Breaking the Million-Token Barrier: The Technical Achievement of Azure ND GB300 v6
Azure ND GB300 v6 Virtual Machines with NVIDIA GB300 NVL72 rack-scale systems achieve unprecedented performance of 1,100,000 tokens/s on Llama2 70B Inference, beating the previous Azure ND GB200 v6 record of 865,000 tokens/s by 27%.
A Quick Guide to Benchmarking AI models on Azure: Llama 405B and 70B with MLPerf Inference v5.1
by Mark Gitau (Software Engineer) Introduction For the MLPerf Inference v5.1 submission, Azure shared performance results on the new ND GB200 v6 virtual machines. A single ND GB200 v6 VM on Azure is powered by two NVIDIA Grace CPUs and four NVIDIA Blackwell B200 GPUs. This document highlights Azure’s MLPerf Inference v5.1 results and outlines the steps to run these benchmarks on Azure. These MLPerf™ benchmark results demonstrate Azure’s commitment to providing our customers with the latest GPU offerings of the highest quality. Highlights from MLPerf Inference v5.1 benchmark results include: Azure had the highest Llama 2 70B Offline submission with 52,000 tokens/s on a single ND GB200 v6 virtual machine. This corresponds to an 8% increase on single node performance since our record which would correspond to 937,098 tokens/s on a full NVL72 rack. Azure results for Llama 3.1 405B are at par with the best submitters (1% difference), cloud and on-premises, with 847 tokens/s. How to replicate the results in Azure Pre-requisites: ND GB200 v6-series (single node): Deploy and set up a virtual machine on Azure Set up the environment First, we need to export the path to the directory where we will perform the benchmarks. For ND GB200 v6-series (single node), create a directory called mlperf in /mnt/nvme Set mlperf scratch space: export MLPERF_SCRATCH_PATH=/mnt/nvme/mlperf Clone the MLPerf repository inside the scratch path: git clone https://github.com/mlcommons/inference_results_v5.1.git Then create empty directories in your scratch space to house the data: mkdir $MLPERF_SCRATCH_PATH/data $MLPERF_SCRATCH_PATH/models $MLPERF_SCRATCH_PATH/preprocessed_data Download the models & datasets Download the models inside the models directory you created in the previous step. This will take a while because the weights are large. Llama 2 70B model Llama 3.1 405B model Download the preprocessed datasets for both models Llama 2 70B datasets Llama 3.1 405B datasets Prepare the datasets for Llama 2 70B: inference_results_v5.1/closed/Azure/code/llama2-70b/tensorrt at main · mlcommons/inference_results_v5.1 prepare the datasets for Llama 3.1 405B: inference_results_v5.1/closed/Azure/code/llama3.1-405b/tensorrt at main · mlcommons/inference_results_v5.1 Build & launch MLPerf container Export the Submitter and System name: export SUBMITTER=Azure SYSTEM_NAME=ND_GB200_v6 Enter the container by entering the closed/Azure directory and running: make prebuild Inside the container, run make build Build engines & run benchmarks Make sure you are still in the closed/Azure directory of the MLPerf repository To build the engines for both Llama 3.1 405B and Llama 2 70B: make generate_engines RUN_ARGS="--benchmarks=llama2-70b,llama3.1-405b --scenarios=offline,server" To run the benchmarks for both Llama 3.1 405B and Llama 2 70B: make run_harness RUN_ARGS=="--benchmarks=llama2-70b,llama3.1-405b --scenarios=offline,server" MLPerf from MLCommons® MLCommons® is an open engineering consortium of AI leaders from academia, research, and industry where the mission is to “build fair and useful benchmarks” that provide unbiased evaluations of training and inference performance for hardware, software, and services—all conducted under predetermined conditions. MLPerf™ Inference benchmarks consist of compute-intensive AI workloads that simulate realistic usage of the systems, making the results very influential in technology management’s buying decisions.Azure’s ND GB200 v6 Delivers Record Performance for Inference Workloads
Achieving peak AI performance requires both cutting-edge hardware and a finely optimized infrastructure. Azure’s ND GB200 v6 Virtual Machines, accelerated by the NVIDIA GB200 Blackwell GPUs, have already demonstrated world record performance of 865,000 tokens/s for inferencing on the industry standard LLAMA2 70B
