Meta has detailed MTIA 300, its first in-house accelerator optimized for training ranking and recommendation models. By Matt Foster

Amazon Web Services has extended Amazon Bedrock AgentCore with runtime instances, a new compute option that gives AI agents persistent infrastructure purpose-built for complex long-running...

Arun Joseph shares real-world insights on scaling enterprise agentic platforms like Deutsche Telekom’s LMOS. He discusses bridging organizational fault lines, replacing tool sprawl with core...

Oracle halved the Always Free Ampere A1 compute allowance from 4 OCPUs and 24 GB RAM to 2 OCPUs and 12 GB RAM with no public announcement. Support agents gave conflicting answers on whether PAYG...

Apple chose Google Cloud to run Private Cloud Compute outside its own data centers for the first time, using NVIDIA Blackwell GPUs, Intel TDX, and Google's Titan chip. Apple maintains an...

Roofline模型用于判断算子是计算密集型还是内存密集型，算术强度（AI）是关键指标，定义为浮点运算数与内存搬运字节数之比。通过双对数图，Roofline展示了性能与算术强度的关系，分为带宽限制区和算力限制区。优化策略包括提高算术强度，采用算子融合和tiling等方法，以减少内存访问并提升性能。

本文介绍了NVIDIA的两个调优工具：Nsight Systems和Nsight Compute。Nsight Systems用于分析程序的时间线，找出占用时间最多的kernel；而Nsight Compute则深入分析单个kernel的性能瓶颈。优化流程是先用Systems定位热点，再用Compute分析原因，如访存延迟或计算瓶颈。在没有完整工具链时，可使用CUDA event进行轻量测量。最终目标是提升GPU性能，优化计算效率。

本文讨论了GPU kernel的调试与数值正确性，主要包括内存/竞态错误和数值错误两类。使用compute-sanitizer工具检查内存问题，并通过与高精度参考实现对比来验证数值正确性。强调浮点运算的非结合性可能导致结果微小差异，需用容差比较。总结了常见的kernel错误及调试方法，确保正确性是关键。

AI’s next breakthrough may not be a smarter model but a cheaper token.

By Alvin Bao, Alex Petrov, Jennifer Lai, Aidan Sherr, and Samartha ChandrashekarAs a part of the journey to transition Netflix’s compute infrastructure to be more Kubernetes-native, we have leaned...

PostgreSQL 14 unified query-id computation across all subsystems, but defaulting to always-on would tax every backend.

Australia has an opportunity to become an Asia–Pacific AI hub, unlocking economic growth and productivity—though a number of constraints need to be addressed.

Uber launches IngestionNext, a streaming-first data lake ingestion platform that reduces data latency from hours to minutes and cuts compute usage by 25%. Built on Kafka, Flink, and Apache Hudi,...

Power and cooling equipment are the backbones of data center infrastructure. Innovations and on-time supply of this technology will become increasingly relevant as the demand for data centers grows.

We are thrilled to announce the newest member of our JupyterLite kernel ecosystem: Xeus-Octave. Xeus-Octave allows you to run GNU Octave code directly on your browser. GNU Octave is a free and...