Comparing graphics cards takes three passes — hardware specs, architecture features, and measured frame rates in the games you actually play.
For the full breakdown, see our best Worlds GPU guide.
Two graphics cards can post wildly different numbers and still trade wins in the same game. Learning how to compare GPUs comes down to reading three separate languages — the spec sheet, the architecture behind it, and real frame rates — and never letting one answer for the other two.
The card with 21,760 CUDA cores is not five times the card with 4,096 stream processors, and 3,352 AI TOPS will not line up against 128 AI accelerators. Those figures describe different machines. Read each layer on its own terms and the confusion falls away.
What Do GPU Specs Actually Tell You?
Specs describe theoretical capability under ideal conditions — they set the ceiling, not the result. The exact model and generation matter before anything else, because board partners ship the same chip with different clocks, coolers, power limits, and dimensions.
Here is how two current cards line up on the fields that carry the most weight.
| Field | GeForce RTX 5090 | Radeon RX 9070 XT |
|---|---|---|
| Architecture | Blackwell | RDNA 4 |
| Compute units | 21,760 CUDA cores | 64 compute units, 4,096 stream processors |
| Ray tracing and AI | 318 ray-tracing TFLOPS, 3,352 AI TOPS | 64 ray accelerators, 128 AI accelerators |
| Clocks | 2.41 GHz boost, 2.01 GHz base | Up to 2,970 MHz boost, 2,400 MHz game |
| Memory | 32 GB GDDR7, 512-bit | 16 GB GDDR6, 256-bit, up to 640 GB/s |
| Board power | 575 W | 304 W typical |
| Recommended PSU | 1,000 W system | 750 W minimum |
- CUDA cores and stream processors count different hardware, so the two numbers cannot be ranked against each other.
- TFLOPS, TOPS, clock speed, and bandwidth are not universal scores.
- “Boost” is a ceiling, not a guaranteed sustained speed.
- Extra VRAM helps with high-resolution textures and some professional work; it does not raise frame rates on its own.
- A 575 W card and a 304 W card demand different PSUs, case airflow, and build budgets.
Architecture And Features Change What The Numbers Mean
Architecture decides which features a card has and how a given frame rate gets produced. DLSS 4 and multi-frame generation are RTX 50-series features, so a chart run with them enabled is not a native-rendering comparison.
Vendor figures carry the same caveat. Both companies publish their own metrics on their own designs — NVIDIA’s RTX 5090 specifications list ray-tracing TFLOPS and AI TOPS, while AMD’s RX 9070 XT page lists ray accelerators and AI accelerators — and neither set doubles as a benchmark.
Feature support is a gate as well. AMD lists RX 9070 XT support for Windows 10 64-bit, Windows 11, and Linux x86 64-bit, so a card that drops your operating system loses part of its value no matter where it sits on a chart.
GPU Comparison Charts: What Decides The Outcome
A benchmark chart means something only when game version, preset, resolution, ray tracing, driver, and upscaling mode all match. Change one of those and the ranking can flip.
- Match the test conditions. Same game patch, preset, resolution, ray-tracing setting, driver, and upscaling mode.
- Separate native from upscaled. DLSS, FSR, and frame generation belong in their own row, never folded into a native result.
- Read the 1% lows. Average FPS hides stutter; frame-time plots expose it.
- Check what is limiting the run. A CPU-bound test at low resolution narrows the gap and flatters both cards.
- Weight several games. Rasterization, ray tracing, encoding, and productivity workloads rarely rank cards the same way.
- Discount vendor slides. They are demonstrations, not independent reviews.
Price the whole system, not just the card. PSU headroom, cooling, case clearance, and monitor resolution decide whether a fast GPU actually delivers, and any price needs a country, retailer, date, and exact SKU attached before it means anything.
Once the shortlist is down to two or three models, matching each one to your monitor is the last step. Our tested roundup of current graphics cards pairs each pick with the resolution it suits.
Layer order beats spec order: name the exact card, confirm the features you need exist on your operating system, then compare frames under matched settings.
Common Questions
Is More VRAM Always Better For Gaming?
No. The 32 GB on an RTX 5090 against the 16 GB on an RX 9070 XT matters most with high-resolution texture packs and some professional workloads, but spare capacity does not lift frame rates by itself. Bandwidth, cache, memory compression, drivers, and the workload decide what the extra memory actually buys you.
Why Do CUDA Cores And Stream Processors Not Compare Directly?
The two figures count different hardware. NVIDIA’s RTX 5090 lists 21,760 CUDA cores, and AMD’s RX 9070 XT lists 64 compute units containing 4,096 stream processors. Each number reflects its own architecture’s design, so placing them side by side says nothing about which card renders faster. Use matched benchmarks instead of core counts.
Do Vendor Performance Claims Count As Benchmarks?
No. Figures such as ray-tracing TFLOPS and AI TOPS come from demonstrations the vendor ran on its own hardware and settings. They help you understand what a card is built to accelerate, not how many frames it produces in your library. Only independent testing with matched settings gives a fair comparison.
Sources
- NVIDIA. “GeForce RTX 5090 Graphics Card.” Source for Blackwell architecture, CUDA core count, ray-tracing and AI figures, clocks, memory, and power requirements.
- AMD. “AMD Radeon RX 9070 XT.” Source for RDNA 4, compute units, ray and AI accelerators, clocks, memory, board power, and operating system support.
- NVIDIA. “GeForce RTX 50 Series.” Source for DLSS 4 and multi-frame generation as RTX 50-series features.