Nvidia's 300 fps in Hogwarts Legacy counts frames the headline leaves out
The 6X multiplier sits in the same sentence as the 300, so at best one frame in six was actually rendered. Independent testers measured the cost of those rendered frames and got 20 to 42 percent, which is a different number for a different question.
Hero image: official Steam store screenshot for Hogwarts Legacy, 1920x1080. It predates the DLSS 5 patch and shows no Neural Rendering.
Avalanche Software shipped the DLSS 5 patch for Hogwarts Legacy on November 5. Nvidia's announcement puts the result at 4K and "up to 300 frames per second" on an RTX 5090, with ray tracing, Ray Reconstruction, Reflex and 6X Multi Frame Generation all switched on. The same week, DSOGaming ran the patch on its own RTX 5090 and reported a 20 percent drop in Hogsmeade and 42 percent in the opening scene. GameGPU, on an RTX 5080, measured 20 to 38 percent across three scenes.
Both sets of numbers hold up. They answer different questions, and only one of them describes how fast the card draws.
What the 300 counts
Nvidia's test bed is spelled out in the post: "Activating DLSS 3D-Guided Neural Rendering, 6X Multi Frame Generation, Ray Reconstruction, NVIDIA Reflex, and ray tracing, sees max setting Hogwarts Legacy run at 4K at up to 300 frames per second on the GeForce RTX 5090." Frame generation at six times is not an inference from a screenshot. It is in the sentence that carries the 300.
That changes what the number measures. Nvidia describes the multiplier elsewhere in the same post as producing "up to five additional frames per rendered frame", so at full strength one frame out of every six on screen came out of the game's renderer and the other five were interpolated. Divide 300 by six and the GPU was drawing roughly 50 frames per second to feed the pipeline. The 1440p claim, 370 fps, divides to about 62, which is the right direction for a lower resolution and a useful sign that the six-times framing is at least internally consistent.
Fifty is a ceiling on the render side, not a reading. Both halves of the claim say "up to", and Dynamic Multi Frame Generation changes the multiplier as the workload moves, so the ratio was probably not six for the entire capture. Where the multiplier steps down, the rendered rate has to climb to hold the total. What survives that caveat is narrower and still worth stating: 300 cannot be a count of frames the 5090 drew. At best one in six did.
Why the two camps never meet
DSOGaming and GameGPU both flipped Neural Rendering on and off and compared the two runs, and both kept the multiplier out of the comparison. DSOGaming reports DLSS Quality at 4K and does not mention frame generation at all. GameGPU calls its numbers "base performance". Either way the percentages price the neural pass in frames the renderer produced, which is why they look like bad news next to a 300 fps press line and are not in conflict with it.
One cross-check makes the split concrete. GameGPU's RTX 5080 renders Hogsmeade at 39 fps at 4K with Neural Rendering active. Gamersky put the same card at the same resolution through the same patch with 6X Multi Frame Generation also enabled and posted an average of 208 fps. Thirty-nine times six is 234, and 208 sits within reach of that. Gamersky's page carries no author and credits its material to "the internet", which reads like a vendor-supplied test, and the scenes and upscaling modes differ from GameGPU's, so treat this as an order-of-magnitude check and not a confirmation. It still shows how one card posts 39 and 208 in the same week with no one lying: one number counted frames leaving the renderer, the other counted frames arriving at the display.
The bottleneck sets the percentage
DSOGaming chose its two scenes deliberately and said which was which. Hogsmeade is CPU-bound, the opening is GPU-bound, and the two drops differ by more than a factor of two. Converting to frame time shows the mechanism. At 123 fps the budget per frame is 8.1 ms; at 98 it is 10.2 ms, so the pass added about 2.1 ms. In the opening, 173 fps is a 5.8 ms budget and 100 fps is 10.0 ms, an addition of 4.2 ms.
A CPU-bound scene leaves the GPU idle for part of every frame, and a GPU pass can occupy some of that gap without moving the clock. A GPU-bound scene has no such gap, so the whole cost lands on the frame. That explains the direction of the difference. It does not explain the size, because the added cost itself doubles between two scenes in one game, and nothing published says which pipeline stage absorbs the difference.
