Posts

AWS Parallel Computing Service supports P6e-GB200 and P6e-GB300 UltraServers - devamazonaws.blogspot.com

AWS Parallel Computing Service (PCS) now supports Amazon EC2 P6e-GB200 and P6e-GB300 UltraServer instances, enabling customers to run large-scale GPU workloads using the NVIDIA Blackwell architecture within Slurm-managed clusters. You can reserve UltraServers through EC2 Capacity Blocks for ML, associate them with a PCS compute node group via an EC2 launch template, and PCS automatically configures Slurm with the correct topology plugin. With P6e-GB200 UltraServers, you can access up to 72 NVIDIA Blackwell GPUs within one NVLink domain to use 360 petaflops of FP8 compute (without sparsity) and 13.4 TB of total high bandwidth memory (HBM3e). P6e-GB300 UltraServers provide 1.5x GPU memory and 1.5x FP4 compute (without sparsity) compared to P6e-GB200. AWS PCS is a managed service that simplifies running and scaling HPC workloads on AWS using Slurm. You can build complete, elastic environments that integrate compute, storage, networking, and visualization tools, while the service handl...
Post a Comment
Read more

[MS] Models don't have preferences, they have context - devamazonaws.blogspot.com

You open a fresh chat, type "What framework should I use for a web app?", and the model says "React." You screenshot it, share it, and write "Claude prefers React." It gets engagement. People nod along. A few reply with their own results. And now we have a consensus: Claude prefers React. Except it doesn't. The model doesn't prefer anything. You're reading the room, not the mind. The genre There's a whole genre of this. Run N prompts in a bare chat window, tabulate the answers, maybe build a heatmap, publish it as a blog post or a thread. "Which frameworks do LLMs prefer?" "What languages do models recommend most?" The format varies, but the method is the same: ask abstract questions in an empty context, count the answers, declare preferences. It's easy to see why this is appealing. It feels scientific: you have a sample size, you have a methodology, and it produces shareable results and charts. The problem isn...
Post a Comment
Read more

[MS] How to Use Deep Agents with Azure Cosmos DB – Plan, act, and verify against operational data - devamazonaws.blogspot.com

Image
Deep Agents is an agent harness built on LangGraph , for agents that need to work through a task over many steps instead of a single LLM call. The agent runs tools, looks at the results, and uses that to pick the next one, keeping a todo list as it goes. On top of that loop the harness brings what a longer-running agent needs. It can load instructions on demand instead of holding everything in the prompt (skills), offload large tool outputs so they don’t fill the context window, and pause for human approval in apps that need an approval gate before data changes. Support Ops Agent is a sample app that puts this to work on a customer-support ticket queue. We can ask it which tickets are at risk, who’s overloaded, or whether a run of similar complaints is really one outage. When a ticket needs to change, it updates the ticket and reads it back to confirm. Most requests become a handful of reads against the queue. Requests that change a ticket add a patch and a verification read. That q...
Post a Comment
Read more

[MS] Retirement of Azure DevOps issuer in Workload identity federation service connections - devamazonaws.blogspot.com

Image
We are announcing the deprecation of the Azure DevOps issuer in workload identity federation (WIF) service connections , with planned retirement on July 1, 2027 . The Azure DevOps issuer uses the https://vstoken.dev.azure.com prefix in federated credentials. This change is part of Microsoft's broader initiative to standardize on the Microsoft Entra issuer across Azure services that implement workload identity federation. Important This deprecation only applies to service connections in Azure public cloud that use single-tenant Microsoft Entra applications or managed identities. Service connections targeting non-public clouds (for example, Azure Government, Azure China, or Azure Stack) and service connections that use multi-tenant applications ( signInAudience: AzureADMultipleOrgs ) are explicitly excluded from today's deprecation announcement. The Azure DevOps issuer will continue to be supported for these scenarios until they're supported by the Microsoft Entra issu...
Post a Comment
Read more

[MS] What 50,000 Runs of a 5-Line Eval Taught Us - devamazonaws.blogspot.com

How AI coding models calibrate effort, token cost, and tool use on even the simplest task, and what that means for model selection and cost. Read the full article Post Updated on June 19, 2026 at 01:00AM Thanks for reading from devamazonaws.blogspot.com
Post a Comment
Read more

[MS] What 50,000 Runs of a 5-Line Eval Taught Us - devamazonaws.blogspot.com

How AI coding models calibrate effort, token cost, and tool use on even the simplest task, and what that means for model selection and cost. Read the full article Post Updated on June 19, 2026 at 01:00AM Thanks for reading from devamazonaws.blogspot.com
Post a Comment
Read more

Amazon RDS for SQL Server increases the maximum size and provisioned performance of General Purpose (gp3) volumes - devamazonaws.blogspot.com

Amazon Relational Database Service (Amazon RDS) for SQL Server now supports higher volume-level limits for General Purpose (gp3) storage. With this update, each gp3 volume can scale up to 64 TiB in size (4X the previous 16 TiB limit), up to 80,000 IOPS (5X the previous 16,000 IOPS limit), and up to 2,000 MiB/s throughput (2X the previous 1,000 MiB/s limit). With these improvements, customers can now run larger Microsoft SQL Server databases on Amazon RDS. Workloads with demanding I/O requirements such as high-throughput OLTP systems and large-scale analytical workloads can take advantage of higher IOPS and throughput on a single volume with simplified storage management, and get better performance for mission-critical SQL Server workloads. Additionally, you can configure additional storage volumes to add up to three gp3 or io2 volumes per DB instance, increasing total capacity up to 256 TiB per instance. There is no change to pricing - customers pay for storage and any additional I...
Post a Comment
Read more