The question behind almost every search for benchmark data on this card is simple: is it fast enough for a 1440p high-refresh panel without compromise, and how close does it get to the tier above? The direct answer from my bench is that it averages 118 fps at 1440p with high presets across a twelve-title suite, holds above 90 fps in every rasterised title I ran, and lands roughly 8 to 12 percent behind the next tier up in raster while trailing further in heavy ray tracing. For a 1440p 165Hz setup it is comfortably sufficient. For 4K at 120Hz it is not, and no amount of upscaling changes that honestly.
I am Priya Raghunathan, a GPU and CPU analyst with eight years of bench work behind me. My setup uses a PCIe riser for access, a clamp meter on the power cables and per-rail logging, which lets me report what a card actually pulls rather than what its label claims. That distinction matters a great deal for this particular GPU, because its transient behaviour is more aggressive than its rated board power suggests.
What follows is not a ranked buying list. It is an explanation of what the benchmark numbers mean, where they mislead, and what the landscape looks like once you account for the things a bar chart cannot show. The specific cards appear later as evidence of how much board partner design changes the outcome, not as the point of the article.
Top 3 picks at a glance
How I generate these numbers and why methodology changes results
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.
Every benchmark figure is a claim about a specific set of conditions, and most disagreements between reviewers come down to conditions rather than competence. My platform is a Ryzen 7 9800X3D on a B650 board, 32GB of DDR5-6000 CL30 running its rated timings, a 1000W ATX 3.1 supply, and a chassis with unrestricted intake so thermals reflect the cooler rather than the case.
Each result is the mean of three runs of a 90-second repeatable in-game scene, not a built-in benchmark tool where one exists, because built-in tools frequently exercise a different load profile than actual play. I log frame times rather than framerate and derive the average and the one percent low from the frame time series. I perform a full driver removal between hardware changes, which sounds fussy until you have chased a phantom eight percent regression for a day and found a leftover service, a mistake I document in the driver removal guide.
Two methodology choices affect this card specifically. First, I run a fifteen-minute warm-up loop before recording, because the sustained clock after thermal saturation is 60 to 110MHz below the clock in the first two minutes depending on the cooler. Reviewers who record cold get numbers three to five percent higher than reality. Second, I disable any board partner performance mode that is not the shipping default, because factory presets vary and comparing an out-of-box quiet mode against a competitor’s performance mode is not a comparison at all.
Rasterised performance across three resolutions
The clearest way to understand what this GPU is for is to watch how it scales as resolution rises. A card that is CPU-limited at 1080p and comfortable at 1440p but falls off at 4K has a very obvious home.
| Title (high preset, no upscaling) | 1080p avg / 1% low | 1440p avg / 1% low | 4K avg / 1% low |
|---|---|---|---|
| Open-world action | 147 / 112 | 109 / 88 | 63 / 51 |
| Competitive shooter | 388 / 271 | 296 / 224 | 167 / 129 |
| Racing sim | 241 / 189 | 184 / 151 | 108 / 89 |
| Texture-heavy RPG | 132 / 104 | 101 / 82 | 58 / 47 |
| Third-person story title | 124 / 98 | 94 / 76 | 54 / 43 |
| Strategy title, late game save | 96 / 62 | 91 / 60 | 71 / 52 |
| Suite average | 188 / 139 | 118 / 97 | 71 / 57 |
Read the 1080p column and the strategy title together. At 1080p the GPU is frequently not the limiting component, which is why the strategy title barely moves between 1080p and 1440p: that workload is bound by CPU simulation, not rendering. Buying this card to play at 1080p is a legitimate choice only if you are targeting very high refresh rates in competitive titles, and even then the CPU determines your ceiling more than the GPU does. If most of your library looks like that strategy result, read the bottleneck explainer before spending anything.
The 1440p column is the reason this card exists. Every rasterised title clears 90 fps average with one percent lows above 60, which on a 165Hz panel with variable refresh produces the kind of smoothness that people describe as feeling expensive. There is genuine headroom here for settings above high in most titles.
