OVERCLOCK.news
Live Feature
Feature hardware

Why a $429 graphics card edged out a $999 one in Topaz's AI video upscaler

In Puget Systems' tests of Topaz Video's Starlight Fast 2 model, the RTX 5060 Ti finished frames slightly faster than the RTX 5080. The two have the same 16GB of memory, and Puget thinks that is why.

By OVERCLOCK September 14, 2026 3 min read 6 sources
BlueskyXRedditMail

Most benchmarks confirm what the price tag already told you. The latest from Puget Systems, the workstation builder whose labs publish their test results, contains one that doesn't. Testing the Starlight AI models in Topaz Video, which upscale and restore video, Puget found that in the model called Starlight Fast 2 a GeForce RTX 5060 Ti, a card that launched at $429, edged out an RTX 5080, which launched at $999.

NVIDIA's promotional image for the GeForce RTX 5060 Ti, a black PC case with a glass side showing the graphics card and three fans, on a background of green light.
The RTX 5060 Ti, the cheapest 16GB card Puget tested. It launched at $429 and, in one Topaz model, finished frames slightly faster than the $999 RTX 5080.Image: NVIDIA

The test, published on 11 September by Puget's Peter Emery, is simple enough to follow. Take 60 frames of 1080p H.264 video at 23.976 frames per second, about two and a half seconds, upscale it 2x to 4K, export it five times back to back as ProRes 422 LT, and measure the seconds each frame takes. Every card ran in the same AMD Threadripper PRO 9965WX workstation, on Windows 11 with Topaz Video 1.6.1 and NVIDIA's 610.62 driver.

In three of the four Starlight models the results ran the way you would expect. In Starlight Mini the RTX 5090 took about 8 seconds a frame and the RTX 5080 about 13, with the RTX 5070 Ti about 10% behind the 5080; the RTX 5070 took about 20 seconds and the RTX 5060 Ti nearly 27. Starlight Sharp put the 5090 at just over a second, the 5080 and 5070 Ti in a near tie under 3 seconds, and the 5060 Ti at about 5. In Precise 2.5, the only one of the four that also runs on AMD cards, the 5090 took 5 seconds, the 5080 was 37% slower, and AMD's Radeon RX 9070 XT took just over 54 seconds, more than three times as long as the slowest NVIDIA card.

Topaz Video Starlight Mini: seconds to process one frameRTX 5090 about 8, RTX 5080 about 13, RTX 5070 about 20, RTX 5060 Ti nearly 27. The RTX 5070 Ti was about 10% slower than the RTX 5080 and is not charted because Puget gives it only as a percentage. RTX 5090 (32GB) 8 RTX 5080 (16GB) 13 RTX 5070 (12GB) 20 RTX 5060 Ti (16GB) 27 seconds per frame, lower is faster
The normal ladder. In Starlight Mini, more expensive cards were faster, in order. The RTX 5070 Ti, not shown, was about 10% behind the RTX 5080. Values are Puget's rounded figures.Source: Puget Systems Labs, 11 September 2026. 60 frames of 1080p video upscaled 2x to 4K, five exports averaged.
View as data
Topaz Video Starlight Mini: seconds to process one frame
Systemseconds per frame, lower is faster
RTX 5090 (32GB)8
RTX 5080 (16GB)13
RTX 5070 (12GB)20
RTX 5060 Ti (16GB)27

Starlight Fast 2 broke the pattern. The 5090 again took just over a second a frame. But the 5060 Ti, the 5070 Ti and the 5080 all landed between 14 and 16 seconds, and the 5060 Ti came out slightly ahead of the 5080, a card with far more computing power.

What Puget measured, seconds per frame
Model Runs on RTX 5090 RTX 5080 RTX 5070 Ti RTX 5060 Ti
Starlight MiniNVIDIA onlyabout 8about 13about 10% slower than the 5080nearly 27
Starlight SharpNVIDIA, 16GB or morejust over 1under 3under 3, nearly tied with the 5080about 5
Starlight Fast 2NVIDIA, 16GB or morejust over 114 to 1614 to 1614 to 16, slightly ahead of the 5080
Starlight Precise 2.5NVIDIA or AMD537% slower than the 5090about 20% behind the 508020% slower than the RTX 5070
Fast 2 is the odd one out. Three 16GB cards bunched between 14 and 16 seconds a frame there, with the cheapest of them slightly ahead. Puget gives some results only as percentages, and they are shown that way.Source: Puget Systems Labs, 11 September 2026. The RTX 5070 ran Mini and Precise 2.5 only; AMD's RX 9070 XT ran Precise 2.5 only.

