标签: communication

2026年10月08日

NCCL M2N: A Layout- and Topology-Aware Collective for Distributed Tensor Resharding

2026年10月08日

T-CCL: Resource Efficient and Performant Collective Communication using Tensor Memory Accelerator

2026年10月07日

Characterizing Parallelism Strategies in LLM Inference: Fundamental Compute-Communication Trade-offs

2026年10月06日

DICE: Detailed Inter-Chiplet End-to-End PHY Modeling

2026年10月06日

3DLS: A 3D Logic-Stacked Architecture for Disaggregated LLM Serving

2026年10月06日

C2C-Explorer: Chip-to-Chip Interconnect DSE for LLM Systems

2026年10月06日

Mozart: Modularized and Efficient MoE Training on 3.5D Wafer-Scale Chiplet Architectures

2026年10月06日

pHost: Distributed Near-Optimal Datacenter Transport Over Commodity Network Fabric

2026年10月06日

Optimization of Collective Reduction Operations

2026年10月06日

RDMA over Ethernet for Distributed AI Training at Meta Scale

2026年10月06日

Near-Optimal Wafer-Scale Reduce

2026年10月06日

Optimizing the Parallelism of Communication and Computation in Distributed Training Platform

2026年10月06日

PRESERVE: Prefetch Weights and KV-Cache in Distributed LLM Serving

2026年10月06日

MPI Reduce/AllReduce Algorithms

2026年10月06日

A Cloud-Scale Characterization of Remote Procedure Calls

2026年10月06日

WSE Reduce Algorithms

2026年10月06日

FlashMoE: Fast Distributed MoE in a Single Kernel

2026年10月06日

WSE Performance Model

2026年10月06日

Distributed GEMM Algorithms

2026年10月06日

FlashMoE Kernel

2026年10月06日

Cerebras Color Mechanism

2026年10月06日

RailWave: Adaptive Spatial and Temporal Scheduling for Expert-Parallel Communication

2026年10月05日

MegaMoE Kernel (Expert Parallelism Overlap)

2026年10月05日

GPU-Initiated Communication: Dissecting Down to the Bone

2026年10月05日

MegaFlux: Skew-Resilient MoE Megakernels via Pipelined Expert Replication

2026年10月01日

Purlin: Separating Orchestration from the Datapath of Collectives

2026年9月14日

Entwine: Tiled Computation and Fine-Grained GPU Communication

2026年9月14日

SAGE: Semantic-Aware Geographic Error Recovery for AI Data Movement

2026年9月10日

Sharing a Fabric with Collective Communication: Two Storage Penalties

2026年9月08日

CIERA: Cross-Iteration Exponent Reuse for Lossless Allgather in Sharded MoE Training

2026年9月08日

REACT: Tuning Collective Patterns to Alleviate Congestion in Shared AI Clusters

2026年9月07日

BASP: Communication-Efficient Batch-Aware Sequence Parallelism for LLM Training

2026年9月07日

CREDIT: Cost-guided Reduction-reuse with Efficient DSMEM Inter-CTA Tiling

2026年9月07日

Einsummable: Automatic Multi-GPU Parallelism via Join-Agg Specs

2026年9月04日

AInfer-PD: Communication-Safe In-Place Prefill–Decode Multiplexing for Distributed MoE

2026年8月31日

Understanding the Synchronization Tax in GPU Scale-Up Domains

2026年8月26日

AI Infra Wiki

探索

标签: communication

NCCL M2N: A Layout- and Topology-Aware Collective for Distributed Tensor Resharding

T-CCL: Resource Efficient and Performant Collective Communication using Tensor Memory Accelerator

Characterizing Parallelism Strategies in LLM Inference: Fundamental Compute-Communication Trade-offs

DICE: Detailed Inter-Chiplet End-to-End PHY Modeling

3DLS: A 3D Logic-Stacked Architecture for Disaggregated LLM Serving

C2C-Explorer: Chip-to-Chip Interconnect DSE for LLM Systems

Mozart: Modularized and Efficient MoE Training on 3.5D Wafer-Scale Chiplet Architectures

pHost: Distributed Near-Optimal Datacenter Transport Over Commodity Network Fabric

Optimization of Collective Reduction Operations

RDMA over Ethernet for Distributed AI Training at Meta Scale

Near-Optimal Wafer-Scale Reduce

Optimizing the Parallelism of Communication and Computation in Distributed Training Platform

PRESERVE: Prefetch Weights and KV-Cache in Distributed LLM Serving

MPI Reduce/AllReduce Algorithms

A Cloud-Scale Characterization of Remote Procedure Calls

WSE Reduce Algorithms

FlashMoE: Fast Distributed MoE in a Single Kernel

WSE Performance Model

Distributed GEMM Algorithms

FlashMoE Kernel

Cerebras Color Mechanism

RailWave: Adaptive Spatial and Temporal Scheduling for Expert-Parallel Communication

MegaMoE Kernel (Expert Parallelism Overlap)

GPU-Initiated Communication: Dissecting Down to the Bone

MegaFlux: Skew-Resilient MoE Megakernels via Pipelined Expert Replication

Purlin: Separating Orchestration from the Datapath of Collectives

Entwine: Tiled Computation and Fine-Grained GPU Communication

SAGE: Semantic-Aware Geographic Error Recovery for AI Data Movement

Sharing a Fabric with Collective Communication: Two Storage Penalties

CIERA: Cross-Iteration Exponent Reuse for Lossless Allgather in Sharded MoE Training

REACT: Tuning Collective Patterns to Alleviate Congestion in Shared AI Clusters

BASP: Communication-Efficient Batch-Aware Sequence Parallelism for LLM Training

CREDIT: Cost-guided Reduction-reuse with Efficient DSMEM Inter-CTA Tiling

Einsummable: Automatic Multi-GPU Parallelism via Join-Agg Specs

AInfer-PD: Communication-Safe In-Place Prefill–Decode Multiplexing for Distributed MoE

Understanding the Synchronization Tax in GPU Scale-Up Domains

HCCL: Collective Communication for Meta MTIA 300