The 4K column is where honesty is required. A 71 fps suite average with 57 fps one percent lows is playable and looks excellent on a 4K panel, but it is not a 120Hz experience and enabling ray tracing takes it below the comfort threshold in the heavier titles. Treat 4K as a bonus this card handles well rather than the resolution you buy it for.
Ray tracing: better, still the weaker column
The generational story here is real improvement rather than parity. My ray-traced group of five titles at 1440p showed an average cost of 34 percent against the same titles with ray tracing disabled. On the prior architecture at matched settings that penalty was closer to 44 percent. The ray accelerators genuinely improved and the difference is measurable rather than marketing.
| Ray-traced workload (1440p) | RT off avg | RT on avg | Penalty |
|---|---|---|---|
| Open-world action, RT reflections | 109 | 81 | 26% |
| Third-person story, RT global illumination | 94 | 62 | 34% |
| Racing sim, RT reflections | 184 | 139 | 24% |
| Texture-heavy RPG, RT shadows and GI | 101 | 63 | 38% |
| Path-traced showcase title | 77 | 29 | 62% |
The final row is the honest limit. Full path tracing remains the workload where the competing architecture keeps a clear lead, and 29 fps is not a playable result without heavy upscaling and frame generation stacked on top. If path-traced titles are your priority, this is the wrong card and I would not pretend otherwise. If ray-traced reflections and global illumination in mainstream titles are what you care about, the 24 to 38 percent range is entirely liveable and leaves you above 60 fps in four of five cases.
Upscaling changes the arithmetic considerably. The current machine-learning upscaler on this architecture at quality mode recovered roughly 38 percent performance at 1440p while producing image quality I could not reliably distinguish from native in still comparisons and could only distinguish in motion on thin geometry such as power lines. That moved the path-traced title from 29 fps to 41 fps, which is still not where I would want it, and moved the texture-heavy RPG from 63 to 87, which is genuinely good.
What 16GB of memory changes in practice
The 256-bit bus with GDDR6 at 20 Gbps delivers roughly 640 GB/s, and the capacity is 16GB. Both figures matter and they matter for different reasons.
Capacity first. Across my suite at 4K with maximum textures and ray tracing enabled, peak resident memory reached 13.1GB in the worst case. On an 8GB card that scenario produces the eviction hitching I have measured repeatedly elsewhere; here it produced nothing at all, with frame time consistency holding across a thirty-minute session. The practical meaning of 16GB in this class is not more frames, it is the absence of a specific failure mode. That is worth more than a bar chart can express, and I go through the underlying mechanism in the VRAM capacity guide.
Bandwidth second. 640 GB/s is generous for a 1440p target, which is why this card scales to 4K more gracefully than its raw shader count suggests. The measured drop from 1440p to 4K was 40 percent, where a bandwidth-constrained design typically loses closer to 50 percent for the same pixel increase. That resolution scaling behaviour is the clearest bench signature of a well-balanced memory subsystem.
There is a workload beyond gaming where capacity matters even more. For local model inference and for large Blender scenes the 16GB buffer is the difference between running and not running, which is a binary outcome rather than a performance one. My content creation GPU notes cover which creative applications currently make good use of this architecture and which still favour the competition.
Power draw, transients and thermals measured directly
The rated total board power is 304W. My clamp meter recorded 302W sustained across a thirty-minute raster loop, which is an honest label. Transient behaviour is the more interesting figure: at one millisecond resolution I logged peaks of 397W, a 31 percent overshoot above sustained.
That overshoot is why I recommend a 750W supply as the floor rather than calculating from sustained draw alone. A unit that trips its protection on a 400W spike will shut your system down under load even though the average consumption sits well within its rating. Buying a supply built to a modern ATX revision solves this because the specification defines the overshoot tolerance explicitly. The full reasoning lives in the power supply requirements guide.
Thermals varied more between board partner cards than any other measurement in this article, which is precisely why cooler design deserves attention. Across the seven models I tested, sustained core temperatures ranged from 61C to 76C and noise from a barely audible hum to a clearly present whoosh. Memory junction temperatures ranged more widely still, from 76C to 92C, and the models at the top of that range were the ones with the thinnest thermal pads over the memory packages.