Performance analysis of DeepSeek R1 AI Inference using vLLM on ND-H100-v5
Introduction The DeepSeek R1 model represents a new frontier in large-scale reasoning for AI applications. Designed to tackle complex inference tasks, R1 pushes the boundaries of what’s possible—but not without significant infrastructure demands. To deploy DeepSeek R1 effectively in an inference service like vLLM, high-performance hardware is essential. Specifically, the model requires two Azure ND_H100_v5 nodes, each equipped with 8 NVIDIA H100 GPUs, totaling 16 H100s. These nodes are interconnected via InfiniBand and NVLink, ensuring the bandwidth and latency characteristics necessary to support the model’s massive memory footprint and parallel processing needs. In this post, we’ll present inference benchmark results for DeepSeek R1, measuring performance across GPU utilization, memory throughput, and interconnect efficiency. While R1 excels in reasoning tasks, it’s important to recognize that such models are not universally optimal. For many general-purpose AI applications, smaller models like Llama 3.1 8B offer a compelling alternative, delivering sufficient accuracy and performance at a fraction of the cost. We explore the performance characteristics of DeepSeek R1 and help you decide when a large reasoning model is worth the investment—and when a leaner solution might be the better choice. Benchmark environment 2 ND_H100_v5 nodes (16 H100) were required to load the DeepSeek R1 model into GPU memory. vLLM was used to set-up the online inference server using a single API server endpoint. vLLM bench (specifically hugging face AI-MO/aimo-validation-aime) was used to run the inference benchmark. Detailed specifications are available in the tables below. Results Reasoning model’s additional chain of thought tokens We show that DeepSeek R1 generates many additional tokens (for chain of thought), which helps reasoning models answer complex logical questions. In this example the DeepSeek R1 generates 1162 completion tokens, but Llama 3.1 8B only generates 37. This example also demonstrates that while non-reasoning models may struggle with this seemingly simple logical question, reasoning models like DeepSeek R1 can solve it—though they require significantly more intermediate tokens to do so. #!/bin/bash curl http://localhost:8000/v1/chat/completions \ -H "Content-Type: application/json" \ -d '{ "model": "deepseek-ai/DeepSeek-R1", "messages": [ {"role": "user", "content": "9.11 and 9.8, which is greater? Explain your reasoning"} ] }' Note: The REST API json output is formatted to make the output more readable. (./run_restapi_query.sh 2> /dev/null | python3 ./convert_json_2_txt.py >& run_query.out_$$) Reasoning: Okay, so I need to figure out whether 9.11 is greater than 9.8 or if 9.8 is greater. Hmm, let me start by looking at both numbers. They both have a 9 in the units place, so that part is equal. Now, I need to compare the decimal parts: .11 and .8. Wait, but decimals can be tricky because the places matter. The first number is 9.11, which has a 1 in the tenths place and a 1 in the hundredths place. The second number is 9.8, which is the same as 9.80 when you add a zero to the hundredths place for comparison. So, 9.80 versus 9.11. Now, looking at the tenths place: 8 is greater than 1. Therefore, 9.80 should be greater than 9.11. But let me double-check because sometimes when numbers have different decimal lengths, it's easy to make a mistake. If I convert both to fractions, maybe that will help. 9.11 is the same as 9 + 0.11, and 0.11 is 11/100. On the other hand, 9.8 is 9 + 0.8, which is 9 + 8/10. To compare 8/10 and 11/100, I can convert them to have the same denominator. 8/10 is equal to 80/100. Now, 80/100 is much larger than 11/100. So yes, 9.8 is indeed greater than 9.11. Another way to see it is by subtracting the two numbers. If I subtract 9.11 from 9.8, what do I get? Let's see: 9.8 - 9.11. Since 9.8 is the same as 9.80, subtracting 9.11 gives 0.69. Since the result is positive, that means 9.8 is larger. Wait, maybe I can also think about money to make it clearer. If I have $9.11 and you have $9.80, which is more? $9.80 is $9 and 80 cents, whereas $9.11 is $9 and 11 cents. Definitely, 80 cents is more than 11 cents, so $9.80 is more. That helps visualize it. Alternatively, using a number line. If I imagine a number line with 9.0, 9.1, 9.2,... up to 10.0. 