Skip to content
Hardware reviewGood

NVIDIA GeForce RTX 5070 Ti · 2025

NVIDIA GeForce RTX 5070 Ti: a cut-down RTX 5080 at a price nobody pays

NVIDIA GeForce RTX 5070 Ti review. 8,960 Blackwell cores and 16 GB of GDDR7 for a suggested 749 dollars, a figure the shops have rarely honoured.

Voxizer EditorialSection editor Pablo Huguet10 min read

Good

Editorial score out of 10

A graphics card can be two products at once: the one on the specification sheet and the one in the shops. The GeForce RTX 5070 Ti has spent its whole life being both, and the two have almost never cost the same money.

7.8
Editorial score
#26
Of 33 hardware reviewed
2025
Released
9.0
Best at: Features

On paper it is unusually generous for its place in the range. NVIDIA did not build it from the smaller die that powers the RTX 5070; it built it from GB203, the same 378 mm² piece of silicon that runs the RTX 5080, with part of it switched off. The memory system came across untouched, and that single decision is why the card sits nearer the model above it than its name suggests.

The suggested price was 749 dollars. NVIDIA never made a Founders Edition to sell at it, and the reporting since February 2025 has largely been an account of what the partner cards charged instead.

What the silicon is

GB203-300 is a trimmed GB203. The whole die holds 45.6 billion transistors on TSMC's 4N process across 378 mm², and the RTX 5080 draws 10,752 shaders from it. The RTX 5070 Ti gets 8,960, which is a little over four fifths of the same chip, along with 70 fourth-generation ray tracing cores and 280 fifth-generation tensor cores. The clocks run to 2.30 GHz base and 2.45 GHz boost.

What comes off in a harvest matters as much as what stays. Alongside the shaders, a slice of the level-two cache goes: the full die carries 64 MB and this configuration uses 48 MB. The boost clock lands lower than the RTX 5080's as well. None of that is unusual for a second-tier part, and none of it touches the memory controllers, which is the part buyers should care about most.

The specification
ItemGeForce RTX 5070 Ti
ArchitectureBlackwell, GB203-300 on TSMC 4N
Shaders8,960 CUDA cores, 70 RT cores, 280 tensor cores
Clocks2.30 GHz base, 2.45 GHz boost
Memory16 GB GDDR7, 256-bit bus, 896 GB/s
Power300 W board power, 750 W system supply recommended
Outputs3x DisplayPort 2.1b, 1x HDMI 2.1b, PCI Express Gen 5
Launch20 February 2025, 749 dollars suggested

The comparison this card invites is therefore upwards rather than downwards. A buyer weighing it against the RTX 5070 is comparing two different chips; a buyer weighing it against the RTX 5080 is comparing two versions of one chip, separated by a fifth of the shaders, a slower boost, 60 W and a suggested price gap of 250 dollars. The second comparison is the honest one.

NVIDIA GeForce RTX 5070 Ti
NVIDIA GeForce RTX 5070 Ti極客灣Geekerwan, CC BY 3.0, via Wikimedia Commons

Memory and the power budget

Sixteen gigabytes of GDDR7 on a 256-bit bus works out at 896 GB/s. Capacity and bandwidth do separate jobs and age at different rates: bandwidth decides how fast the card can be fed today, while capacity decides which settings it can still reach in three years. Bandwidth here is plentiful and modern. Capacity is where the argument starts, because 16 GB is exactly what the RTX 4080 SUPER shipped with in January 2024, and the textures have not stood still since.

At 1440p, with or without upscaling, that is comfortable. At 4K with path tracing, high-resolution texture packs and a frame generation mode holding several frames in flight at once, 16 GB is the first thing in this card that runs out. It is not a flaw so much as a ceiling with a date on it, and the date is closer than a 2025 card should make it.

