Let’s build a high-performance fuzzer with GPUs!

Summary: Trail of Bits built a GPU-based fuzzer that translates embedded binaries to PTX via its Remill tool and runs them across thousands of GPU threads, reaching 361K executions/second/dollar in an early benchmark against libFuzzer's 5.2M. After profiling showed the GPU using only 3% of its compute capacity, interleaving each warp's 32 threads into a shared memory slab improved performance by an order of magnitude and raised compute utilization to 33%.
What you should know: The fuzzer is still missing a mutator and trails libFuzzer by roughly an order of magnitude on cost-normalized throughput, so the 10x-per-dollar goal remains a preliminary claim rather than a measured result.

Read the original at Trail of Bits Blog ↗