9.11 is just a little past 9.1, while 9.8 is almost at 10.0. So clearly, 9.8 is much further to the right, meaning it's a larger number. I think that covers different methods: decimal place comparison, converting to fractions, subtraction, real-life money examples, and number line visualization. All of these methods point to 9.8 being greater than 9.11. I don't see any contradictions here. Maybe I can also consider writing both numbers with the same number of decimal places to make it even clearer. 9.11 is already two decimal places, and 9.8 can be written as 9.80. Now, comparing 9.80 and 9.11 digit by digit: - The units place: 9 vs. 9 (equal) - Tenths place: 8 vs. 1 (8 is greater) - Since the tenths place of 9.80 is greater, the entire number is larger, regardless of the hundredths place. So yes, that confirms it again. The key idea here is that when comparing decimals, you start from the leftmost digit after the decimal point and move right until you find a difference. In this case, the tenths place already shows that 9.8 is larger. I guess another common mistake could be thinking that because 11 is more than 8, 0.11 is more than 0.8, but that's not correct because the places are different. Tenths versus hundredths. 0.8 is equivalent to 0.80, which is way bigger than 0.11. To prevent that mistake, always align the decimal points and add trailing zeros if necessary. Then compare each digit one by one from left to right. That method avoids confusion. So, all in all, through multiple approaches, 9.8 is definitely greater than 9.11. Answer: To determine which number is greater between **9.11** and **9.8**, follow these steps: **Align the decimal places** by rewriting 9.8 as **9.80**. **Compare digit by digit** from left to right: - **Units place**: Both have **9** (equal). - **Tenths place**: **8** (in 9.80) vs. **1** (in 9.11). Since \(8 > 1\), **9.80 is greater** at this point. - No need to compare hundredths place once a difference is found. **Conclusion**: \(9.8\) (or \(9.80\)) is greater than \(9.11\) because **8 tenths** (\(0.8\)) is larger than **1 tenth + 1 hundredth** (\(0.11\)). \(\boxed{9.8 \text{ is greater}}\) Completion Info: Finish reason: stop Prompt tokens: 21 Completion tokens: 1162 Total tokens: 1183 When running using Llama 3.1 8B (Not reasoning model). Answer: To determine which is greater, we need to compare the two numbers. 9.11 is greater than 9.8 because 9.11 is a larger decimal value. Completion Info: Finish reason: stop Prompt tokens: 51 Completion tokens: 37 Total tokens: 88 Throughput and latency results Cost comparison In this cost analysis we use the ND-H100-v5 and ND-H200-v4 pay as you go pricing in south central US region and the measured total throughput tokens/sec to compute the $/(1K tokens). Note: ND-H200-v5 pricing was estimated at 20% more than ND-H100-v5 pricing. Analysis The DeepSeek R1 is a large, complex reasoning model that is costlier and slower than smaller models. It needs 16 H100 GPUs for FP8 precision and generates many more intermediate tokens in its chain of thought process—about 31 times more than Llama 3.1 8B—but at a much slower rate (~54 times slower). Its latency is also higher, with TTFT and ITL being roughly 6 and 3 times slower, respectively. The DeepSeek R1 model has small intranode and internode network requirements (~14% of available InfiniBand network bandwidth was used, and < 1% of available NVLink bandwidth is used. GPUs with higher memory bandwidth and higher FLOPS would help improve its performance. The cost analysis shows that the cost to generate DeepSeek R1 tokens is ~54 times more expensive than Llama 3.1 8B on the same 16 H100 GPU’s and ~34 times more expensive on 8 H200 GPU’s. DeepSeek R1 model is very capability, but due to its higher TCO it should be only used in specific AI applications that require its strong reasoning abilities. Conclusion The DeepSeek R1 model demonstrates exceptional reasoning capabilities, but its deployment demands substantial infrastructure and incurs high latency and cost. While it excels in generating detailed chains of