The Signal

Signals from Delphixit Nova — research notes, tool releases, and the plays that let us extract high performance from GPUs, codecs, and sound stacks. Our focus: custom toolkits for Vulkan, D3D12, Metal, FFmpeg-powered media pipelines, and low-level audio for professional production.

Active Development

Architecting GToolKit: A New Vulkan-First Foundation

GToolKit

We are building GToolKit from the ground up to solve a specific problem: generic engines hide too much. By abstracting Vulkan, D3D12, and Metal into a thin, explicit layer, we gain deterministic control over memory aliasing and barrier placement. This is the architectural baseline for the upcoming alpha release.

Dev Log #001

Media Pipelines

Oct 2025

Optimizing Hardware Transcoding Pipelines

Standard FFmpeg pipelines often incur a heavy penalty moving frames from VRAM to system RAM and back. We engineered a zero-copy path using CUDA/VAAPI interop and Vulkan Video extensions. This allows us to keep 4K/8K frame buffers strictly on the GPU during the entire decode-filter-encode loop, reducing PCI bus saturation by over 60%.

Stack

FFmpegVulkan VideoDMA

Systems

Sep 2025

Deterministic Audio Scheduling

Achieving synchronization across WASAPI (Windows), CoreAudio (macOS), and ALSA (Linux) is notoriously difficult due to varying buffer callback behaviors. We implemented a custom lock-free ring buffer that abstracts these OS-specific irregularities, ensuring frame-perfect audio-visual alignment even under heavy CPU load.

Stack

PortAudioC++Real-time

Graphics Engineering

Sep 2025

Unified Shader Reflection & Layouts

Managing descriptor sets across HLSL and GLSL manually is error-prone. We built a toolchain that uses SPIRV-Cross to extract reflection data at build time, automatically generating C++ structs that match your shader layouts. This guarantees type safety between CPU host code and GPU device code, eliminating silent validation errors.

Stack

SPIR-VReflectionTools

Systems

Aug 2025

Custom Allocators for Render Graphs

General-purpose allocators (malloc/new) introduce fragmentation and lock contention in multi-threaded rendering. We switched to linear and stack-based allocators for our frame graph execution. This improved cache locality for command recording and reduced frame-time spikes caused by OS-level page faults.

Stack

C++23Memory Management

Rendering

Aug 2025

Sparse Residency & Texture Streaming

Loading 8K texture sets for open-world environments instantly exceeds VRAM limits. We implemented a virtual texturing system using Vulkan sparse binding. This allows us to stream only the visible mip-levels into memory on demand, decoupling asset quality from hardware memory constraints.

Stack

VulkanSparse Binding

Access GToolKit Source

We are opening the repository to select engineers. Leave your GitHub email to be added to the allowlist.