Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.28671

Accelerating CPU Speculative Decoding via Unified Graph Execution and Transactional KV Isolation

Zhangdie Yuan
View original source ↗Read source document ↗Export metadata ↓

Abstract

Speculative decoding accelerates LLM inference, but deploying it on CPUs introduces severe systems bottlenecks: context-swapping latency between draft and target graphs, and memory allocation overhead during KV cache rollbacks. We introduce SubTernary, an execution framework that eliminates these costs through Unified Graph Caching and Transactional KV-Cache Isolation. By compiling both models into a single memory arena with a persistent thread pool and routing speculative tokens to a static scratchpad for zero-copy pointer resets, SubTernary avoids dynamic reallocation and thread synchronization entirely. We integrate a 256-bit AVX2 ternary SIMD kernel as a low-bit case study. Ablations reveal our systems runtime drives 35.0% of the throughput gains, while the ternary kernel adds 6.0%. SubTernary achieves a $1.35\times$ speedup over llama.cpp speculative decoding and $3.99\times$ over autoregressive decoding, maintaining $O(1)$ rollback latency.

Explore related subjects

cs.AR

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Zhangdie Yuan. 2026-10-17. Accelerating CPU Speculative Decoding via Unified Graph Execution and Transactional KV Isolation. https://arxiv.org/abs/2610.28671

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

TensorPool: A 3D-Stacked 8.4TFLOPS/4.3W Many-Core Domain-Specific Processor for AI-Native Radio Access Networks

The upcoming integration of AI in the physical layer (PHY) of 6G radio access networks (RAN) will enable a higher quality of service in challenging transmission scenarios. However, deeply optimized AI-Native PHY models impose higher computational complexity compared to conventional baseband, challenging deployment under the sub-msec real-time constraints typical of modern PHYs. Additionally, following the extension to terahertz carriers, the upcoming densification of 6G cell-sites further limits the power consumption of base stations, constraining the budget available for compute ($\leq$ 100W). The desired flexibility to ensure long term sustainability and the imperative energy-efficiency gains on the high-throughput tensor computations dominating AI-Native PHYs can be achieved by domain-specialization of many-core programmable baseband processors. Following the domain-specialization strategy, we present TensorPool, a cluster of 256 RISCV32IMAF programmable cores, accelerated by 16 256 MACs/cycle (FP16) tensor engines with low-latency access to 4MiB of L1 scratchpad for maximal data-reuse. Implemented in TSMC's N7, TensorPool achieves 3643 MACs/cycle (89% tensor-unit utilization) on tensor operations for AI-RAN, 6$\times$ more than a core-only cluster without tensor acceleration, while simultaneously improving GOPS/W/mm$^2$ efficiency by 9.1$\times$. Further, we show that 3D-stacking the computing blocks of TensorPool to better unfold the tensor engines to L1-memory routing provides 2.32$\times$ footprint improvement with no frequency degradation, compared to a 2D implementation.

cs.AR↗
arXiv2026

M3D-FLOW: A Monolithic 3-D-Stackable Processing-Near-DRAM Architecture for Generative Normalizing Flow Applications

Normalizing Flows (NFs) have shown strong results in generative modeling, recently reaching state-of-the-art image generation performance. However, their repetitive, sequential operations limit the computational and memory efficiency achievable with GPUs and TPUs. We propose M3D-FLOW, a novel monolithic 3D-stackable (M3D) DRAM-based processing-near-memory (PNM) architecture that eliminates memory bottlenecks using a Cu-Cu hybrid-bonded logic die integrated with an M3D DRAM die. The M3D DRAM provides dense bank integration and a tiered organization with variable access latencies that M3D-FLOW exploits for the efficient processing of different NF workloads. To maximize compute density while minimizing area, energy, and thermal overhead, M3D-FLOW employs hybrid binary--stochastic floating-point multipliers, performing mantissa multiplication in the stochastic domain and exponent operations in the binary domain. A pipelined execution strategy maps sequential NF stages across dynamically partitioned PE groups, and together with the stochastic hardware, enables M3D-FLOW to achieve at least $6.6\times$ higher performance and $3\times$ better energy efficiency than prior PNM accelerators, with only a 2.2% bits-per-dimension (BPD) increase over bfloat16 and 4.1% over FP32. An analysis of stochastic hardware variants further shows that shorter stochastic bitstreams reduce execution time by up to $3.3\times$ at the cost of higher approximation error, providing a concrete design guide for latency--accuracy trade-offs.

cs.AR↗
arXiv2026

Reorienting the Vector Axis: Restructuring FlashAttention for Long-Vector Processors

A Transformer is built around attention, whose cost grows with the square of the sequence length and dominates the layer on long sequences. FlashAttention made that cost manageable by tiling the attention matrix and updating the softmax as it goes, but almost all work since has targeted GPUs. On long-vector processors one vector register can hold up to 512 single-precision numbers. The savings of a vector instruction are proportional to how much of the register it uses, so the choice of which loop to vectorize is a first-order decision. We port PyTorch's CPU FlashAttention kernel to a RISC-V CPU with 2KB vector registers and find that it reaches 16.2% of peak floating-point throughput. The kernel walks each tile one query row at a time: the product against V is vectorized along the 64-element head dimension, the softmax's per-row maximum and sum are reductions that collapse a full register to a scalar, and the software exponential's polynomial coefficients compete with tile data for vector registers and cause spilling. We propose two techniques. The first transposes the score tile and the running output so that queries lie along the contiguous axis. The algorithm and its values are unchanged. A single parameter, the number of queries in a tile, now sets the vector length of every part of the kernel and can be matched to the machine. Both reductions also disappear. The second moves the polynomial's coefficients into scalar registers, removing the spills and cutting the softmax's vector memory instructions by 78.7%. On silicon the two together cut whole-kernel cycles by 1.84 to 2.08x at a head dimension of 64, raising floating-point utilization from 16.2% to 29.7% of peak. An FPGA of the same design ranks the changes identically and scores them higher, at 2.71x.

cs.AR↗