What those three cards share is memory. Fast 2 needs at least 16GB of video memory, which is why Puget left the 12GB RTX 5070 out, and the 5060 Ti, 5070 Ti and 5080 each have exactly 16GB. The 5090 has 32GB. Puget is careful about what it concludes: "This result could point to a memory bottleneck." Its reading is that when memory is what limits the work, a fast card like the 5080 loses more of its advantage than a slower one like the 5060 Ti, because more of its computing power is left with nothing to do.

NVIDIA's product image of the GeForce RTX 5080 Founders Edition graphics card, a silver and black card with two fans, over a field of curved green lines.
The RTX 5080 has the same 16GB. In Starlight Fast 2, where every card needs at least 16GB, its extra computing power did not show.Image: NVIDIA

A bottleneck is simply the slowest step in a process, and speeding up any other step changes nothing. A graphics card can only compute on data it holds in its own memory, so if the work is limited by memory rather than by the processor, three cards with the same 16GB hit the same wall and their extra cores stop mattering. That is what Fast 2 looks like. The 5090, with twice the memory, finished a frame in just over a second against 14 to 16 seconds for the other three, which fits the theory, although it also has far more computing power, so on its own it cannot prove it. Puget has not published memory measurements, so the explanation stays a strong possibility rather than a finding.

For anyone choosing a card for AI video work, the lesson is that memory decides first what you can run at all, and then, sometimes, how fast. The 12GB RTX 5070 could not run two of the four Starlight models, and in the one most limited by memory, the cheapest 16GB card here kept pace with cards that launched at up to twice its price and more. It is one application, one clip and one lab, and Puget's own conclusion is hedged. Still, it is a good reason to check how much memory a card has before counting its cores. And at today's prices the gap is wider than the launch prices suggest: TechSpot's September tracking found the RTX 5060 Ti 16GB at about $730 in the US and the RTX 5080 at about $1,550.

Sources

Every figure in this article traces to one of the following. Where a manufacturer and an independent test disagree, both are listed and the disagreement is stated in the text rather than resolved silently.

  1. Puget Systems Labs Best Consumer GPUs For Topaz's Starlight Models Peter Emery, 11 September 2026: the test method and system, every seconds-per-frame figure and percentage in the piece, the 16GB minimum for Sharp and Fast 2, the RTX 5070's exclusion, AMD compatibility with Precise 2.5 only, and the memory bottleneck hypothesis.
  2. TweakTown GPU prices are now 28% above MSRP on average, RTX 5090 explodes to 136% above MSRP Reporting TechSpot's price tracking: launch prices of $429 for the RTX 5060 Ti 16GB and $999 for the RTX 5080, and US street prices of about $730 and $1,550 in early September 2026.
  3. NVIDIA GeForce RTX 5090 specifications Standard memory configuration of 32GB GDDR7.
  4. NVIDIA GeForce RTX 5080 specifications Standard memory configuration of 16GB GDDR7. Also the product image in figure 5.
  5. NVIDIA GeForce RTX 5070 family specifications RTX 5070 Ti with 16GB GDDR7; RTX 5070 with 12GB GDDR7.
  6. NVIDIA GeForce RTX 5060 family specifications RTX 5060 Ti in 16GB and 8GB versions; Starlight Fast 2 and Sharp need at least 16GB, so the card Puget ran in them is the 16GB version. Also the image in figure 1.

This is original work by OVERCLOCK, not a summary of another outlet's reporting. Specifications are as published by the manufacturer unless a measured figure is named as such; vendor performance claims are labelled with their precision and are not comparable across different numeric formats. Prices and availability change, and are accurate as of publication.

More hardware