thought, its throughput and efficiency lag significantly behind smaller models like Llama 3.1 8B. For applications requiring deep logical analysis, DeepSeek R1 is a powerful tool. However, for general-purpose inference tasks, more lightweight models offer better performance and cost-effectiveness. Strategic use of DeepSeek R1 should be reserved for scenarios where its advanced reasoning justifies the resource investment. References Deepseek R1 model on Hugging Face https://huggingface.co/deepseek-ai/DeepSeek-R1 vLLM GitHub repository https://github.com/vllm-project/vllm Azure ND H100 v5 documentation https://learn.microsoft.com/en-us/azure/virtual-machines/nd-h100-v5-series FlashInfer GitHub repository https://github.com/flashinfer-ai/flashinfer DeepGEMM GitHub repository https://github.com/deepseek-ai/DeepGEMM AI-MO validation dataset on Hugging Face https://huggingface.co/datasets/AI-MO/aimo-validation-aime Appendix Install vLLM curl -LsSf https://astral.sh/uv/install.sh | sh uv venv myvllm --python 3.11 --seed source myvllm/bin/activate uv pip install vllm --torch-backend=auto git clone https://github.com/flashinfer-ai/flashinfer.git --recursive uv pip install ninja cd flashinfer uv pip install --no-build-isolation --verbose . Install DeepSeek DeepEP git clone https://github.com/vllm-project/vllm.git cd ~/vllm/tools/ep_kernels export CUDA_HOME=/usr/local/cuda-12.8 TORCH_CUDA_ARCH_LIST="9.0" (For Hopper) bash install_python_libraries.sh 2.&1 | tee install_python_libraries.log_$$ sudo bash configure_system_drivers.sh 2>&1 | tee configure_system_drivers.log_$$ sudo reboot Install DeepSeek DeepGEMM git clone --recursive https://github.com/deepseek-ai/DeepGEMM.git cd deepGEMM ./install.sh 2>&1 | tee install.log_$$ Configure DeepSeek R1 with vLLM on 2 ND_H100_v5 Second node configuration Execute this script on second node before the script on the primary node. #!/bin/bash MODEL="deepseek-ai/DeepSeek-R1" PORT=8000 export VLLM_LOGGING_LEVEL=INFO export HF_HUB_CACHE=/home/azureuser/cgshared/hf_cache #export VLLM_ALL2ALL_BACKEND=deepep_high_throughput export VLLM_ALL2ALL_BACKEND=deepep_low_latency export VLLM_USE_DEEP_GEMM=1 export GLOO_SOCKET_IFNAME=eth0 vllm serve $MODEL --port $PORT --tensor-parallel-size 1 --enable-expert-parallel --data-parallel-size 16 --data-parallel-size-local 8 --data-parallel-start-rank 8 --data-parallel-address 10.0.0.6 --data-parallel-rpc-port 23345 --headless --max-model-len 32768 --reasoning-parser deepseek_r1 Primary node configuration #!/bin/bash MODEL="deepseek-ai/DeepSeek-R1" PORT=8000 export VLLM_LOGGING_LEVEL=INFO export HF_HUB_CACHE=/home/azureuser/cgshared/hf_cache #export VLLM_ALL2ALL_BACKEND=deepep_high_throughput export VLLM_ALL2ALL_BACKEND=deepep_low_latency export VLLM_USE_DEEP_GEMM=1 export GLOO_SOCKET_IFNAME=eth0 vllm serve $MODEL --port $PORT --tensor-parallel-size 1 --enable-expert-parallel --data-parallel-size 16 --data-parallel-size-local 8 --data-parallel-address 10.0.0.6 --data-parallel-rpc-port 23345 --api-server-count 1 --max-model-len 32768 --reasoning-parser deepseek_r1 Install vLLM benchmark environment cd vllm uv pip install vllm[bench] Run vLLM benchmark #!/bin/bash vllm bench serve \ --backend vllm \ --model deepseek-ai/DeepSeek-R1 \ --endpoint /v1/completions \ --dataset-name hf \ --dataset-path AI-MO/aimo-validation-aime \ --ramp-up-strategy linear \ --ramp-up-start-rps 1 \ --ramp-up-end-rps 10 \ --num-prompts 400 \ --seed 42Inference performance of Llama 3.1 8B using vLLM across various GPUs and CPUs
Introduction Following our previous evaluation of Llama 3.1 8B inference performance on Azure’s ND-H100-v5 infrastructure using vLLM, this report broadens the scope to compare inference performance across a range of GPU and CPU platforms. Using the Hugging Face inference benchmarker, we assess not only throughput and latency but also the cost-efficiency of each configuration—an increasingly critical factor for enterprise deployment. As organizations seek scalable and budget-conscious solutions for deploying large language models (LLMs), understanding the trade-offs between compute-bound and memory-bound stages of inference becomes essential. Smaller models like Llama 3.1 8B offer a compelling balance between capability and resource