Artificial intelligence workloads, patterns, model deployment, and orchestration on AKS.
View All Tags
Dynamic Resource Allocation (DRA) is often mentioned in discussions about GPUs and specialized devices designed for high-performance AI and video processing jobs. But what exactly is it?
This blog post is co-authored with Rohan Varma, Saurabh Aggarwal, Anish Maddipoti, and Amr Elmeleegy from NVIDIA to showcase solutions that help customers run AI inference at scale using Azure Kubernetes Service (AKS) and NVIDIA’s advanced hardware and distributed inference frameworks.
Modern language models now routinely exceed the compute and memory capacity of a single GPU or even a whole node with multiple GPUs on Kubernetes. Consequently, inference at the scale of billions of model parameters demands multi-node, distributed deployment. Frameworks like the open-source NVIDIA Dynamo platform play a crucial role by coordinating execution across nodes, managing memory resources efficiently, and accelerating data transfers between GPUs to keep latency low.
However, software alone cannot solve these challenges. The underlying hardware must also support this level of scale and throughput. Rack-scale systems like Azure ND GB200-v6 VMs, accelerated by NVIDIA GB200 NVL72, meet this need by integrating 72 NVIDIA Blackwell GPUs in a distributed GPU setup connected via high-bandwidth, low-latency interconnect. This architecture uses the rack as a unified compute engine and enables fast, efficient communication and scaling that traditional multi-node setups struggle to achieve.
For some more demanding or unpredictable workloads, even combining advanced hardware and distributed inference frameworks is not sufficient on its own. Inference traffic spikes unpredictably. Fixed, static inference configurations and setups with predetermined resource allocation can lead to GPU underutilization or overprovisioning. Instead, inference infrastructure must dynamically adjust in real time, scaling resources up or down to align with current demand without wasting GPU capacity or risking performance degradation.
To effectively address the variability in inference traffic in distributed deployments, our approach combines three key components: ND GB200-v6 VMs, the NVIDIA Dynamo inference framework, with an Azure Kubernetes Service (AKS) cluster. Together, these technologies provide the scale, flexibility, and responsiveness necessary to meet the demands of modern, large-scale inference workloads.
At the core of Azure’s ND GB200-v6 VM series is the liquid-cooled NVIDIA GB200 NVL72 system, a rack-scale architecture that integrates 72 NVIDIA Blackwell GPUs and 36 NVIDIA Grace™ CPUs into a single, tightly coupled domain.
The rack-scale design of ND GB200-v6 unlocks model serving patterns that were previously infeasible due to interconnect and memory bandwidth constraints.
NVIDIA Dynamo is an open source distributed inference serving framework that supports multiple engine backends, including vLLM, TensorRT-LLM, and SGLang. It disaggregates the prefill (compute-bound) and decode (memory-bound) phases across separate GPUs, enabling independent scaling and phase-specific parallelism strategies. For example, the memory-bound decode phase can leverage wide expert parallelism (EP) without constraining the compute-heavy prefill phase, improving overall resource utilization and performance.
Dynamo includes an SLA-based Planner that proactively manages GPU scaling for prefill/decode (PD) disaggregated inference. Using pre-deployment profiling, it evaluates how model parallelism and batching affect performance, recommending configurations that meet latency targets like Time to First Token (TTFT) and Inter-Token Latency (ITL) within a given GPU budget. At runtime, the Planner forecasts traffic with time-series models, dynamically adjusting PD worker counts based on predicted demand and real-time metrics.
The Dynamo LLM-aware Router manages the key-value (KV) cache across large GPU clusters by hashing requests and tracking cache locations. It calculates overlap scores between incoming requests and cached KV blocks, routing requests to GPUs that maximize cache reuse while balancing workload. This cache-aware routing reduces costly KV recomputation and avoids bottlenecks, which in turn improves performance, especially for large models with long context windows.
To reduce GPU memory overhead, the Dynamo KV Block Manager offloads infrequently accessed KV blocks to CPU RAM, SSDs, or object storage. It supports hierarchical caching and intelligent eviction policies across nodes, scaling cache storage to petabyte levels while preserving reuse efficiency.
Dynamo’s disaggregated execution model is especially effective for large, dynamic inference workloads where compute and memory demands shift across phases. The Azure Research paper "Splitwise: Efficient generative LLM inference using phase splitting" demonstrated the benefits of separating the compute-intensive prefill and memory-bound decode phases of LLM inference onto different hardware. We will explore this disaggregated model in detail in an upcoming blog post.
Let’s put Dynamo’s features in context by walking through a realistic app scenario and explore how its framework addresses common inference challenges on AKS.
Imagine you operate a large e-commerce platform (or provide infrastructure for one), where customers browse thousands of products in real time. The app runs on AKS and experiences traffic surges during sales, launches, and seasonal events. The app also leverages LLMs to generate natural language outputs, such as:
This architecture powers user experiences like: “Customers who viewed this camera also looked at these accessories, chosen for outdoor use and battery compatibility.” Personalized product copies are dynamically rewritten for different segments, such as “For photographers” vs. “For frequent travelers.”