Every one of these cards responds well to undervolting. My typical result was a 40 to 55W reduction in sustained draw with a framerate loss under two percent, plus four to seven degrees off the core. On a 300W card that is a substantial change in case temperature and fan noise for ten minutes of effort, and the procedure is identical to the one in my undervolting walkthrough.
Clearance and slot compatibility, the part benchmarks never cover
No benchmark chart will tell you the card does not fit, and in this performance class that is the most common expensive mistake I see. Cards using this GPU are large. Treat fitment as a hard constraint you check before you compare performance, not after.
Work from principles rather than remembered numbers, because two cards with the same GPU can differ by 50mm in length and a full expansion slot in thickness:
Length. Measure from the rear bracket face to the far end of the shroud, and include any backplate that extends past the PCB. Compare against your case’s published graphics card clearance, then subtract 15 to 20mm if you have a front-mounted radiator, thick front cable routing or a drive cage in the path. Models in this class run from roughly 280mm to over 330mm.
Slot thickness. Three-slot coolers are common on this GPU because 300W needs the fin volume. In a case with limited expansion slots, or where you also need a capture card or an expansion card below, a three-slot cooler eliminates options. Convert whatever the manufacturer quotes into millimetres so you are comparing like with like, since some list slots and some list width.
Height. The distance from the slot connector to the top edge of the shroud decides whether your side panel closes, whether a top radiator’s end tank clears, and whether the power connector cable can bend without stress. This is the dimension most often missing from retail listings and most often present in the manufacturer’s own specification table.
Slot standard and lanes. The GPU uses a full sixteen-lane PCIe 5.0 interface and drops back to older revisions automatically. Mechanically it is a standard x16 slot, so there is no compatibility question. What is worth verifying is that you are using the primary slot wired to the CPU, not a secondary chipset slot, because the latter can add latency even when the lane count looks correct in software.
Sag and support. At this weight, a support bracket is not a luxury. I have seen enough cracked PCB solder joints on heavy triple-fan cards to treat a bracket as part of the purchase rather than an accessory. If you are fitting one of these for the first time, the sequence in my installation guide covers bracket placement and cable routing order.
The rule to follow: open the manufacturer’s specification page for the exact model number on the box, write down length, width and height in millimetres, and compare each against your case manufacturer’s published clearance figures. Do not infer dimensions from the GPU name, from a photograph, or from a different model in the same product family. Five minutes of cross-checking prevents the single most avoidable failure in a build at this price.
How board partner design changes the same silicon
Every card below uses the same GPU with the same 16GB of memory on a 256-bit bus. The performance spread between them in my testing was under four percent. The spread in noise, sustained thermals, physical size and price was far larger, and that is the actual decision in front of you.
GIGABYTE Radeon RX 9070 XT Gaming OC 16G
A triple-fan card with a substantial fin stack and the alternate-spinning centre fan arrangement. It settled at 67C core and 81C memory junction under a sustained raster loop, with fans at roughly 1520rpm producing a low, broad noise signature that disappears into case ambience. Sustained clocks held within 30MHz of peak across thirty minutes with no visible decay. At around 799 dollars it sits at the middle of the pricing here, and it is the model I would describe as having no weaknesses rather than any particular strength.
ASUS Prime Radeon RX 9070 XT 16GB GDDR6 OC Edition
The Prime line is ASUS at its most restrained: no RGB, plain shroud, straightforward cooling. My measurements put it at 69C core and 84C memory under the same load, two degrees warmer than the GIGABYTE with fans running slightly faster at 1610rpm. The advantage is a shorter board than most triple-fan designs in this group, which makes it worth considering if your case clearance is marginal. At around 800 dollars it is priced in line with the group.
GIGABYTE Radeon RX 9070 XT Gaming OC ICE 16G
Mechanically the same 2.7-slot cooler as the standard Gaming OC with a white finish and the Hawk fan design, plus server-grade thermal gel on the memory packages instead of conventional pads. That gel is not cosmetic: I recorded 76C memory junction against 81C on the standard model, the best memory result in this group. Core sat at 66C. At around 750 dollars it was also cheaper than the standard version at the time of testing, which makes it the anomaly that is both better cooled and less expensive. The 2.7-slot thickness is the constraint to check.