demand, but the underlying hardware and software stack can dramatically influence both performance and operational cost. This report presents a comparative analysis of inference performance across multiple hardware platforms, factoring in: Token throughput and latency across chat, classification, and code generation workloads. Resource utilization, including KV cache utilization and efficiency. Cost per token, derived from cloud pricing models and hardware utilization metrics. By combining performance metrics with cost analysis, we aim to identify the most effective deployment strategies for enterprise-grade LLMs, whether optimizing for speed, scalability, or budget. Benchmark environment Inference benchmark The Hugging face Inference benchmarking code was used for the AI Inference benchmark. Three different popular AI inference profiles were examined. Chat: Probably the most common use case, question and answer format on a wide range of topics. Classification: Providing various documents and requesting a summary of its contents. Code generation: Providing code and requesting code generation, e.g. create a new function. Profile Data set Input prompt Output prompt Chat hlarcher/inference-benchmarker/share_gpt_turns.json N/A min=50, max=800, variance=100 Classification hlarcher/inference-benchmarker/classification.json Min=8000, max=12000, variance=5000 Min=30, max=80, variance=10 Code generation hlarcher/inference-benchmarker/github_code.json Min=3000, max=6000, variance=1000 Min=30, max=80, variance=10 Huggingface Lama 3.1 8B models used Precision Model Size (GiB) meta-llama/Llama-3.1-8B-Instruct FP16 14.9 vLLM parameters Default value gpu_memory_utilization 0.9 max_num_seqs 1024 max_num_batched_tokens 2048 (A100), 8192 (H100,H200) enable_chunked_prefill True enable_prefix_caching True VM Configuration GPU ND-H100-v5, ND-H200-v5, HD-A100-v4 (8 H100 80GB &40GB) running HPC Ubuntu 22.04 (Pytorch 2.7.0+cu128, GPU driver: 535.161.08 and NCCL 2.21.5-1). 1 GPU was used in benchmark tests. CPU Ubuntu 22.02 (HPC and Canonical/jammy) Results GPU Profile Avg prompt throughput Avg generation throughput Max # Requests waiting Max KV Cache usage % Avg KV Cache hit rate % H100 Chat ~2667 ~6067 0 ~14% ~75% Classification ~254149 ~1291 0 ~46% ~98% Code generation ~22269 ~266 ~111 ~93% ~1% H200 Chat ~3271 ~7464 0 ~2% ~77% Classification ~337301 ~1635 0 ~24% ~99% Code generation ~22726 ~274 ~57 ~46% ~1% A100 Chat ~1177 ~2622 0 ~2% ~75% Classification ~64526 ~333 0 ~45% ~97% Code generation ~7926 ~95 ~106 ~21% ~1% A100_40G Chat ~1069 ~2459 0 ~27% ~75% Classification ~7846 ~39 ~116 ~68% ~5% Code generation ~7836 ~94 ~123 ~66% ~1% Cost analysis Cost analysis used pay-as-you-go pricing for the south-central region and measured throughput in tokens per second to calculate the metric $/(1K tokens). CPU performance and takeaways The Huggingface AI-MO/aimo-validation-aime data was by vllm bench to test the performance of Llama 3.1 8B on various VM types (left graph below). It is a struggle (insufficient FLOPs and memory bandwidth) to run Llama 3.1 8B on CPU VM’s, even the best performing CPU VM (HB176-96_v4) throughput and latency is significantly slower than the A100_40GB GPU. Tips Enable/use AVX512 (avx512f, avx512_bf16, avx512_vnni etc) (See what is supported/available via lscpu) Put AI model on single socket (if it has sufficient memory). For larger models you can use tensor parallel to split the model across sockets. Use pinning to specify which cores the threads will run on (in vLLM, VLLM_CPU_OMP_THREADS_BIND=0-22) Specify large enough KVCache (on CPU memory). In vLLM, VLLM_CPU_KVCACHE_SPACE=100) Analysis Throughput & Latency H200 outperforms all other GPUs across all workloads, with the highest prompt and generation throughput. H100 is a close second, showing strong performance especially in classification and code generation. A100 and A100_40G lag significantly behind, particularly in classification tasks where throughput drops by an order of magnitude (on A100_40G, due to smaller GPU memory and lower KV Cache hit percentage). KV Cache Utilization H200 and H100 show efficient cache usage with high hit rates (up to 99%) and low waiting requests. (The exception is code generation which has low hit rates (~1%)) A100_40G suffers from high KV cache usage and low hit rates, especially in classification and code generation, indicating memory bottlenecks. The strain on the inference server is observed by the higher number of waiting requests. Cost Efficiency Chat profiles: The A100 GPU (40G) offers the best value. Classification profiles: The H200 is most cost-effective. Code-generation profiles: The H100 provides the greatest cost efficiency. CPU vs GPU Llama 3.1 3B can run on CPU VM’s but the throughput and latency are so poor compared to GPU’s if does not make an practical or financial sense to do so. Smaller AI models (<= 1B parameters) may be OK on CPU’s for some light weight inference serves (like Chat). Conclusion The benchmarking results clearly demonstrate that hardware choice significantly impacts the inference performance and cost-efficiency of Llama 3.1 8B deployments. The H200 GPU consistently delivers the highest throughput and cache efficiency across workloads, making it the top performer overall. H100 follows closely, especially excelling in code generation tasks. While A100 and A100_40G offer budget-friendly options for chat workloads, their limitations in memory and cache performance make them less suitable for more demanding tasks. CPU virtual machines do not offer adequate performance—in terms of throughput and latency—for running AI models comparable in size to Llama 3.1 8B. These insights provide a practical foundation for selecting optimal infrastructure based on inference workload type and cost constraints. References Hugging Face Inference Benchmarker https://github.com/huggingface/inference-benchmarker Datasets used for benchmarking: Chat: hlarcher/inference-benchmarker/share_gpt_turns.json Classification: hlarcher/inference-benchmarker/classification.json Code Generation: hlarcher/inference-benchmarker/github_code.json Model: meta-llama/Llama-3.1-8B-Instruct on Hugging Face https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct vLLM Inference Engine https://github.com/vllm-project/vllm Azure ND-Series GPU Infrastructure https://learn.microsoft.com/en-us/azure/virtual-machines/nd-series PyTorch 2.7.0 + CUDA 12.8 https://pytorch.org NVIDIA GPU Drivers and NCCL Driver: 535.161.08 NCCL: 2.21.5-1 https://developer.nvidia.com/nccl Azure Pricing Calculator (South-Central US Region) https://azure.microsoft.com/en-us/pricing/calculator CPU - vLLM Appendix Install vLLM on CPU VM’s git clone https://github.com/vllm-project/vllm.git vllm_source cd vllm_source edit Dockerfiles (vllm_source/docker/Dockerfile.cpu) cp Dockerfile.cpu Dockerfile_serve.cpu change last line to “ENTRYPOINT ["/opt/venv/bin/vllm","serve"]” cp Dockerfile.cpu Dockerfile_bench.cpu change last line to “ENTRYPOINT ["/opt/venv/bin/vllm","bench","serve"]” Build images (enable AVX512 supported features (see lscpu)) docker build -f docker/Dockerfile_serve.cpu --build-arg VLLM_CPU_AVX512BF16=true --build-arg VLLM_CPU_AVX512VNNI=true --build-arg VLLM_CPU_DISABLE_AVX512=false --tag vllm-serve-cpu-env --target vllm-openai . docker build -f docker/Dockerfile_bench.cpu --build-arg VLLM_CPU_AVX512BF16=true --build-arg VLLM_CPU_AVX512VNNI=true --build-arg VLLM_CPU_DISABLE_AVX512=false --tag vllm-bench-cpu-env --target vllm-openai . Start vllm server Remember to set <YOUR HF TOKEN> and <CPU CORE RANGE> docker run --rm --privileged=true --shm-size=8g -p 8000:8000 -e VLLM_CPU_KVCACHE_SPACE=<SIZE in GiB> -e VLLM_CPU_OMP_THREADS_BIND=<CPU CORE RANGE> -e HF_TOKEN=<YOUR HF TOKEN> -e LD_PRELOAD="/usr/lib/x86_64-linux-gnu/libtcmalloc_minimal.so.4:$LD_PRELOAD" vllm-serve-cpu-env meta-llama/Llama-3.1-8B-Instruct --port 8000 --dtype=bfloat16 Run vLLM benchmark Remember to set <YOUR HF TOKEN> docker run --rm --privileged=true --shm-size=4g -e HF_TOKEN=<YOUR HF TOKEN> -e LD_PRELOAD="/usr/lib/x86_64-linux-gnu/libtcmalloc_minimal.so.4:$LD_PRELOAD" vllm-bench-cpu-env --backend vllm --model=meta-llama/Llama-3.1-8B-Instruct --endpoint /v1/completions --dataset-name hf --dataset-path AI-MO/aimo-validation-aime --ramp-up-strategy linear --ramp-up-start-rps 1 --ramp-up-end-rps 2 --num-prompts 200 --seed 42 --host 10.0.0.4Performance of Llama 3.1 8B AI Inference using vLLM on ND-H100-v5