Behind the scenes, it requires a multi-stage LLM pipeline: retrieving product/user context, running prompted inference, and generating natural language outputs per session.
Heavy Prefill + Lightweight Decode = GPU Waste
Generating personalized recommendations requires a heavy prefill stage (processing more than 8,000 tokens of context) but results in short outputs (~50 tokens). Running both on a single GPU can be inefficient.
Dynamo Solution: The pipeline is split into two distinct stages, each deployed on separate GPUs. This allows independent configuration of GPU count and model parallelism for each phase. It also enables the use of different GPU types—for example, GPUs with high compute capability but lower memory for the prefill stage, and GPUs with both high compute and large memory capacity for the decode stage.
In our e-commerce example, when a user lands on a product page:
Prefill runs uninterrupted on dedicated GPUs using model parallelism degrees optimized for accelerating math-intensive attention GEMM operation. This enables fast processing of 8,000 tokens of user context and product metadata.
Decode runs on a GPU pool with different counts and parallelism degrees designed and tuned to maximize memory bandwidth and capacity for generating the short product blurb.
Result: This approach maximizes GPU utilization and reduces per-request cost.
Meeting SLOs and handling traffic spikes without overprovisioning
Your SLO might define time-to-first-token < 300ms and 99th percentile latency < 500ms, but maintaining this across dynamic workloads is tough. Static GPU allocation leads to bottlenecks during traffic spikes, causing either SLO violations or wasted capacity.
Dynamo Solution: Continuously monitors metrics and auto-scales GPU replicas or reallocates GPUs between prefill and decode stages based on real-time traffic patterns, queue depth, and latency targets.
In our e-commerce example:
Result: SLOs are met consistently without over or under provisioning, controlling costs while maintaining performance.
Recomputing shared context is wasteful
Many requests within the same session reuse the same product or user context but unnecessarily recompute the KV cache each time, wasting valuable GPU resources that could be spent serving other user requests.
Dynamo Solution: LLM-aware routing maintains a map of KV cache across large GPU clusters and directs requests to the GPUs that already hold the relevant KV cache, avoiding redundant computation.
In our e-commerce example:
Result: Faster response times, lower latency, reduced GPU usage.
KV cache growth blows past GPU memory
With many concurrent sessions and large input sequence lengths, the KV cache (product data + user history) can exceed available GPU memory. This can trigger evictions, leading to costly re-computations or inference errors.
Dynamo Solution: The KV Block Manager (KVBM) offloads cold/unused KV cache data to CPU RAM, NVMe, or networked storage freeing valuable GPU memory for active requests.
In our e-commerce example:
Result: Higher concurrency at lower cost, without degrading user experience.
We set out to deploy the popular open-source GPT-OSS 120B reasoning model using Dynamo on AKS on GB200 NVL72, adapting the SemiAnalysis InferenceMAX recipe for a large scale, production-grade environment.
Our approach: leverage Dynamo as the inference server and swap GB200 NVL72 nodes in place of NVIDIA HGX™ B200, scaling the deployment across multiple nodes.
Our goal was to replicate the performance results reported by SemiAnalysis, but at a larger scale within an AKS environment, proving that enterprise-scale inference with cutting-edge hardware and open-source models is not only possible, but practical.
Ready to build the same setup? Our comprehensive guide walks you through each stage of the deployment:
Find the complete recipe for GPT-OSS 120B at aka.ms/dynamo-recipe-gpt-oss-120b and get hands-on with the deployment guide at aka.ms/aks-dynamo.
By following this approach, we achieved 1.2 million tokens per second, meeting our goal of replicating SemiAnalysis InferenceMAX results at enterprise scale. This demonstrates that Dynamo on AKS running on ND GB200-v6 instances can deliver the performance needed for production inference workloads.
This work reflects a deep collaboration between Azure and NVIDIA to reimagine how large-scale inference is built and operated, from the hardware up through the software stack. By combining GB200 NVL72 nodes and the open-source Dynamo project on AKS, we’ve taken a step toward making distributed inference faster, more efficient, and more responsive to real-world demands.
This post focused on the foundational serving stack. In upcoming blogs, we will build on this foundation and explore more of Dynamo's advanced features, such as Disaggregated Serving and SLA-based Planner. We'll demonstrate how these features allow for even greater efficiency, moving from a static, holistic deployment to a flexible, phase-splitted architecture. Moving forward, we also plan to extend our testing to include larger mixture-of-experts (MoE) reasoning models such as DeepSeek R1. We encourage you to try out the Dynamo recipe in this blog on AKS and share your feedback!
