News For Dev-ops

Welcome to the D3D12 LinAlg Matrix Preview release! Today, we are excited to announce the preview release for the D3D12 Linear Algebra APIs! This feature set unlocks comprehensive hardware acceleration for Matrix-oriented operations across various use cases. Previously, we announced the WaveMMA and Cooperative Vectors features which supported narrow matrix operation use cases; the LinAlg feature set being announced today subsumes these APIs into a singular set of orthogonal APIs. With today's announcement, we are enabling developers to both efficiently drive neural rendering techniques directly from individual shader threads in real-time graphics pipelines and utilize higher bandwidth matrix MMA operations for ML and image processing applications, all in a singular combined API. The application of machine learning techniques is now ubiquitous across the industry. For graphics development, neural network based rendering methods, which we’ve been calling neural rendering, are quick...

In our exploration of calling conventions for various processors on Windows, we learned that in many cases, some of the parameters are passed in registers. Suppose that there is a function that takes two parameters, but you know that the function ignores the second parameter if the first parameter is positive. What happens if you call the function with just one parameter (say, passing zero). The function should ignore the second parameter, so why does it matter that you didn't pass one? Even though the function doesn't use the parameter, it still may decide to use the storage for that parameter as a conveniently provided scratch space. For example: int blah(int a, int b) { if (a <= 0) { int c = f1(); f2(a); return c; } else { return f3(a, b); } Is it okay to call blah with zero as its only parameter? You aren't passing b , but the function doesn't use b , so why does it matter? Formally, the C and C++ languages say that i...

Amazon SageMaker HyperPod now supports G7e and r5d.16xlarge instances. SageMaker HyperPod is a purpose-built infrastructure for developing, training, and deploying foundation models at scale. It provides a resilient and performant environment with built-in fault tolerance, automated cluster recovery, and optimized distributed training libraries, reducing the undifferentiated heavy lifting of managing large-scale AI/ML infrastructure.  G7e instances are powered by NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs and deliver up to 2.3x better inference performance than G6e instances, allowing you to process more requests per second while reducing latency. With up to 768 GB of total GPU memory, G7e instances let you deploy larger language models or run multiple models on a single endpoint. You can use these instances for deploying LLMs, agentic AI, multimodal generative AI, and physical AI models. G7e instances are also well suited for cost-efficient single-node fine-tuning or trai...

AWS announces availability of new Linux bundles for Amazon WorkSpaces Personal, including Rocky Linux 9, Red Hat Enterprise Linux 9, and Ubuntu 24.04. With these bundles, customers can launch WorkSpaces powered by the latest enterprise-grade Linux operating systems and take advantage of modern versions of Linux packages only available in these updated releases. While Rocky Linux 8, Red Hat Enterprise Linux 8, and Ubuntu 22.04 powered WorkSpaces bundles remain available, the new OS options bring access to the latest software ecosystems, improved security postures, and extended long-term support lifecycles offered by each respective distribution. These new bundles also provide a migration path for Amazon Linux 2 customers ahead of its end of life in June 2026. You can get started using managed Rocky Linux 9, Red Hat Enterprise Linux 9, or Ubuntu 24.04 WorkSpaces bundles by selecting one when creating a new Linux WorkSpace. These new bundles are available in all AWS Regions where Amaz...

AWS Lambda now supports Provisioned Mode for event source mappings (ESMs) that subscribe to Apache Kafka event sources in the Asia Pacific (Taipei), AWS GovCloud (US-East), and AWS GovCloud (US-West) Regions. Provisioned Mode allows you to optimize the throughput of your Kafka ESM by provisioning event polling resources that remain ready to handle sudden spikes in traffic, helping you build highly responsive and scalable event-driven Kafka applications with stringent performance requirements. Customers building streaming data applications often use Kafka as an event source for Lambda functions, relying on Lambda's fully managed ESM to automatically scale polling resources in response to events. However, for event-driven Kafka applications that need to handle unpredictable bursts of traffic, lack of control over the throughput of ESM can lead to delays in your users' experience. Provisioned Mode for Kafka ESM enables customers to fine-tune the throughput of their Amazon Managed...

One of the handy helpers in the Windows Implementation Library (WIL) is wil:: scope_ exit . We've used it to simulate the finally keyword in other languages by arranging for code to run when control leaves a scope. I've identified three places where exceptions can occur when using scope_ exit . auto cleanup = wil::scope_exit([captures] { action; }); One is at the construction of the lambda. What happens if an exception occurs during the initialization of the captures? This exception occurs even before scope_ exit is called, so there's nothing that scope_ exit can do. The exception propagates outward, and the action is never performed. Another is at the point the scope_ exit tries to move the lambda into cleanup . In a naïve implementation of scope_ exit , the exception would propagate outward without the action ever being performed. The third point is when the scope_ exit is destructed. In that case, it's an exception thrown from a destructor. Since destruc...