Engineering Bounty Board
Resolve critical industry bottlenecks in distributed systems, quantitative finance, zero-knowledge proofs, and hardware acceleration. Verified solutions are rewarded directly from protocol treasury pools.
Zero-Overhead Memory Recycling for SIMD AVX-512 Ring Buffers
High-frequency trade execution gateways encounter cache invalidation spikes when ring buffers cycle over 64MB boundary segments. The objective is to design an automated memory recycling allocator with sub-50ns recycle latency under continuous saturation.
- Maximum allowed allocation latency: 45 nanoseconds (P99.9)
- Zero syscall invocations during warm execution paths
- Compatible with AMD EPYC and Intel Xeon Scalable architectures
- Full formal proof or benchmark test harness provided in Rust or C++
Recursive Plonky3 Folding Circuit for 50,000 Keccak-256 Hashes
Rollup state transitions require high-throughput Keccak-256 precompile verification. Current STARK provers require excessive RAM for witness generation. We invite submissions to implement a recursive folding scheme compressing proof generation time below 4.0 seconds on commodity 16-core CPUs.
- Peak RSS memory usage strictly under 16 GB
- Prover execution time < 4000ms for 50k Keccak leaves
- Open-source Apache-2.0 or MIT license
- Deterministic verification checkable on Ethereum Sepolia
Cross-Platform Metal & WebGPU Matrix-Vector Quantized Shader
Current browser-based LLM inference engines lose 40% memory bandwidth converting FP4/INT4 tensors. Develop a unified WGSL and Metal kernel that executes fused INT4 dequantization directly inside SIMD subgroup shared memory.
- Achieve >= 85% of theoretical peak device memory bandwidth
- Zero host-device buffer duplication
- Support both Safari Metal WebGPU and Chrome Dawn runtimes
Asynchronous UDP Leader Election for Lossy 35% Drop Edge Swarms
Design a decentralized leader election protocol that retains deterministic safety and liveness across ad-hoc satellite constellations with asymmetric packet loss reaching 35% without triggering cascade timeouts.
- Formal TLA+ or Coq specification required
- Maximum recovery window: < 350ms post leader severance
- Zero reliance on NTP centralized timestamp servers