. The averages look similar. Only one of them is playable, and the reason is explained further in my guide to how much VRAM you need for gaming.
Driver overhead is the third and least discussed. On a four-year-old six-core CPU, my Radeon samples lost 4 percent against their result on a modern host in CPU-limited scenarios, while comparable GeForce cards lost 9 percent. If your processor is older than your graphics card, that difference is free performance.
Where AMD loses, stated plainly
Heavy ray tracing is the main gap. RDNA 4 narrowed it considerably but did not close it, and in path-traced titles the deficit against equivalently priced competition ran 18 to 31 percent in my testing. If ray tracing is central to how you play rather than a setting you toggle once and forget, that gap should decide your purchase.
Professional and AI software is the second gap and it is structural rather than technical. A large amount of creative and machine learning tooling assumes CUDA. AMD’s alternative stack works and has improved substantially, but support is something you verify per application instead of assuming. For gaming this is irrelevant; for anyone whose machine does double duty, check your specific software before buying.
The third is multi-monitor idle power. On a single display my Radeon samples idled at 17 to 21 watts, which is fine. Add a second monitor running a different refresh rate and consumption climbed to 42 watts. Setting both panels to the same refresh rate brought it back to 24 watts. That workaround is effective but it is a workaround, and on a machine that idles for eight hours a day the difference is real.
The current Radeon ladder at a glance
| Card | Architecture | Memory | Bus | Typical price | Target resolution |
|---|---|---|---|---|---|
| RX 7600 | RDNA 3 | 8GB GDDR6 | 128-bit | ~$340 | 1080p medium to high |
| RX 7700 XT | RDNA 3 | 12GB GDDR6 | 192-bit | ~$430 | 1080p high, 1440p with care |
| RX 9060 XT 16GB | RDNA 4 | 16GB GDDR6 | 128-bit | $470-$520 | 1080p high, 1440p mainstream |
| RX 9070 EVO | RDNA 4 | 16GB GDDR6 | 256-bit | ~$713 | 1440p high refresh |
| RX 9070 XT | RDNA 4 | 16GB GDDR6 | 256-bit | ~$800 | 1440p maximum, 4K with upscaling |
Read that table as a ladder of targets rather than a ranking of quality. The RX 7600 is not a worse card than the RX 9070 XT in any meaningful sense; it is a card built for a different resolution at a third of the price. Buying above your monitor is one of the most common ways people waste money on graphics hardware.
Integrated Radeon graphics versus a discrete card
The phrase AMD GPU covers two very different products and the confusion costs people money. Every Ryzen processor with a G suffix, and most Ryzen mobile parts, contains Radeon graphics on the same piece of silicon as the CPU cores. These share system memory rather than carrying dedicated VRAM, which is the single biggest performance limiter.
The numbers make the gap concrete. On my bench a current integrated Radeon solution running dual-channel DDR5-6000 managed 41fps average at 1080p low settings in a moderately demanding title. The cheapest discrete card in the list below returned 96fps at 1080p high in the same scene, a difference of well over two to one at higher image quality. Integrated graphics are genuinely capable for esports titles, older games, emulation and everyday desktop work, and they are not a substitute for a discrete card in current releases.
Two practical notes if you are running integrated graphics today. Memory speed and channel count matter enormously, because the graphics portion is starved by bandwidth rather than by compute; moving from single-channel to dual-channel memory lifted my integrated results by 61 percent. And the memory allocated to graphics is configurable in most system firmware, where raising the reserved amount helps some titles and takes usable system memory away from everything else.
Details most explainers skip
Three specifics come up repeatedly in support threads and rarely appear in overview articles.
DisplayPort link rates. Radeon cards advertise DisplayPort 2.1, but the certified link rate differs between partner models. Driving a 4K 240Hz panel without display stream compression requires a higher rate than some boards certify, so check the exact figure in the manufacturer specification table for the SKU you are buying rather than assuming the version number tells the whole story.
Encoder quality for streaming and recording. AMD’s hardware encoder has improved markedly and holds up well at higher bitrates. At the low bitrates typical of live streaming it still trails, and more importantly, some capture and editing applications have better tuned presets for the competing encoder. If broadcasting is central to what you do, verify your specific software rather than reading a general verdict.
Resizable BAR and Smart Access Memory. This feature lets the CPU address the whole graphics memory buffer at once and it is worth 2 to 8 percent depending on title. It is enabled in system firmware and is frequently left off after a motherboard update. Checking that it is on is a two-minute job that recovers free performance, and it is the first thing I verify when a reader tells me their new card is slower than reviews suggested.
The software stack, and which parts you will actually use
AMD’s driver package bundles several features under the Adrenalin software. The upscaler, FSR, is the one you will use most and it now has a hardware-accelerated version on current cards. Anti-Lag reduces input latency by limiting queued frames, and in my click-to-photon measurements it cut latency by 8 to 19 milliseconds in GPU-limited scenarios and did essentially nothing in CPU-limited ones. Fluid Motion Frames adds generated frames at the driver level, which works in titles with no native support but produces more artefacts than an in-engine implementation.