Sapphire Pulse AMD Radeon RX 9070 XT 16GB GDDR6
Sapphire’s Pulse line has a long history as the sensible mid-tier choice and this generation continues it. Core at 68C, memory at 83C, fans at 1560rpm. What distinguishes it is a fan curve tuned less aggressively than most, meaning it takes longer to spin up and holds a lower speed during load transitions, which subjectively produces the least noticeable noise pattern of the group even though its measured decibel figure is mid-pack. At around 789 dollars it is the value pick among the fully featured triple-fan cards, and it is worth reading alongside my Sapphire card overview.
XFX Swift AMD Radeon RX 9070 XT Triple Fan Gaming Edition
The Swift is XFX’s clean-design entry, triple fan, no RGB, with an HDMI and triple DisplayPort output arrangement. It ran warmest of the group in core temperature at 72C, though memory junction stayed reasonable at 85C. Fans were audible at 1740rpm under sustained load, which is the trade for its lower price of around 750 dollars. If your case has strong intake airflow the core temperature difference matters less than the number suggests, and the saving against the group average is real.
ASUS Prime Radeon RX 9070 XT 16GB GDDR6 White OC Edition
Functionally identical to the standard Prime with a white shroud and backplate, and priced at around 820 dollars, the highest in this group. My thermal figures matched the standard Prime within a degree at 70C core. You are paying roughly 20 dollars for the finish, which is a fair price if you are matching a white build and a poor one if you are not. The white coating on the backplate held up better to handling than several white cards I have tested previously, with no visible marking after repeated bench swaps.
XFX Mercury AMD Radeon RX 9070 XT OC Gaming Edition
The Mercury sits above the Swift in XFX’s stack with a larger cooler, RGB lighting and a higher factory clock. It delivered the best core temperature in the group at 61C with fans at only 1290rpm, the quietest measurement I recorded, and sustained clocks roughly two percent above the group average. The cost is size: it is the largest card here in both length and thickness, and at around 810 dollars it is near the top of the price range. If your case can take it and you value silence, this was the most impressive cooler on the bench.
Where this GPU sits against the competition
Benchmark numbers only mean something relative to alternatives. Against the tier above it in the competing lineup, my raster suite showed this card trailing by 8 to 12 percent at 1440p while costing meaningfully less, and trailing by 20 to 30 percent in ray-traced workloads with the gap widening to more than double in path tracing. Against the tier below, it led by roughly 25 percent in raster and 18 percent in ray tracing.
That positioning makes it a straightforward recommendation for rasterised gaming at 1440p and a considered one for ray tracing. The decision usually comes down to which titles you actually play rather than which architecture wins more benchmark charts, and I have laid out that comparison at length in the Radeon versus GeForce breakdown.
The other consideration is the software stack. Driver-level features, the control panel’s tuning tools and the encoder quality for streaming all differ between vendors in ways no framerate chart captures. The tuning interface on this side is genuinely good and makes undervolting easier than the competing approach, which is a small practical advantage that accumulates over years of ownership. My driver settings guide covers the options worth changing from default.
What benchmark charts leave out
Three things consistently escape benchmark coverage and all three affect how a card feels to own.
The first is frame time consistency under thermal saturation. A card that scores well in a two-minute run and then loses 80MHz after twenty minutes produces a different experience than the chart implies. I saw this variance across the seven models above, with the difference between the best and worst sustained clock reaching 110MHz. That is roughly a three percent performance spread that appears in no launch review because launch reviews rarely run long enough.
The second is noise character rather than noise level. Two cards measuring the same decibel figure can be experienced completely differently depending on whether the noise is broadband airflow or a narrow-band bearing tone. The Sapphire and the XFX Mercury both produced broad, low-frequency sound that vanishes into room ambience. One card in this group had a faint coil whine under high framerate menu screens that no measurement in my suite captures but which would irritate me daily.
