Published by Sean Carr
Why We Optimize Natively: Hardware Block Compression Over Neural Upscaling
Direct empirical bake-off between open GPU block compression (BCn) and neural texture codecs (NTC).
At a Glance
Payload Reduction
6.4x (84.4%)
Payload reduced from 85.3 MiB (RGBA8 level-0+mips) to 13.34 MiB (.ktx2).
Height Map Fidelity
58.12 dB (BC4)
BC4 PSNR was 58.12 dB vs NTC low (50.46 dB) and NTC default (53.90 dB).
Normal Map Fidelity
57.77 dB (BC5)
BC5 PSNR was 57.77 dB vs NTC low (48.07 dB) and NTC default (53.30 dB).
Device VRAM Saved
960 -> 736 MiB
Engine device VRAM dropped by 224 MiB (23.3% heap savings) on identical scenes.
| Telemetry Metric | Measured Value | Context & Observability |
|---|---|---|
| Payload Reduction | 6.4x (84.4%) | Payload reduced from 85.3 MiB (RGBA8 level-0+mips) to 13.34 MiB (.ktx2). |
| Height Map Fidelity | 58.12 dB (BC4) | BC4 PSNR was 58.12 dB vs NTC low (50.46 dB) and NTC default (53.90 dB). |
| Normal Map Fidelity | 57.77 dB (BC5) | BC5 PSNR was 57.77 dB vs NTC low (48.07 dB) and NTC default (53.30 dB). |
| Device VRAM Saved | 960 -> 736 MiB | Engine device VRAM dropped by 224 MiB (23.3% heap savings) on identical scenes. |
Published by Sean Carr
View & print executive brief version →Executive Summary & Core Finding
Neural texture compression (NTC) and temporal upscaling promise massive memory reductions, but introduce proprietary SDK dependencies and neural reconstruction overhead. In direct empirical bake-offs on 2048² PBR maps, standard GPU hardware block compression achieved a 6.4x VRAM reduction while outperforming neural codecs on normal and height fidelity.
Method
Bake-off executed on four 2048² Substance-baked PBR maps (glossiness, roughness, height, normal).
Compared uncompressed RGBA8 against open Khronos KTX2 hardware block compression (BC4/BC5/BC7) and NVIDIA Neural Texture Compression (NTC) at low and default bitrates.
Reconstruction quality evaluated using mathematical Peak Signal-to-Noise Ratio (PSNR) across color, height, and normal vector distributions.
Resident VRAM measured directly via Vulkan Memory Allocator (vmaGetHeapBudgets) on actual GPU heaps.
What We Measured
Compression ratio, PSNR fidelity, and VRAM residency on 2048² PBR asset sets
Payload Reduction
6.4x (84.4%)
Payload reduced from 85.3 MiB (RGBA8 level-0+mips) to 13.34 MiB (.ktx2).
Height Map Fidelity
58.12 dB (BC4)
BC4 PSNR was 58.12 dB vs NTC low (50.46 dB) and NTC default (53.90 dB).
Normal Map Fidelity
57.77 dB (BC5)
BC5 PSNR was 57.77 dB vs NTC low (48.07 dB) and NTC default (53.30 dB).
Device VRAM Saved
960 -> 736 MiB
Engine device VRAM dropped by 224 MiB (23.3% heap savings) on identical scenes.
Experimental Notes & Observability Details
- Hardware block compression decompresses directly in GPU texture filtering silicon with zero shader compute overhead.
- Whole-island 35-texture projection (reducing asset weight from 858.6 MB down to 142.0 MB) was validated by a 9-map terrain subset dropping from 192.0 MiB to 32.0 MiB.
What We Think It Suggests
- Standard GPU hardware block compression formats (BC4/BC5) can achieve higher mathematical reconstruction fidelity on geometric maps than neural texture compression, with zero runtime SDK dependency.
- Native hardware formats eliminate proprietary driver lock-in while reducing VRAM footprints by over 80%.
Limits
- Asset scope: Evaluated on four 2048² Substance-baked PBR maps; highly complex organic or photographic texture sets may exhibit different compression trade-offs.
- Fixed resolution: Benchmarked at native 2048² texture resolution; does not measure extreme mipmap streaming under severe memory oversubscription.
Detailed Analysis & Field Observations
In the push toward modern real-time photorealism, neural texture compression (NTC) has emerged as an attractive pitch: train a neural network to compress multi-channel material textures into extreme low-bitrate latent spaces. However, neural compression requires dedicated ML inference shaders during the render pass, introduces proprietary vendor runtime dependencies, and can blur critical high-frequency geometric displacement.
On July 30, 2026, we conducted an empirical comparison between modern open GPU hardware block compression (BC4, BC5, BC7 packaged via Khronos KTX2) and neural texture codecs across a set of four 2048² Substance-baked PBR material maps.
The uncompressed RGBA8 textures consumed 85.3 MiB of VRAM (including mipmaps). Packaging the assets into native hardware block compression formats reduced the payload to 13.34 MiB, a 6.4x reduction (84.4% byte reduction). Measuring actual heap allocations via vmaGetHeapBudgets confirmed that resident device VRAM dropped from 960 MiB down to 736 MiB, freeing 224 MiB of memory on the GPU.
Crucially, mathematical PSNR analysis revealed that native block compression delivered superior fidelity on functional surface maps compared to neural compression. On the height map, BC4 achieved 58.12 dB PSNR, outperforming low-bitrate NTC (50.46 dB) and default NTC (53.90 dB). On the normal map, two-channel BC5 achieved 57.77 dB, significantly exceeding default NTC (53.30 dB).
Furthermore, hardware block compression decompresses natively in the GPU's fixed-function texture filtering hardware with zero runtime ALU or tensor compute cost. Neural decompression, by contrast, consumes valuable compute passes that compete with lighting and shadows.
In this test, native hardware block formats provide clean mathematical reconstruction without proprietary lock-in or runtime inference overhead.