Month: 2025-08 — Concise monthly summary focusing on performance optimization work in matter-labs/zksync-airbender. Key feature delivered: - NTT kernel performance optimization for nvcc 13.0 on CC 8.9 GPUs, targeting improved throughput of Number Theoretic Transform computations. This work reduces register spilling and tunes loop unrolling pragmas to prevent performance degradation caused by nvcc 13.0 on high-end GPUs. Commit associated: 0c3fb49338f869a614cee84b7075e3f8ffdac847. Major bugs fixed: - No major bug fixes recorded in the provided data for this month. Overall impact and accomplishments: - Deliverable contributes to faster NTT computations on compute capability 8.9 hardware, strengthening the performance of the zk proof pipeline and overall prover throughput. - Aligns with ongoing optimization strategy for GPU kernels in zksync-airbender, supporting scale and reliability under nvcc 13.0 toolchain. Technologies/skills demonstrated: - GPU kernel tuning (NTT) and memory traffic optimization - nvcc 13.0 compatibility considerations and register spilling mitigation - Performance profiling and hardware-specific optimization for CC 8.9" , "key_achievements": [ "NTT kernel performance optimization for nvcc 13.0 on CC 8.9 GPUs (commit 0c3fb49338f869a614cee84b7075e3f8ffdac847).", "Reduced register spilling and tuned loop unrolling pragmas to counter nvcc 13.0-related performance regression.", "Improved speed of NTT computations on compute capability 8.9 hardware, contributing to faster zk-prover throughput." ]}"} ``` But we must ensure valid JSON. The tool accepts summary:string and key_achievements:string[]. We must escape newline? It's a string. The above has quotes and newlines; it's acceptable as JSON string. The content includes triple quotes; ensure not include extraneous characters. Let's craft clean JSON again. The summary field is a plain string; we shouldn't include stray backticks or extra code blocks. Let's craft a clean object. We'll provide summary with line breaks; it's fine as string. Ensure the JSON is valid. We'll recompose carefully. Also the

Monthly performance summary for 2025-07 focusing on feature delivery, reliability, and efficiency across matter-labs/zksync-os and matter-labs/zksync-airbender. Key outcomes include documentation updates for Ethereum proving (evm_replay requirement) to ensure accurate EVM block simulation, GPU prover performance optimizations, and stability improvements on H100 GPUs to enhance reliability of driver kernel loading. What was delivered: - Documented that the evm_replay feature is required for accurate EVM block simulation and updated proving_ethereum.md accordingly. Commit: a1075bdfc135b39bbbf5627be8e0e1b23efd5bd7. Repo: matter-labs/zksync-os. - GPU Prover Performance Enhancements: Optimized deep_quotient_kernel to reduce register spills and refactored term processing; introduced adaptive parallelism for batch reduce to improve occupancy and execution time. Commits: 74f1890a3473cc0d3a1eaf39bc20813095006394; 3fdbdd4140d5bf7f344d0d7bc83f47a18385b4a5. Repo: matter-labs/zksync-airbender. - Driver Kernel Loading Stability on H100: Increased device slack memory from 32 to 40 to accommodate memory requirements and improve stability, addressing driver kernel loading failures. Commit: d4fc8163d0d6934323af3bcef2b1bafa9064865d. Repo: matter-labs/zksync-airbender. Overall impact: - Increased correctness and confidence in EVM simulation, enabling more reliable test coverage and fewer edge-case regressions. - Substantial performance gains in GPU proving, leading to faster proof generation and better throughput for large-scale deployments. - Improved hardware reliability on cutting-edge H100 GPUs, reducing downtime and facilitating broader hardware support. Technologies/skills demonstrated: - Documentation discipline and change management; git-based traceability; cross-repo coordination. - GPU kernel optimization, memory management, and performance tuning; occupancy and parallelism strategies. - Stability tuning for driver interactions on high-end GPUs; attention to hardware-specific constraints. Business value summary: - By delivering clarity in EVM proving requirements, boosting prove-times via GPU optimizations, and stabilizing driver behavior on H100, the team reduced risk, improved throughput, and supported scalable deployment of zk proofs in production.

Month: 2025-06 Key features delivered: - GPU NTT performance optimizations: Optimize NTT computations on the GPU by leveraging L2 cache persistence and a 2-stream ping-pong approach to minimize tail effects and improve production NTT sizes. Includes overlapping kernel executions and managing working sets within the GPU's L2 cache. Impact: higher throughput for large proofs with reduced variance. - Fibonacci benchmarks aligned with zkvm-perf: Update Fibonacci benchmarks to align with the zkvm-perf benchmark, introducing larger input sizes and clearer documentation. Updates CI workflows, README, and core Fibonacci logic to improve reproducibility and benchmarking. - CLI enhancement: show total GPU-proof time: Accumulate and print the total time spent on GPU runs (base and recursion layers) via a total_proof_time metric for end-to-end performance visibility. - Documentation improvements: CUDA usage and Ethereum block proving examples: Improve documentation to clarify CUDA_LAUNCH_BLOCKING usage for reproducibility and memory footprint considerations, including precise commands and output locations; and update Ethereum block proving examples with corrected paths and a more representative default block for RISC-V simulations. Major bugs fixed: - Deterministic delegation data ordering: Fix an assertion failure arising from non-deterministic ordering when unpacking delegation data from HashMaps by sorting the keys before unpacking to ensure consistent ordering by ID for large proofs. - GPU prover memory synchronization to prevent race conditions: Address a race condition in the GPU prover's memory commitment process by adding a synchronization step to ensure the execution stream is synchronized before proceeding, improving stability on fast GPUs and large proofs. - NTT kernel register spilling fixes across CUDA versions: Address register spilling issues in NTT kernels across CUDA 12.8 on GPUs 5090 and L4, adjusting unroll directives to equalize performance between bitrev-to-nonbitrev and nonbitrev-to-bitrev transforms. Includes architecture-specific unroll tuning for stability and performance. Overall impact and accomplishments: - End-to-end performance visibility and production readiness: The combined GPU-accelerated optimizations, stable memory synchronization, and robust benchmarking updates deliver higher throughput, more reliable large-proof generation, and clearer performance metrics for optimization cycles. - Consistency and reproducibility: Deterministic data handling and aligned benchmarks reduce variability, enabling more predictable performance tuning and CI-based validation. Technologies/skills demonstrated: - CUDA and GPU kernel optimization (L2 cache-aware, multi-streaming, overlapping kernels) - Memory synchronization and race-condition mitigation in GPU streams - Deterministic data handling for large proofs - Benchmark engineering and CI workflow enhancements - Documentation and developer experience improvements