RTX 40 running DLSS 5: Jensen says 15% overhead, so why are modders getting 39%?

Jensen has consistently claimed that DLSS 5 running on RTX 40 has "performance overhead as low as 15%." Sounds great—but this week the modding community dug out the leaked DLSS 5 files from NBA 2K27, hacked them apart, and threw them onto an RTX 40

💡 What You Will Learn

RTX 40 running DLSS 5: Jensen says 15% overhead, so why are modders getting 39%? Jensen has consistently claimed that DLSS 5 running on RTX 40 has "performance overhead as low as 15%." Sounds great—b

📜 Table of Contents

• These Two Numbers Are Measuring the Same Thing (Just Different Math)
• Jensen's 15%: NVIDIA's Projected Figure After Ada Optimization Is Complete
• Modder's 39%: A Half-Baked Version — No Tuning, No Room for Optimization
• Community Versions Will Keep Improving, But RTX 40 Users Don't Want to Wait

RTX 40 Running DLSS 5: Jensen Says 15% Overhead — So Why Are Modders Getting 39%?

Jensen has been claiming all along that DLSS 5 running on RTX 40 has "performance overhead as low as 15%." Sounds great — but this week, the modder community dug out the leaked DLSS 5 files from NBA 2K27, tweaked them, threw them onto an RTX 4090, and ran it. Tom's Hardware and TweakTown measured a frame rate drop of 39% (calculated by raw frame rate decline), which translates to a 58% "target frame rate loss ratio." Same dataset, two different calculations, more than double the difference.

So here's the question — is Jensen lying? Or is the version modders tested different from what NVIDIA plans to release?

These Two Numbers Are Measuring the Same Thing (Just Different Math)

Let's get the arithmetic straight first, or we can't have this conversation.

Mainland Chinese media (including Kuaikeji, NetEase, Sohu) mostly report by "raw frame rate decline": 130 FPS - 82 FPS = 48 FPS, 48 divided by 130 ≈ 37%. TweakTown and Tom's Hardware calculate by "target frame rate loss": 48 divided by 82 ≈ 58%. Both numbers are real — it depends on how you define "overhead." I used the mainland China convention, writing "dropped about 40%" in the headline version — based on 37% overhead.

But that's still 22 percentage points away from "15%." Those 22 points can't be explained by "different math" — it's two different versions.

Jensen's 15%: NVIDIA's Projected Figure After Ada Optimization Is Complete

Here's what NVIDIA officially said: "Our current priority is optimizing RTX 50 series performance through model updates, expected to roll out later this fall. After RTX 50 series performance is fully tuned, we will begin extending official support to the RTX 40 series."

Note the key words: model updates, tuning, expected. What this means:

• The DLSS 5 currently running on RTX 40 is the Blackwell-optimized version — written specifically for RTX 50's 5th-gen Tensor Cores (NVIDIA's own AI compute units);
• Ada architecture (what RTX 40 uses) only has 4th-gen Tensor Cores — a full tier behind in hardware compute;
• What NVIDIA needs to do is recompile a model for Ada — not a simple toggle, but a low-level CUDA (NVIDIA's GPU programming interface) binary (machine code) rewrite.

That "rewritten version" is the target NVIDIA dares to call "15% overhead." The version modders are running now is the Blackwell version — which NVIDIA never intended for RTX 40 — crammed in by force. It's essentially making last-gen hardware run current-gen software. Dropping 40% of your frames is being generous.

Modder's 39%: A Half-Baked Version — No Tuning, No Room for Optimization

What Uncle Burrito did isn't technically complex: he replaced the CUDA binary originally compiled for Blackwell inside the DLSS 5 dynamic link library (that .dll file) with an Ada-compatible version, letting DLSS 5's neural rendering (NVIDIA's tech for real-time image reconstruction using neural networks) run on RTX 40.

With that change, DLSS 5's core features (neural rendering, frame generation) work — but at a cost:

1. Inference path isn't optimized — Blackwell's Tensor Cores run FP8 (an 8-bit floating point precision commonly used for AI model inference) fast and with low VRAM usage; Ada running FP8 has no hardware-level acceleration and has to brute-force it with general compute, so frames get cut;
2. VRAM bandwidth is the bottleneck — The RTX 4090's memory bandwidth (GDDR6X running at 21 Gbps) is about 75% of the RTX 5090's (GDDR7 running at 28 Gbps). DLSS 5's intermediate tensors (temporary data stored during neural network computation) are large — this alone is a wall you can't get past;
3. No time for model pruning — The modder community's attitude is "if it runs, it's good enough." Nobody's spending three months doing Ada optimization for NVIDIA.

TweakTown's measured 82 FPS is already the result of 4K resolution + DLSS 5 neural rendering — max settings, frames cut by 40%. If modders were willing to spend time on Ada optimization (rewriting operators, adjusting tensor layouts, sacrificing some image quality), shaving off another 10-15 percentage points of overhead isn't impossible — but nobody's going down that path, because NVIDIA's official optimized version is coming soon.

Community Versions Will Keep Improving, But RTX 40 Users Don't Want to Wait

The most dramatic part isn't the 15% vs 39% gap — it's when the official version actually arrives. NVIDIA says "later this fall." Whether that's October or December, nobody knows.

For RTX 40 users, these two months are a vacuum:

• Wait for NVIDIA → get a stable version, but the timeline is vague;
• Run the modder version → usable tonight, but every time NVIDIA updates drivers you have to wait for modders to adapt (this DLSS-G 310.9 lockout is exactly that);
• Upgrade to RTX 50 → problem solved, but the RTX 5090 is currently at a 41% markup — MSRP (manufacturer's suggested retail price) is long gone.

Jensen's "15% overhead" this time isn't a lie — it's a pie in the sky for the future. The problem is RTX 40 users have to wait two months to eat that pie. Who's responsible for the experience during those two months? The community modders have already taken action.

That RTX 4090 of yours — will it stick around until NVIDIA finishes tuning? Or are you running the MOD tonight and enjoying it first?