Introduction The pace of development in large language models (LLMs) has continued to accelerate as the global AI community races toward the goal of artificial general intelligence (AGI). Today’s most advanced models boast trillions of parameters, pushing the boundaries of what machines can understand and generate. However, this scale comes at a steep cost—both in terms of training and inference—due to the immense GPU resources required to host and operate these models. Yet, innovation is not limited to those with access to the largest AI supercomputers. DeepSeek have demonstrated that it is possible to build highly competitive models without relying on the latest, most expensive infrastructure. At the same time, a renewed wave of open-source collaboration is challenging the closed-source strategies of leading AI companies, offering more accessible and customizable alternatives. For enterprise customers, the focus is shifting toward practical, cost-effective solutions. Rather than deploying trillion-parameter giants, many organizations are turning to smaller models—such as those with around 8 billion parameters—that strike a balance between accuracy and efficiency. These models are not only easier to fine-tune and deploy but also significantly reduce the cost per token, making them ideal for real-world business applications. In this paper, we explore the capabilities of the Llama 3.1 8B model as a representative example of a modern, enterprise-ready LLM. We benchmark its inference performance on Azure’s ND-H100 v5 infrastructure using the vLLM engine and present our findings along with recommendations for enterprise deployment. AI Inference architecture Inference in transformer-based large language models (LLMs) is typically divided into two primary stages: prefill and decode. Understanding the characteristics and resource demands of each stage is essential for optimizing performance, especially when deploying models like Llama 3.1 8B in enterprise environments. Prefill Stage: Compute-Bound Initialization The prefill stage is responsible for processing the input prompt. It involves tokenizing the input and performing a forward pass through the model to populate the key-value (KV) cache. This stage is compute-intensive, as it requires full attention computation across all input tokens. The performance bottleneck here is typically the GPU's compute throughput, especially for long prompts or large batch sizes. Decode Stage: Memory-Bound Token Generation Once the KV cache is populated, the decode stage begins. This stage generates one token at a time, using the cached context to predict the next token. The decode step is memory-bound, as it relies heavily on fast access to the KV cache. When the model achieves a KV cache hit, it can skip re-computation of earlier tokens, significantly reducing latency and compute load. This makes cache efficiency a critical factor in real-time inference performance. Fig 1. High level architecture of AI Inference, showing efficient use of KV cache can increase token throughput and reduce AI inference latency. Overall Inference Characteristics In general, AI inference is memory-bound, particularly during the decode phase. The ability to serve multiple concurrent requests efficiently depends on how well the system can manage memory bandwidth and cache locality. As such, optimizing memory access patterns and minimizing cache misses are key to achieving high throughput and low latency. Techniques for Optimization To maximize GPU utilization and token throughput while minimizing latency, several architectural strategies are employed: KV Cache Management: Efficient reuse and eviction policies to maintain high cache hit rates. vLLM uses PagedAttention (which is inspired by virtual memory and paging techniques used in operating systems), to manage the KV cache using blocks/pages. This allows vLLM to efficiently/dynamically utilize HBM memory and minimizes memory fragmentation. Batching and Scheduling: Grouping similar requests to improve parallelism and reduce idle GPU time. VLLM has a few parameters to control batching/parallelism. MAX_NUM_SEQS: How many input requests to process in parallel. MAX_NUM_BATCHED_TOKENS: The number of tokens to process in parallel (Forward pass in Deep neural network) Note: