T. Molom-Ochir, B. F. Morris, et al.
arXiv:2607.22869 · 2026-07-24T19:17:37Z
The paper maps the phases of Monte Carlo tree search to separate in-memory primitives, including CAM, logic, RRAM crossbars, and SRAM, rather than treating the irregular algorithm as a single crossbar workload. Using 22 nm estimates calibrated with fabricated RRAM-array parameters, it reports about 60 mW for 9×9 Go and 96× CPU energy efficiency. The claim is compelling because it tackles an irregular search algorithm, but the end-to-end system is still an architecture evaluation rather than measured silicon.
论文将蒙特卡洛树搜索的各阶段分别映射到CAM、逻辑、RRAM交叉阵列和SRAM等存内原语，而不是把这类不规则算法视为单一交叉阵列工作负载。基于采用已制备RRAM阵列参数校准的22 nm估算，其对9×9围棋报告约60 mW功耗和相对CPU 96倍的能效。该工作的重要性在于处理了不规则搜索算法，但端到端系统仍属架构评估而非实测芯片。
E.-M. Huang, A.-C. Chang, et al.
arXiv:2607.23099 · 2026-07-25T08:20:47Z
DA-MoE chooses fused-MoE GPU kernels from the live expert-routing histogram instead of only from token-count buckets. On HumanEval-X serving traces, the authors report geometric-mean latency gains of 1.16× for DeepSeek-V3 and 1.29× for Kimi K2, with peaks of 1.40× and 1.56×. It identifies routing skew as a first-class systems variable, although the benefit depends on representative offline tuning and the serving mix matching those distributions.
DA-MoE根据实时专家路由直方图选择融合MoE GPU内核，而非仅依据token数量分桶。在HumanEval-X服务轨迹上，作者报告DeepSeek-V3的几何平均延迟改善为1.16倍、Kimi K2为1.29倍，峰值分别达1.40倍和1.56倍。该工作将路由偏斜提升为一等系统变量，但收益依赖有代表性的离线调优及实际服务负载与这些分布的匹配度。
O. Mohite
arXiv:2607.22785 · 2026-07-24T11:42:17Z
FusionML shows that MLX lazy-graph scheduling can serialize CPU–GPU work unless an eager materialization boundary is inserted, then uses contention-aware row splitting for transformer prefill. Across five Apple chips, it reports 1.15–1.38× faster decoder-block prefill and 1.18–1.25× faster time to first token for Qwen2.5-7B, with unchanged decode throughput. The useful result is that shared memory alone does not guarantee operator-level parallelism.
FusionML指出，若不插入即时物化边界，MLX的惰性图调度会将CPU-GPU工作串行化；随后它为Transformer prefill采用了考虑争用的行切分。在五款Apple芯片上，作者报告解码器块prefill加速1.15至1.38倍，Qwen2.5-7B的首token时间加速1.18至1.25倍，而decode吞吐不变。其关键结论是：统一内存本身并不保证算子级并行。