Graphics card spec sheets are unusually easy to misread, because the numbers printed largest — core counts and boost clocks — are only comparable within a single generation, while the number that most often decides whether a game runs well is printed small. Here is what each specification is actually telling you.
CUDA cores, clocks and why they only compare within a generation
A CUDA core is a small parallel arithmetic unit; a modern GPU has thousands running the same shader program over different pixels. Within one generation, core count scales almost linearly with performance — the RTX 5070’s 6,144 cores really are about a third faster than the 5060 Ti’s 4,608. Across generations the comparison collapses, because each generation changes what a core does per clock. A Blackwell core is not an Ada core, so counting them against an RTX 40-series card tells you very little.
Memory: GDDR7, bus width, and the bandwidth that decides 4K
Bandwidth is bus width ÷ 8 × memory speed, and it is the specification that separates the tiers on this page. A 128-bit bus with 28 Gbps GDDR7 gives 448 GB/s; the same memory on a 256-bit bus gives 896 GB/s. That is why the RTX 5070 Ti handles 4K and the RTX 5060 Ti, which has the identical memory type, does not — at 4K there are four times as many pixels to feed, and a narrow bus starves first. GDDR7 also matters in itself: the RTX 5050 and the RX 9060 XT both use older GDDR6 and land at 320 GB/s despite the same bus width as the 448 GB/s RTX 5060.
VRAM capacity is a cliff, not a slope
Bandwidth degrades gracefully; capacity does not. While a game’s working set fits in VRAM, frame times are steady. The moment it does not, the card starts fetching textures across the PCIe bus from system memory and frame times spike into visible stutter — not a smaller average frame rate, but a juddering one. This is why an 8GB and a 16GB card built on identical silicon can differ by 40% in ray-traced 1440p, and why the capacity number deserves more weight than its position on the box suggests.
Ray tracing, RT cores and what it costs
Rasterisation approximates light with decades of clever shortcuts. Ray tracing simulates individual light paths, which is enormously more expensive — enabling it can halve frame rates even on dedicated hardware. RT cores are fixed-function units that accelerate the specific geometric test of whether a ray hits a triangle. NVIDIA is on its fourth generation of them and holds a roughly one-tier lead over RDNA 4, which is the strongest technical argument for GeForce at these prices.
DLSS 4, Multi Frame Generation, and the latency question
DLSS renders at a lower internal resolution and reconstructs the full-resolution image with a neural network; at 1440p quality mode the output is frequently indistinguishable from native. Multi Frame Generation goes further and synthesises up to three intermediate frames between two rendered ones. That multiplies the frame counter but not your input response — the game still samples your mouse at the underlying rate, and generated frames add a little latency. It is transformative for single-player AAA at 4K and the wrong tool for competitive shooters.
Board power, transients and the 12V-2x6 connector
The rated TGP is an average. Real cards draw brief transient spikes well above it — a 360W RTX 5080 can pull over 500W for microseconds, which is why an underspecified or ageing supply trips its protection and the machine reboots under load rather than failing gracefully. The 12V-2x6 connector (the revised 12VHPWR) carries up to 600W on one cable; seat it until it clicks and avoid tight bends within 35mm of the plug.
GIGABYTE