Where the budget goes

Memory capacity
7/10
16 GB is easy at 1440p and the first wall at 4K
Memory bandwidth
8/10
896 GB/s across the same 256-bit bus the RTX 5080 uses
Board power
5/10
300 W, against 360 W for the RTX 5080 and 575 W for the RTX 5090
Display output
9/10
Three DisplayPort 2.1b outputs, enough for a 4K 480 Hz panel
Supply headroom
6/10
750 W recommended; the cable on your power supply is the thing to check

The power figure is the quiet success. Three hundred watts is restrained for the performance class, and NVIDIA asks for a 750 W system supply rather than anything exotic. The box includes an adapter that takes two 8-pin PCIe cables; a modern supply with a single 300 W PCIe Gen 5 cable does the job with one lead instead of a bundle. Anyone reusing a supply bought before 2023 should count their spare connectors before ordering.

Watch
NVIDIA GeForce RTX 5070 Ti — trailer

Frames that are not rendered

The RTX 50 series launched alongside DLSS 4 and Multi Frame Generation, and it changed what a frame rate is. The card renders a frame, then inserts up to three more between that one and the next. A number produced that way is not four times the rendering work; it is one rendered frame doing the work of four displayed ones. The picture is smoother, and the responsiveness is still governed by how often a real frame arrives. Reflex trims the latency that comes with it; it does not rewrite the arithmetic.

A quoted frame rate now says as much about which switches were left on as about the card underneath.

This is why any comparison between an RTX 5070 Ti and an older card needs its settings stated before it means anything. Set against a GeForce RTX 4090 running DLSS 3, the Blackwell card can post a larger number while drawing fewer real frames per second, and a reader who does not know that will draw the wrong conclusion. The feature is genuinely good on a high refresh monitor with a solid base frame rate. It is worth much less when the base rate is already low, which is the situation it is most often advertised as fixing.

The feature set has also kept growing, which counts in the card's favour. DLSS 5 arrived in September 2026 with 3D-Guided Neural Rendering and landed on every RTX 50 card at once rather than on the expensive ones first. A card bought in 2025 does more in 2026 than it did on the day it shipped.

Gallery · 3 images極客灣Geekerwan, CC BY 3.0, via Wikimedia Commons

The price nobody pays

Seven hundred and forty-nine dollars was the suggested price, and no card was ever built to carry it. NVIDIA made reference designs for the RTX 5090, the RTX 5080 and eventually the RTX 5070, and skipped this one entirely. Every RTX 5070 Ti is a partner board with a partner's cooler, a partner's factory overclock and a partner's price, and the cheapest of those has only intermittently touched the figure in the press release.

The reported history since is not encouraging. Stock disappeared at launch, and street prices sat above the suggested number through 2025. Through early 2026, as GDDR7 was pulled towards more profitable work, several partners were reported to have placed the card at end of life, others denied any such thing, and retailers were widely reported as unable to get stock either way. Whatever the internal truth, the shelves thinned and the asking prices went up. None of this is a judgement on the silicon; it is a judgement on what buying the silicon has actually involved.

There is one comparison that settles the value question faster than any other, and it is not against a rival. It is against the card NVIDIA was replacing.

Head to head

Value
5.67.0
Design
7.88.5
Features
9.09.3
Performance
8.68.8

NVIDIA GeForce RTX 4080 SUPER edges it by 0.4. See them side by side

The RTX 4080 SUPER launched in January 2024 at 999 dollars with 16 GB on the same 256-bit bus and a 320 W rating. Thirteen months later this card arrived with the same memory capacity, a faster memory type, a lower power figure, a suggested price 250 dollars below it and a frame generation mode the older card cannot run. On the specification sheet that is a clean generational win. In a shop where neither part sells at its suggested price, it is a much harder sum, and it explains why the card's reputation and the card's value have never quite matched.

Who should buy it

This is a 1440p card with genuine 4K ability, and it is the cheapest place in the Blackwell range to get ray tracing that does not feel like a compromise. Anyone pairing it with a fast processor such as the Ryzen 7 9800X3D and a high refresh monitor is buying into the part of the range where the frame generation work pays off rather than papers over a problem.