GameGPU's numbers make the picture messier. Its largest relative drop, 38 percent, came in a dialogue scene running at 71 fps before the patch, and a light scene is the last place a fixed-cost GPU pass should hurt most. Its smallest, 20 percent, came in Hogsmeade, the same CPU-bound area where DSOGaming measured 20 percent on a different card. Two testers, two GPUs, the same figure in the same place, which points at the scene rather than at either benchmark rig.
So the spread is not a dispute. Twenty percent and 42 percent are the same pass billed against frame budgets with different slack in them, and any single figure quoted for DLSS 5 in this game is a figure for one scene.
The shots the launch post does not include
Nvidia's post ships two Hogwarts Legacy comparisons, both chosen by Nvidia. DSOGaming captured more angles and found the other side of the ledger. In its last two comparisons the effect changes nothing visible while still costing frames, and the floor texture reads sharper with Neural Rendering off. In the first three, grass ambient occlusion is weaker with the feature on, which inverts part of the pitch, since better AO is one of the things the technique is sold on. In Hogsmeade, where the improvement is real, white outlines appear around characters. DSOGaming notes that those outlines are the first thing its own DLSS 5 mod setups strip out, and that an official implementation exposes no control to remove them.
That does not make the lighting improvement fake. The Hogsmeade comparison in the same set is the strongest case for the feature, and it holds up. It does mean the images in a launch post are a chosen subset, which is ordinary marketing and also the reason to read them as an advertisement for the best case rather than a survey of the game.
The cost was never a surprise
Neural Rendering takes a finished frame, reads motion vectors and scene geometry, and writes a new one on the same GPU that just drew the original. It competes for the same silicon, so a frame rate drop follows from switching it on. GameGPU logged power climbing to 364 W with the die at 68 C while the feature was active, which is what a new full-frame pass looks like on a power meter.
Nvidia's account of the past seven months says the same thing from the other direction. DLSS 5 went from needing two RTX 5090s to running on one, a tenfold gain, and the October model cut its parameter count by 24 percent to roughly 118 million while trimming 4K memory use by 20 percent. A technique that required two flagship cards in spring fits on one in autumn. The cost did not vanish. It was reduced until a single card could pay it, and the 20 to 42 percent is what is left of the bill.
The comparison with Nvidia's other segmentation fight is worth one line. Multi Frame Generation was locked to Blackwell until a community mod ran it on an RTX 3090, as we covered in the DLSS frame generation report. That argument was about which cards are allowed to run the multiplier. This one is about what happens to the frames the multiplier is handed, and the two questions have separate answers.
Owners of the cards in question are also paying more for them than the launch prices suggest, a subject we tracked in the delisting story. Both test rigs here use hardware from the top tier of our Q4 table, which is the tier where a 42 percent loss is easiest to absorb and also the tier least likely to need the frames.
What we have not run
Our bench did not touch this patch. Every frame rate in this article belongs to Nvidia, DSOGaming, GameGPU or Gamersky, and each is attributed where it appears.
Reproducing the split takes an RTX 5090 or 5080, driver 621.14, the Game Ready driver Nvidia shipped on November 4, and a capture tool that records per-frame present times instead of an average, PresentMon or an equivalent. It also takes a 2x2 matrix: Neural Rendering on and off against Multi Frame Generation on and off, in the same scene at the same time of day, with the multiplier readout logged next to the frame times. Without that multiplier in the log, a presented-frame count and a rendered-frame count look identical, which is the whole confusion this article is trying to unpick.
The patch is live for every RTX 50 series desktop card and laptop, RTX Spark PCs, and GeForce NOW Ultimate. RTX 40 and 30 owners get no toggle to test either way. For anyone with a 50 series card, the option is called Neural Rendering in the graphics menu, and the case for keeping it on is best judged on the scenes you actually play, since the published spread between 20 and 42 percent is wide enough to contain both answers.
