DLSS 5 mod brings next-gen tech to old Ampere GPUs, but frame rates are horrible — most games tank to single digits, high-end GPUs can hit up to 40 FPS in some cases
- ID
- 19619
- Status
- summarized
- Published
- 31 Aug 2026, 12:30 AM
- Fetched
- 31 Aug 2026, 2:33 AM
- Provider
- Tom's Hardware
- Category
- technology
- Original URL
- https://www.tomshardware.com/pc-components/gpus/dlss-5-mod-brings-next-gen-tech-to-old-ampere-gpus-but-frame-rates-are-horrible-most-games-tank-to-single-digits-high-end-gpus-can-hit-up-to-40-fps-in-some-cases
- Source URL
- https://www.tomshardware.com/feeds/all
Summary
- Score
- 4.0
- Created
- 31 Aug 2026, 2:34 AM
- Tags
- Audience
- developersai_ml_learners
What happened
A modder has ported DLSS 5 next-gen upscaling tech to older Nvidia Ampere architecture GPUs (RTX 30 series), but performance is abysmal — most games drop to single-digit frame rates, with high-end Ampere cards reaching only up to 40 FPS in best cases. The mod demonstrates that DLSS 5's underlying transformer-based model is computationally heavy enough that older tensor cores struggle to run it in real time.
Why it matters
If you're running AI/ML workloads or inference on older Ampere GPUs, this signals that newer transformer-based upscaling and generation models may be impractical on that hardware without significant optimization. Developers and ML practitioners using RTX 30-series cards for local inference should factor in that next-gen model architectures may exceed their hardware's real-time compute budget.
Discussion angle
What this tells us about the compute ceiling of older tensor cores for real-time transformer inference — and whether ML practitioners should treat Ampere as end-of-life for cutting-edge model deployment.