aiqbench analysis
APEX4 addresses the hardware-specific bottleneck of W4A4 quantization by dynamically adapting kernel granularity to the ratio of Tensor Core to CUDA Core throughput, enabling pure INT4 inference on diverse GPUs without modifying the underlying vLLM stack. This approach recovers performance on high-throughput A100 instances while delivering significant speedups on consumer and data center cards, proving that W4A4 viability depends on intra-SM compute balance rather than being universally infeasible. For practitioners, this offers a drop-in solution to bypass mixed-precision fallbacks and achieve near-FP16 accuracy with substantial latency reductions across Ampere and Ada architectures.
APEX4: Efficient Pure W4A4 LLM Inference via Intra-SM Compute Rebalancing
arXiv:2606.08761v1 Announce Type: cross Abstract: W4A4 quantization promises full utilization of INT4 Tensor Cores, yet group dequantization overhead on CUDA Cores has driven existing systems to mixed-precision fallbacks. We present the first systematic study of how intra-SM compute balance governs this bottleneck. Through controlled benchmarks across four GPUs from Ampere and Ada architectures, we identify the Tensor Cores to CUDA Cores throughput ratio ($\rho$) as the primary hardware indicator: the W4A4-g128 kernel yields $2.0$--$2.5\times$ speedup on RTX~3090 ($\rho=16$) yet degrades to $0.43$--$0.47\times$ on A100 ($\rho=64$) in compute-bond scenarios, establishing W4A4 viability as platform-dependent rather than universally infeasible. Guided by this finding, we build \textbf{APEX4}, which co-designs pure INT4 GEMM kernels with $\rho$-aware granularity adaptation to mitigate the CUDA Cores dequantization bottleneck. APEX4 achieves perplexity within 0.63 of FP16 on LLaMA-2-70B and outperforms W4Ax Atom-g128 by 4.0\%--4.4\% in zero-shot accuracy. Deployed as a drop-in replacement in unmodified vLLM, it delivers up to $1.66\times$ end-to-end speedup on L40S ($\rho=8$), and $1.78\times$ on RTX~3090 ($\rho=16$), $2.09\times$ on A40 ($\rho=16$), while recovering A100 ($\rho=64$) to $1.20$--$1.40\times$ via the mixed-granularity mode.
Abstract:W4A4 quantization promises full utilization of INT4 Tensor Cores, yet group dequantization overhead on CUDA Cores has driven existing systems to mixed-precision fallbacks. We present the first systematic study of how intra-SM compute balance governs this bottleneck. Through controlled benchmarks across four GPUs from Ampere and Ada architectures, we identify the Tensor Cores to CUDA Cores throughput ratio ($\rho$) as the primary hardware indicator: the W4A4-g128 kernel yields $2.0$--$2.5\times$ speedup on RTX~3090 ($\rho=16$) yet degrades to $0.43$--$0.47\times$ on A100 ($\rho=64$) in compute-bond scenarios, establishing W4A4 viability as platform-dependent rather than universally infeasible. Guided by this finding, we build \textbf{APEX4}, which co-designs pure INT4 GEMM kernels with $\rho$-aware granularity adaptation to mitigate the CUDA Cores dequantization bottleneck. APEX4 achieves perplexity within 0.63 of FP16 on LLaMA-2-70B and outperforms W4Ax Atom-g128 by 4.0\%--4.4\% in zero-shot accuracy. Deployed as a drop-in replacement in unmodified vLLM, it delivers up to $1.66\times$ end-to-end speedup on L40S ($\rho=8$), and $1.78\times$ on RTX~3090 ($\rho=16$), $2.09\times$ on A40 ($\rho=16$), while recovering A100 ($\rho=64$) to $1.20$--$1.40\times$ via the mixed-granularity mode.
Distributed, Parallel, and Cluster Computing (cs.DC); Artificial Intelligence (cs.AI)
arXiv:2606.08761 [cs.DC]
(or arXiv:2606.08761v1 [cs.DC] for this version)
https://doi.org/10.48550/arXiv.2606.08761 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 271 words
Related coverage
DI Series: Serving Kimi-K3 MXFP4 Wide-EP Disaggregated on AMD Instinct MI300X and MI325X
AMD · Oct 8, 2026
Expenditure Horizon: Measuring Optimization Ability, with an Application to NanoGPT
METR · Jul 21, 2026
Para-Pipe: Exploiting Hierarchical Operator Parallelism of ML Computational Graphs on SoCs
arXiv · Sep 4, 2026
ROCm 7.14: TheRock Goes Production and Expands AMD’s AI Software Platform
AMD · Jul 15, 2026