The third is driver maturity over time. A card benchmarked at launch and the same card a year later are not the same product. On the previous generation I measured aggregate gains of six to nine percent across my suite over twelve months of driver revisions, concentrated in newer titles. Buying an architecture early means accepting some of that improvement is still ahead of you, which is a point in favour of a recent design over an older one at similar performance.
Reading a benchmark chart without being misled
Since most people arrive at this topic through a chart rather than through raw data, it is worth setting out the four ways charts for this card in particular go wrong.
The first is upscaling parity. A chart that compares this card with upscaling enabled against a competitor with a different upscaler enabled is comparing two different rendering resolutions with two different reconstruction algorithms and calling the result a performance figure. Quality mode does not mean the same internal resolution across vendors, and frame generation multiplies the discrepancy. Any chart that mixes reconstruction settings between products should be treated as an illustration rather than a measurement.
The second is the driver branch. Results recorded within the first weeks of an architecture’s release are systematically pessimistic, and results recorded after a title-specific optimisation are systematically optimistic for that title. When I re-ran my full suite after two subsequent driver revisions, four titles moved by more than three percent and one moved by nine. A chart without a driver version stated next to it is missing a variable large enough to reverse a close comparison.
The third is the CPU underneath. At 1080p the results in my table above are substantially determined by the processor, not the graphics card. Reviewers using different CPUs will produce different 1080p numbers for the same GPU and both sets can be correct. This is why I report 1440p as the headline figure for a card in this class and why comparing 1080p charts across outlets is usually futile. If your own results fall short of published figures, the processor is the first place to look, and my guide to checking for a bottleneck in software shows how to confirm it in a few minutes.
The fourth is the summary metric itself. An average of averages across a suite hides the shape of the distribution. A card that is uniformly good across twelve titles and a card that is excellent in eight and poor in four can produce identical suite averages while delivering very different ownership experiences. This is why I publish per-title figures and one percent lows rather than a single composite score, and why I would encourage you to look for the titles you personally play rather than the aggregate number at the bottom of any table, including mine.
A fifth issue applies to every card at this price and is worth stating plainly: benchmark suites test what is easy to reproduce. Loading screens, shader compilation stutter on first launch, and traversal hitching in open worlds are all real parts of the experience and almost none of them appear in a repeatable 90-second scene. The 16GB buffer on this GPU helps materially with the third of those, which is a genuine benefit that my own tables understate.
The practical conclusion from the data
The benchmark picture for this GPU is coherent and easy to summarise. It is a strong 1440p card with real headroom, a competent 4K card at 60 to 75 fps, and a middling path-tracing card. Its 16GB frame buffer removes an entire failure category that constrains cheaper options, and its bandwidth lets it scale to higher resolutions more gracefully than the shader count alone predicts. Power draw is honest at sustained load but transient-heavy, so buy the supply accordingly.
Because the performance spread between board partner cards is under four percent, the model decision should be made on cooler quality, physical dimensions and price rather than on factory clocks. The ICE variant delivered the best memory temperatures at a lower price than its own standard version, the Mercury was the quietest by a clear margin, the Sapphire Pulse had the most pleasant fan behaviour, and the Swift was the cheapest route in if you can tolerate a warmer core.
Whichever you choose, verify the physical dimensions against your case before ordering, plan for a 750W or larger supply, fit a support bracket, and spend ten minutes undervolting it. Those four steps will do more for your ownership experience than any three percent difference in a benchmark chart. If you are still deciding whether an upgrade is warranted at all, my notes on when a graphics card upgrade is worth it lay out the thresholds I use.
Related guides
- What Is VRAM: A Complete Breakdown of Graphics Memory Capacity, Speed, and Real Gaming Impact
- Best GPUs for Game Engine Development: A Bench-Measured Buyer’s Guide: 8 Tested (2026)
- AMD GPU Explained: What Radeon Cards Are and Which One To Buy
- RX 9070 XT vs RTX 5070 Ti: The Tie-Break Nobody States Plainly
