Nvidia Optimizes DLSS 4, Reduces VRAM Usage by 20% for Better Performance

Nvidia's latest DLSS update, DLSS 4, has been officially moved out of beta, and the company has introduced significant optimizations to reduce VRAM usage. Initially reported by VideoCardz, the new DLSS SDK version 310.3.0 improves VRAM efficiency by 20% for the Transformer model, bringing its overall memory impact closer to that of the earlier CNN model. Previously, the DLSS Transformer model required nearly twice the VRAM compared to the CNN model it replaced. With these recent optimizations, the Transformer model now uses just 40% more VRAM than the CNN model. At 1080p, the new Transformer model consumes 85.77 MB of VRAM, down from 106.9 MB in the old model. For context, the CNN model consumes 60.83 MB at the same resolution. The memory usage scales linearly with resolution, so at 4K, the Transformer model's VRAM requirement has dropped from 387.21 MB to 307.37 MB, a reduction of 80 MB. | Resolution | New Transformer Model | Old Transformer Model | CNN Model | |------------|---------------------|---------------------|-----------| | 1080p | 85.77 MB | 106.9 MB | 60.83 MB | | 1440p | 143.54 MB | 181.11 MB | 97.79 MB | | 4K | 307.37 MB | 387.21 MB | 199.65 MB | | 8K | 1,225.17 MB (1.2 GB) | 1,517.60 MB (1.5 GB) | 778.3 MB | While a 20% reduction in VRAM might seem minor, especially considering modern graphics cards come equipped with several gigabytes of memory, the improvement is notable. Users are unlikely to see a significant difference in real-world applications, except possibly at very high resolutions like 8K, where the Transformer model can consume over a gigabyte of VRAM even after optimization. Nvidia's DLSS 4 features a Transformer upscaling model with twice the number of parameters compared to its CNN-based predecessor. This increased complexity allows DLSS 4 to achieve near-native image quality, making the technological advancements worthwhile despite the initial higher VRAM requirements. It's worth noting that the VRAM improvements specifically target the upscaling model and do not affect frame generation. DLSS 4 already offers major advancements in frame generation, reducing VRAM usage by 30% compared to DLSS 3. For example, in the game Warhammer 40,000: Darktide, DLSS 4's frame generation cuts VRAM consumption by 400 MB at 4K resolution, compared to DLSS 3. These optimizations demonstrate Nvidia's commitment to enhancing the performance and efficiency of DLSS, ensuring that it remains a leading technology in the field of AI-driven upscaling and frame generation. Future updates may further refine the VRAM footprint, as seen with previous versions of DLSS, such as 3.8.10, which also saw VRAM reductions. For the latest news, analysis, and reviews on DLSS and other Nvidia advancements, follow Tom's Hardware on Google News and make sure to hit the Follow button.