The nephew says you need an RTX 5070 Ti and the store has an RTX 5060 Super and a 4070 for less money, and the salesperson explains that one has 12 gigabytes and the other has 8 but the 8 is faster memory. None of this is a language anyone taught you, and the price difference is four hundred dollars. The name on the box is actually a very precise code. Once you can read it, the whole shelf sorts itself out in about a minute.
Reading the name, left to right
Take RTX 5070 Ti and break it into four pieces.
RTX vs GTX. RTX cards have dedicated hardware for ray tracing and for AI upscaling. GTX is the older line without it. If you are buying new, everything is RTX; GTX only matters when someone is selling you a used card.
The first one or two digits — the generation. 50 is the current family, 40 the one before, 30 before that. A higher generation brings better efficiency and newer features, not automatically more raw speed.
The last two digits — the tier within that generation. This is the part that tells you where the card sits:
- 50 or 60 — entry level. Fine at 1080p.
- 70 — the sensible middle. Comfortable at 1440p.
- 80 — high end. 4K.
- 90 — the halo card. Priced accordingly, and bought mostly by people doing professional work.
The suffix. Ti and Super both mean “the same tier, tuned up” — more cores, sometimes more memory. A 5070 Ti sits between a 5070 and a 5080. No suffix is the base version.
The comparison that confuses everyone
A current-generation mid card against a previous-generation high card. Say a 5060 versus a 4070.
The rule of thumb that holds up: one tier up in the previous generation is roughly equal to the tier below it in the current one. A 4070 and a 5060 Ti land in similar territory. The newer card wins on power draw and on features; the older one often wins on price per frame.
So the older card is frequently the better buy, provided it has enough memory. Which brings us back to the only number that reliably predicts whether you will be annoyed in three years.
VRAM, and why it is the real specification
Video memory holds textures, models and — increasingly — AI models. When a card runs out, performance does not degrade gracefully. It falls off a cliff, because the system starts swapping data over a much slower connection.
- 8 GB — the current floor. Fine at 1080p today, already marginal for new titles at high texture settings.
- 12 GB — the comfortable middle. 1440p, and enough for small local AI models.
- 16 GB — buys you years. 4K, and the point where serious local AI becomes practical.
- 24 GB and up — professional work, large AI models, video production.
Memory type matters less than quantity for most people. GDDR7 is faster than GDDR6, but more slow memory beats less fast memory once you cross the limit.
DLSS, and why it changes the maths
DLSS renders the game at a lower resolution and uses AI to reconstruct it at your screen’s resolution. In practice it is a large free performance gain with a quality cost most people cannot see in motion.
It matters for buying because a cheaper RTX card with DLSS can outperform a more expensive card without it, in the games that support it. Frame generation — on the 40 and 50 series — goes further by inventing intermediate frames. It looks excellent when the base frame rate is already decent, and it feels strange when the base rate is low.
What to actually buy
- 1080p, normal games: current-generation 60-class, 8 GB. The cheapest sensible new card.
- 1440p: 70-class, 12 GB minimum. This is the sweet spot for most people and most monitors.
- 4K: 80-class, 16 GB.
- Local AI as well as games: go by memory, not tier. 16 GB is the threshold where useful models fit.
- Just a desktop that drives two monitors: you do not need any of this. Integrated graphics on a modern processor handles it.
The three things people get wrong
- The power supply. A card has a recommended PSU wattage and a required connector. A 5070-class card wants 650 watts and the newer 12-pin connector; plugging it in through three daisy-chained adapters is how systems shut down under load.
- Physical size. Modern cards are long and three slots thick. Measure the case before ordering. “It did not fit” is one of the most common returns.
- The rest of the machine. A fast card attached to a ten-year-old processor is throttled by the processor. If the rest of the system is old, the card is not the upgrade.
And the thing nobody says out loud: a card two tiers down from the one you were talked into will run everything you actually play, and the difference will be invisible unless you are looking for it with a frame counter on screen.