gpu performance engineers

Abstract: Autoregressive large language model inference repeatedly invokes the target model to generate one token at a time, making generation sensitive to GPU memory movement and sequential execution. This study evaluates two-token multi-token prediction (MTP) against autoregressive decoding in a controlled single-request deployment on an NVIDIA A10G GPU. A 360-request benchmark covered plain-text, reasoning-intensive, and tool-calling workloads, while runtime telemetry, Nsight Systems, PyTorch Profiler, and selected Nsight Compute measurements were used to explain the observed performance. MTP increased output throughput by \(1.91\times\) to \(2.19\times\) across all prompts and reduced time to first output by 10.0--14.2\%. Median mean acceptance length ranged from 2.370 to 2.595 tokens per verification iteration. Profiling showed that MTP introduced a longer and more complex execution path, including proposal, sampling, attention, gathering, and reduction operations. However, it required 56.4--78.1\% fewer executions of the selected repeating CUDA Graph per generated token. The dominant MTP GEMM kernel was not faster than the dominant autoregressive GEMV kernel, and selected instances of both approached the A10G memory-bandwidth limit. These results show that MTP improved inference through amortization: greater token progress reduced repeated GPU execution sufficiently to outweigh the additional speculative-execution cost.

Abstract: Long-context LLM inference pushes modern GPU serving stacks into an attention-bound regime, where both compute and memory are dominated by the softmax-GEMM pipeline. On NVIDIA Blackwell GPUs, FP4 Tensor Cores offer high matmul throughput, but we find that fully FP4 attention often fails to translate this throughput into end-to-end speedups due to non-matmul costs: online quantization after softmax, tensor/shared-memory data movement, and contention on the softmax path. We present MpFA, a training-free FlashAttention kernel optimized for Blackwell. Guided by hardware characterization, MpFA uses mixed precision: NVFP4 for QK and FP8 for PV (QK4PV8). This preserves low-bit QK throughput while avoiding the conversion and scaling overheads of FP4 PV. To recover accuracy without further stressing the softmax pipeline, MpFA introduces rank-one smoothing compensation implemented as an additional Tensor Core MMA. MpFA further improves performance with a fine-grained asynchronous pipeline, tensor-memory reuse, and adaptive parallel partitioning across prefill and decode. On an NVIDIA B200 and across 16K-128K contexts, MpFA improves prefill throughput over state-of-the-art BF16/FP8 baselines and increases end-to-end output throughput by 2.81$\times$ over BF16 FA4 across Llama-3.1-8B and Qwen3-14B. Across five benchmark suites and two models, rank-one compensation recovers 62.5% of the accuracy loss with about 2.0% kernel overhead.

Abstract: LLM agents automate GPU kernel optimization by repeatedly composing candidates and measuring their duration on real GPUs. Existing systems preserve measurement fidelity by reserving a GPU for an entire agent session or benchmarking command. However, this results in poor utilization because only a small fraction of command execution requires exclusive GPU access. Sharing GPUs could recover this idle capacity, but introduces contention that compromises measurement fidelity and misdirects the agent's search. We present KREX, a runtime for concurrent kernel agent benchmarking with region-granular exclusivity. KREX lets agents mark critical regions involving timing-sensitive operations within a benchmarking command. The runtime then enforces exclusivity within marked regions and allows concurrent execution outside them, achieving high throughput while preserving measurement fidelity. To enforce in-region exclusivity, KREX blocks new competing GPU submissions and drains outstanding work before freezing sibling processes and isolating CPU cores, protecting both GPU execution and the host threads that drive measurements. To maximize off-region concurrency, KREX reuses GPU contexts in persistent context processes to avoid repeated, node-wide serialized context creation. We evaluate KREX on NVIDIA and AMD GPUs. Compared with command-granular exclusivity baselines, KREX delivers up to $3.4\times$ the benchmarking throughput with a negligible p95 timing inflation of $0.30\%$, $1.58\%$, and $3.90\%$ for kernels longer than 10 ms, 1 ms, and 0.1 ms, respectively.

Abstract: Modern GPU kernels fuse increasingly more work into a single kernel, and intra-kernel tracing has become the mainstream method to profile them. Tracing inserts probes into the kernel to record its runtime states, and the fidelity of the trace determines the efficiency of performance optimization. Unfortunately, existing tools insert probes before compilation. These tools interfere with the compiler's optimizations, so they trace a different binary from the one the GPU executes. They also add significant runtime overhead. Xtrace is the first GPU kernel tracing system with near-zero compile-time interference and minimized runtime overhead. Xtrace inserts probes directly into the compiled kernel binary. It reuses only the registers that hold dead values at the insertion address and resolves all hazards with the compiler's hazard tables. It further schedules the instruction order, register allocation, and control bits to minimize the runtime overhead the probe introduces. Xtrace supports 19 NVIDIA and AMD GPU architectures, and is publicly available for use at https://g-watch.github.io. We evaluate Xtrace on major production large language model (LLM) kernels against the state-of-the-art tracers Neutrino and IKET from NVIDIA. On H100, B300, and MI300X GPUs, Xtrace preserves 94-98% of the instructions of the kernel, while existing tools preserve only 8-48%. Xtrace adds only 0.9-2.8% overhead, while existing tools add 3.8-75.6%. Xtrace guides a coding agent to reach the same FlashAttention-3 performance with 3.9x fewer iterations than existing traces do. Thanks to our binary-level instrumentation, Xtrace also traces the faster closed-source cuDNN kernel, which guides the agent to lift the open-source FlashAttention-4 by 5.2-13.3% in throughput.

Abstract: Modern GPUs require not only SIMT-style parallelism but also software-managed concurrency between compute and data movement to reach maximum performance. Performance engineers must reason about subdividing work into the hierarchy of computation resources (threads, warps, warpgroups, blocks, clusters), and, in many cases, also must use asynchronous tensor core and memcpy instructions on different levels of the memory hierarchy (registers, tensor core accumulators, shared memory, global memory). Unlike CPUs, where out-of-order execution is managed by hardware and hidden from programmers, GPUs expose explicit instruction reordering to software through these asynchronous instructions. Well-established GPU programming languages generally offer either direct low-level control without safety guarantees (e.g., CUDA C++ inline assembly or intrinsics) or easier-to-analyze, high-level abstractions (e.g., Triton's tile-based model) that hide asynchronous instructions in the compiler backend, which may prevent performance engineers from maximizing performance by tuning critical details. We propose Exo-GPU, an imperative, low-level language that creates minimal abstraction over CUDA. Our key idea is to treat parallelism and synchronization as mere annotations on sequential code rather than as fundamental control flow primitives, enabling verification that these constructs do not alter the program semantics. The benefit is twofold: programmers can reason about code without hidden control flow or mutation, while allowing the Exo-GPU compiler to verify sequential-parallel equivalence--guaranteeing that parallel execution is functionally equivalent to its sequential interpretation. We used Exo-GPU to author GEMM kernels for the H100 GPU, using wgmma, TMA, and split-k. Our kernels achieved over 80% of theoretical peak on large problem sizes, in some cases outperforming the vendor-provided CUBLAS library.