remote direct memory access
PulseAugur coverage of remote direct memory access — every cluster mentioning remote direct memory access across labs, papers, and developer communities, ranked by signal.
1 day(s) with sentiment data
-
LLM inference efficiency explored via quantization, memory, parallelism, and scheduling
Researchers are exploring various methods to enhance the efficiency and scalability of Large Language Model (LLM) inference. One approach focuses on energy efficiency through quantization techniques, demonstrating signi…
-
ByteDance deploys RDMA-based ByteExpress for faster AI training and inference
ByteDance has developed ByteExpress, a software communication layer built on top of hardware RDMA, to improve the efficiency of large-scale inference and training workloads. This system addresses challenges like variabl…
-
GPU compute leasing costs cut by optimizing data transfer efficiency
This article discusses optimizing data transfer efficiency in GPU compute leasing, a critical factor for reducing costs and improving performance in AI workloads. It highlights that GPU compute is often billed by the ho…
-
Mac user plans DIY AI cluster using RDMA and Thunderbolt 5
A user is planning to build a local AI cluster using multiple Mac Minis and Mac Studios connected via Thunderbolt 5. This setup will leverage Remote Direct Memory Access (RDMA) to allow these machines to pool their memo…
-
Meta AI unveils MetaRoCE for AI-scale Ethernet networking
Meta AI has developed MetaRoCE, a new transport protocol designed to improve the efficiency of AI training on large-scale Ethernet networks. Unlike traditional RoCE, which relies on lossless networks, MetaRoCE treats th…
-
New research optimizes disaggregated LLM inference with topology-aware data movement
A new research paper introduces a topology-aware data movement system designed to optimize disaggregated LLM inference. The system addresses the challenge of transferring KV caches between separate GPU pools by discover…
-
China unveils domestic GPU-RDMA direct connect for AI computing
At the 2026 World Artificial Intelligence Conference (WAIC 2026), Qimor (奇异摩尔) unveiled its full-stack, domestically produced ultra-node interconnection solution. This solution aims to address the "single point strong, …
-
STAGE framework synthesizes LLM execution graphs for distributed workloads · 2 sources tracked
A new framework called STAGE has been developed to synthesize high-fidelity execution graphs for large language models (LLMs) and Mixture-of-Experts (MoEs). This framework aims to optimize distributed AI workloads by mo…
-
USB4 RDMA implementation could boost local LLM performance
An experimental implementation of Remote Direct Memory Access (RDMA) over USB4 has been demonstrated, potentially enabling high-speed data transfer between devices connected via USB4. This development, detailed in a blo…
-
Modal launches ultra-low-latency servers for high-performance applications
Modal has introduced a new feature called Modal Servers, designed to provide ultra-low-latency server hosting for applications requiring high performance, such as LLM inference for interactive agents. This new offering …
-
OpenURMA: Open-Source Implementation of Huawei's Unified Bus Protocol Released
Researchers have developed OpenURMA, an open-source implementation of Huawei's Unified Bus (UB) protocol, designed to improve datacenter network performance. This implementation aims to address bottlenecks in modern RDM…
-
T-Head unveils Panmai 920 smartNIC, completing its AI infrastructure chip lineup
Pingtan, a subsidiary of Alibaba, has launched its first intelligent network card, the "Panmai 920," designed to address bottlenecks in AI computing infrastructure. This new network card utilizes advanced PCIe 5.0 and 1…
-
NVIDIA and Siemens Healthineers develop AI for adaptive ultrasound imaging
NVIDIA and Siemens Healthineers have developed a new AI model called NV-Raw2Insights-US that processes raw ultrasound data directly, rather than relying on traditional image reconstruction methods. This approach allows …