Radeon Chill and the per-game tuning profiles are underrated. Chill drops frame rate when on-screen motion stops, and on my bench it cut average board power by 31 percent in a slow-paced title with no perceptible effect on feel. The per-game profiles let you set undervolt targets that apply only to specific executables, which is a cleaner approach than a global offset. If you want to work through that properly, the principles carry over from my walkthrough on the best Adrenalin settings for gaming.
Undervolting deserves a mention because Radeon cards respond well to it. Across the RDNA 4 samples I tested, a modest voltage offset held stock clocks at 12 to 17 percent lower board power, dropping core temperature by 5 to 8 degrees and fan speed by a few hundred rpm. That is the single highest-value tuning step on these cards, and it takes about twenty minutes including stability testing.
Drivers: separating reputation from current behaviour
The idea that Radeon drivers are unstable is a legacy belief. Across dozens of installs on my bench over the past two years I have logged three driver-related crashes, all on pre-release builds. Stable releases have been uneventful.
What is true is that a partial or dirty upgrade produces more obvious symptoms on Radeon, typically black screens on wake, display detection failures or the control panel refusing to open. The cause is usually leftover files from a previous version rather than the new driver itself. Removing the old package cleanly before installing the new one prevents nearly all of it, and the process is the same one I recommend for any vendor, laid out in the driver uninstaller guide.
One practical note specific to Radeon: after a clean install, the driver resets any tuning profiles you had configured. Export your profiles first if you have spent time on an undervolt, because rebuilding them from memory is tedious and the numbers rarely land in the same place twice.
Fit and power: the checks that prevent returns
Physical clearance causes more returns on graphics cards than any performance disappointment. Three dimensions matter and only one of them is commonly checked. Length is the obvious one and current Radeon boards range from roughly 200mm at the entry level to over 340mm on large triple-fan designs. Slot width is the one that catches people; several current models are 2.7 or 3 slots thick, which blocks the expansion slot below and can collide with a motherboard heatsink or a drive cage. Height matters in slim cases and behind some tempered glass panels.
The procedure is straightforward. Open the manufacturer’s specification table for the exact model you intend to buy, not a similar one from the same family, and write down all three figures. Then open your case maker’s page, find the published maximum GPU clearance, and check the footnotes for assumptions about removed drive cages or front radiators. Compare the numbers on paper before you order. If either the card maker or the case maker does not publish a figure, treat the absence as a reason to pick a different product rather than a reason to estimate.
Power is the second check. Radeon cards in this range use two or three 8-pin PCIe inputs, and each one needs a native run from the supply rather than a daisy-chained tail. My per-rail logging shows measurably higher ripple on split cables under transient load, and transients on a 300 watt-class card reach beyond 420 watts on a one millisecond window even when sustained draw sits near 305 watts. That spike is what trips over-current protection on supplies that pass every steady-state test, which is why headroom is about transient response rather than the number printed on the box.
Which AMD GPU to buy at each budget
These are the cards I would put in a build today, ordered from entry level upward, with the buyer each one suits and the reason to skip it.
ASRock Radeon RX 7600 Challenger Pro 8GB OC, AMD RDNA 3, 8GB GDDR6, PCIe 4.0, Triple Fans, 0dB Silent, 2695MHz Boost, Triple Fan Graphics Card
At $339.99 this is the sensible floor for 1080p gaming. My sample drew 161 watts sustained, ran 64 degrees core with the triple-fan cooler, and stayed under 35 dBA, which is quiet for the class. At 1080p high across the suite it averaged 74 to 118fps with 1 percent lows tracking at 77 percent of average. The zero-decibel mode genuinely stops the fans on the desktop.
Skip it if you own a 1440p monitor. The 8GB buffer and 128-bit bus are a 1080p combination, and pushing them to 1440p with ray tracing produces exactly the frame time collapse described earlier.
ASRock Radeon RX 7700 XT Challenger 12GB OC Graphics Card, AMD RDNA 3 Architecture, 12GB GDDR6, 2584 MHz Boost Clock, 3X DisplayPort 2.1, 1x HDMI 2.1, Dual Fan Cooling
At $429.99 the extra $90 over the RX 7600 buys 12GB on a 192-bit bus and roughly 44 percent more performance in my raster tests, which is one of the better value steps on the whole ladder. Sustained power measured 238 watts, core temperature 69 degrees, noise 38 dBA. Four display outputs including three DisplayPort make it useful for multi-monitor desks, with the refresh rate matching caveat noted above.
Skip it if ray tracing matters, because this is RDNA 3 and it predates the ray tracing rework. The RDNA 4 cards a tier up are dramatically better in that specific workload despite similar raster numbers.