It's been a few months since the AKS-MCP server was announced. Since then, there have been several updates and improvements. The MCP server can be easily installed on a local machine using the AKS Extension for VS Code, or via the GitHub MCP registry, or even using the Docker MCP hub.
In this blog post, I'll show you one approach to running the AKS MCP server: deploying it inside an AKS cluster as a Streamable HTTP service. This pattern demonstrates how MCP servers can be centrally managed and made accessible to multiple clients—including AI assistants, automation tools, and even autonomous agents.
In this blog, we'll guide you through setting up an OpenAI API compatible inference endpoint with llm-d and integrating with retrieval- augmented generation (RAG) on AKS. This blog will showcase its value in a key finance use case: indexing the latest SEC 10-K filings for the two S&P 500 companies and querying them. We’ll also highlight the benefits of llm-d based on its architecture and its synergy with RAG.
Recently, we released the AKS-MCP server, which enables AKS customers to automate diagnostics, troubleshooting, and cluster management using natural language.
One of its key capabilities is real-time observability using inspektor_gadget_observability MCP tool, which leverages a technology called eBPF to help
customers quickly inspect and debug applications running in AKS clusters.
At KubeCon India earlier this month, the AKS team shared our newest Agentic AI-powered feature with the broader Kubernetes community: the CLI Agent for AKS. CLI Agent for AKS is a new AI-powered command-line experience designed to help Azure Kubernetes Service (AKS) users troubleshoot, optimize, and operate their clusters with unprecedented ease and intelligence.
We're excited to announce the launch of the AKS-MCP Server. An open source Model Context Protocol (MCP) server designed to make your Azure Kubernetes Service (AKS) clusters AI-native and more accessible for developers, SREs, and platform engineers through Agentic AI workflows.
AKS-MCP isn't just another integration layer. It empowers cutting-edge AI assistants (such as Claude, Cursor, and GitHub Copilot) to interact with AKS through a secure, standards-based protocol—opening new possibilities for automation, observability, and collaborative cloud operations.
Temporal is an open source platform that helps developers build and scale resilient Enterprise and AI applications. Complex and long-running processes are easily orchestrated with durable execution, ensuring they never fail or lose state. Every step is tracked in an Event History that lets developers easily observe and debug applications. In this guide, we will help you understand how to run and scale your workers on Azure Kubernetes Service (AKS).
XL-size large language models (LLMs) are quickly evolving from experimental tools to essential infrastructure. Their flexibility, ease of integration, and growing range of capabilities are positioning them as core components of modern software systems.
Massive LLMs power virtual assistants and recommendations across social media, UI/UX design tooling and self-learning platforms. But how do they differ from your average language model? And how do you get the best bang for your buck running them at scale?
Let’s unpack why large models matter and how Kubernetes, paired with NVMe local storage, accelerates intelligent app development.
High performance computing (HPC) workloads, like large-scale distributed AI training and inferencing, often require fast, reliable data transfer and synchronization across the underlying compute. Model training, for example, requires shared memory across GPUs because the parameters and gradients need to be constantly shared. For models with billions of parameters, the available memory in a single GPU node may not be enough, so "pooling" the memory across multiple nodes also requires high memory bandwidth due to the sheer volume of data involved. A common way to achieve this at scale is with a high-speed, low-latency network interconnect technology called InfiniBand (IB).
With the fast-paced advancement of AI workloads, building and fine-tuning of multi-modal models, and extensive batch data processing jobs, more and more enterprises are leaning into Kubernetes platforms to take advantage of its ability to scale and optimize compute resources. With AKS, you can manage up to 5,000 nodes (upstream K8s limit) in a single cluster under optimal conditions, but for some large enterprises, that might not be enough.
We've released new guidance for running Ray on AKS!
Data is often at the heart of application design and development - it fuels user-centric design, provides insights for feature enhancements, and represents the value of an application as a whole. In that case, shouldn’t we use data science tools and workflows that are flexible and scalable on a platform like Kubernetes, for a range of application types?
In collaboration with David Espejo and Shalabh Chaudhri from Union.ai, we’ll dive into an example using Flyte, a platform built on Kubernetes itself. Flyte can help you manage and scale out data processing and machine learning pipelines through a simple user interface.
You may have heard of the Kubernetes AI Toolchain Operator (KAITO) announced at Ignite 2023 and KubeCon Europe this year. The open source project has gained popularity in recent months by introducing a streamlined approach to AI model deployment and flexible infrastructure provisioning on Kubernetes.
With the v0.3.0 release, KAITO has expanded the supported model library to include the Phi-3 model, but the biggest (and most exciting) addition is the ability to fine-tune open-source models. Why should you be excited about fine-tuning? Well, it’s because fine-tuning is one way of giving your foundation model additional training using a specific dataset to enhance accuracy, which ultimately improves the interaction with end-users. (Another way to increase model accuracy is Retrieval-Augmented Generation (RAG), which we touch on briefly in this section, coming soon to KAITO).