The condition attached to all of that is the price on the day. At or near 749 dollars this is an easy recommendation. Several hundred dollars above it, the case collapses, because the money buys a meaningful step up the same die instead. The card has not changed since February 2025; the question of whether to buy one is answered by a shop window rather than by a specification sheet, and that is an unusual and slightly dispiriting thing to have to say about good hardware.

For you if

  • You play at 1440p, or at 4K with upscaling, and want ray tracing without RTX 5080 money
  • You have a 750 W supply with the right cable and a case that fits a large partner card
  • You can find one at or near its suggested price, which is the entire argument

Not for you if

  • You play at 4K with path tracing and texture packs, where 16 GB is the first wall
  • You already own an RTX 4080 or an RTX 4080 SUPER and would be paying a premium to stand still
  • You need a card this month and the only stock on the shelf carries a heavy mark-up

Everything quoted here is NVIDIA's own published specification rather than any measurement, and the manufacturer's page is where to check it against the supply and the case you intend to use.

What works

  • Built from the RTX 5080's own GB203 die rather than the smaller RTX 5070 one
  • 16 GB of GDDR7 on the full 256-bit bus, 896 GB/s of bandwidth
  • 300 W board power, with a 750 W supply enough to run it
  • Still collecting features; DLSS 5 reached every RTX 50 card in 2026

What doesn’t

  • NVIDIA never made a Founders Edition, so the 749-dollar sticker is theoretical
  • Reported street prices have spent most of its life well above that figure
  • 16 GB matches what the RTX 4080 SUPER already shipped with in 2024
GoodHow we score

The verdict

As engineering the RTX 5070 Ti is the most sensible card in the Blackwell range: a harvested RTX 5080 with the memory system left intact, a restrained 300 W budget and a feature set that has grown since launch rather than stalled. As a purchase it has been an ordeal. NVIDIA published a price and then built nothing to sell at it, partners set their own, supply of GDDR7 went elsewhere, and for long stretches the card has been either absent or expensive. The score reflects both halves. Find one near its suggested price and it is the obvious choice at that level; find one at the prices the market has mostly asked, and the money is better spent a tier up.

Quick answers

Is NVIDIA GeForce RTX 5070 Ti worth buying?

Voxizer scores NVIDIA GeForce RTX 5070 Ti (2025) 7.8 out of 10 — good. As engineering the RTX 5070 Ti is the most sensible card in the Blackwell range: a harvested RTX 5080 with the memory system left intact, a restrained 300 W budget and a feature set that has grown since launch rather than stalled. As a purchase it has been an ordeal.

Who will enjoy NVIDIA GeForce RTX 5070 Ti?

You play at 1440p, or at 4K with upscaling, and want ray tracing without RTX 5080 money. You have a 750 W supply with the right cable and a case that fits a large partner card. You can find one at or near its suggested price, which is the entire argument.

Who should skip NVIDIA GeForce RTX 5070 Ti?

You play at 4K with path tracing and texture packs, where 16 GB is the first wall. You already own an RTX 4080 or an RTX 4080 SUPER and would be paying a premium to stand still. You need a card this month and the only stock on the shelf carries a heavy mark-up.

What is NVIDIA GeForce RTX 5070 Ti like, in short?

A harvested GB203, the RTX 5080's own die, with 8,960 CUDA cores of that card's 10,752. 16 GB of GDDR7 on the full 256-bit bus, 896 GB/s, rated at 300 W of board power. No Founders Edition was ever built, so the 749-dollar suggested price had nothing selling at it. Multi Frame Generation makes a quoted frame rate a different measurement from a rendered one.

What readers say

Written by Voxizer members and checked by our moderators before they appear.

Written reviews

–

No written reviews yet

Write a review
9–10
0
7–8.5
0
5–6.5
0
3–4.5
0
0–2.5
0

Nobody has reviewed NVIDIA GeForce RTX 5070 Ti yet. Be the first.

All reviews →

Verified sources

Published by the Voxizer editorial team — David Abellán, Ignacio Pineda, Pablo Huguet. Section editor for hardware: Pablo Huguet.

Revision history

  1. Published

Scores are revised when the facts behind them change, and the change is recorded here. How we score.