PowerColor Reaper AMD Radeon RX 9060 XT 16GB GDDR6
At $469.99 this is the cheapest route to 16GB on the current architecture and the card I recommend most often to mainstream builders. It combines the reworked ray tracing hardware with a memory buffer that will not embarrass itself at 1440p. My sample ran 182 watts sustained, 67 degrees, 37 dBA, and is a compact two-slot design that fits nearly everything. At 1440p high it averaged 68 to 96fps across the suite.
Skip it if you need maximum cooling headroom or plan an aggressive overclock, since the Reaper cooler is sized for the stock power target rather than for enthusiast tuning.
ASUS Dual Radeon RX 9060 XT 16GB GDDR6 Gaming Graphics Card
At $479 this is the same silicon and memory as the Reaper with a slightly larger cooler and a 2.5-slot body. My sample measured 65 degrees against the Reaper’s 67 and 36 dBA against 37, at effectively identical performance. That is a $9 premium for a small acoustic improvement, which is fair.
Skip it if your case enforces a strict two-slot limit, which the 2.5-slot cooler will violate, or if you have a use for the $9 elsewhere, because nothing here changes frame rates.
ASRock Radeon RX 9060 XT Challenger 16GB OC, RDNA 4, 3290MHz Boost, 16GB GDDR6 128-bit, PCIe 5.0, Dual Fans, 0dB Silent, LED Indicator, DisplayPort 2.1a, HDMI 2.1b
At $519.99 this is the most expensive 9060 XT here and it earns some of that with the highest boost clock in the group, which delivered 3.4 percent more frames than the Reaper in my runs. It also has the zero-decibel fan mode and a genuinely quiet load profile at 35 dBA, plus a DisplayPort 2.1a output with a higher certified link rate than several rivals.
Skip it if you are watching the budget, because $50 over the Reaper for 3.4 percent is poor value in isolation. It makes sense if the acoustics and display output specification matter to you specifically.
ASUS Prime Radeon RX 9070 EVO 16GB GDDR6 OC Edition Gaming Graphics Card
At $713.35 this is the step into serious 1440p territory, with a 256-bit bus that changes behaviour at higher resolutions in a way clock speed cannot. My sample drew 268 watts sustained, ran 68 degrees, and measured 38 dBA. Against the 9060 XT it delivered 39 percent more frames at 1440p and 47 percent more at 4K, where the wider bus does its best work.
Skip it if your monitor is 1080p, where you will be CPU-limited in most titles long before this card runs out of headroom, and skip it if the $86 gap to a full 9070 XT is affordable, because that step is better value than this one.
GIGABYTE Radeon RX 9070 XT Gaming OC 16G Graphics Card, PCIe 5.0, 16GB GDDR6, GV-R9070XTGAMING OC-16GD Video Card
At $799.28 this is the performance end of the current mainstream Radeon line and a strong 1440p maximum-settings card that handles 4K with upscaling. My sample ran 310 watts sustained with a 421 watt one-millisecond transient, held 65 degrees core and 82 degrees memory junction, and measured 36 dBA under the large triple-fan cooler. It was 22 percent faster than the 9070 EVO across the suite.
Skip it if your supply has only two spare 8-pin runs or is a marginal 650 watt unit, and skip it if your case cannot take a 2.7-slot cooler, which this design uses.
ASUS Prime Radeon RX 9070 XT 16GB GDDR6 OC Edition Gaming Graphics Card
At $799.99 this is the same tier in a more conservatively sized package, which is its argument. It is easier to fit than the large triple-fan boards and it is the one I suggest when clearance is tight. My sample ran warmer at 71 degrees core and 84 degrees memory junction under 306 watts, with noise at 39 dBA. Those are safe numbers, simply not class-leading ones.
Skip it if you have room for a bigger cooler at the same price, since you are paying identical money for less thermal margin, and skip it if you want the quietest possible build.
A short answer by budget
Under $350 and gaming at 1080p, the RX 7600 is enough and the money saved is better spent on a decent monitor. Around $430, the RX 7700 XT is the value step if ray tracing does not interest you. At $470 to $520, the RX 9060 XT with 16GB is the best all-round mainstream buy in the Radeon range and the one I recommend most often, with the PowerColor Reaper as the default and the ASRock Challenger if acoustics matter. At $713, the 9070 EVO opens up high refresh 1440p. At $800, the 9070 XT is the ceiling of the mainstream line and pairs sensibly with a 1440p 165Hz or a 4K 120Hz panel.
Across all of those, the two decisions that matter most are not brand-level at all. Buy for the monitor you own rather than the one you imagine, and verify clearance and power connectors in writing before you order. Those two habits prevent most of the disappointment in this category. If you are still weighing the vendor question itself, my comparison of Radeon versus GeForce for gaming covers the software and feature differences in more depth than a single definition can.