Larger values may not always be optimal; you could improve token throughput at the expense of latency. Weight and Activation Quantization (fp8): Reducing the precision of AI model weights and activations can give more memory to load AI models or have a larger KV cache. Lowering the precision also allows computations to be performed on more efficient GPU (higher FLOPS) computational units. Parallelization techniques: Tensor parallelism, Pipeline parallelism, Expert parallelism or Data parallelism can be used to split larger models across multiple Nodes/GPUs. Tensor parallelism distributes the model across the GPUs, with each GPU handling multiple layers of the model. Pipeline parallelism involves dividing the model, where a group of nodes (or GPUs) is responsible for processing its assigned DDN layer. Expert parallelism supports Mixture of Experts (MoE) models where different expert networks can be distributed across GPU’s. Data parallelism replicates the entire model across multiple GPU sets and processes different batches of requests in parallel Speculative Decoding: Predicting multiple tokens ahead to reduce the number of forward passes. Prefill/Decode Decoupling: Recent advancements, such as those implemented in vLLM (and NVIDIA Dynamo), decouple the prefill and decode stages, allowing each to be assigned dedicated GPU or CPU resources. This separation enables better resource allocation and parallelism, especially in multi-tenant or high-throughput environments. By leveraging these techniques, vLLM provides a highly efficient inference engine that is well-suited for serving modern LLMs like Llama 3.1 8B. This makes it a compelling choice for enterprise applications where cost, latency, and scalability are critical considerations. Benchmark environment Inference benchmark The Huggingface Inference benchmarking code was used for the AI Inference benchmark. Three different popular AI inference profiles were examined. Chat: Probably the most common use case, question and answer format on a wide range of topics. Classification: Providing various documents and requesting a summary of its contents. Code generation: Providing code and requesting code generation, e.g. create a new function. Profile Data set Input prompt Output prompt Chat hlarcher/inference-benchmarker/share_gpt_turns.json N/A min=50, max=800, variance=100 Classification hlarcher/inference-benchmarker/classification.json Min=8000, max=12000, variance=5000 Min=30, max=80, variance=10 Code generation hlarcher/inference-benchmarker/github_code.json Min=3000, max=6000, variance=1000 Min=30, max=80, variance=10 Huggingface Lama 3.1 8B models used Precision Model Size (GiB) meta-llama/Llama-3.1-8B-Instruct FP16 14.9 neuralmagic/Meta-Llama-3.1-8B-Instruct-FP8 FP8 8.4 nvidia/Llama-3.1-8B-Instruct-FP8 FP8 8.4 vLLM parameters Default value gpu_memory_utilization 0.9 max_num_seqs 1024 max_num_batched_tokens 8192 enable_chunked_prefill True enable_prefix_caching True Dynamo parameters Default values block-size 64 max-model-len 16384 kv_connector DynamoNixlConnector router round-robin remote-prefill True conditional-disagg True max-local-prefill-length 10 max-prefill-queue-size 2 max-num-batched-tokens 16384 Environment Local No-operation (Planner) True Results Fig 2: AI Inference performance comparison between the chat, classification and code generation profiles on ND-H100-v5 (8 H100). Profile Avg prompt throughput Avg generation throughput Max # Requests waiting Max KV Cache usage % Avg KV Cache hit rate % Chat ~7500 ~17600 0 ~2% ~78% Classification ~55300 ~2900 0 ~5% ~100% Code generation ~130400 ~1450 ~ 37 ~10% ~2% Fig 3: Show the impact of modifying MAX_NUM_BATCHED_TOKENS on the code-generation inference benchmark. (This parameter has a greater impact on the code generation benchmark compared to the chat/classification because of the low KV cache hit percentage.) Fig 4: Code generation inference benchmark run on 1 H100 showing the performance impact of using fp8 quantization. Fig 5: Code generation benchmark run on 1, 2, 4 & 8 H100. The results indicate that higher token throughput could be achieved running inference using 1 copy of model on each GPU instead of distributing model (via tensor parallel) amongst the 8 GPUs. Fig 6: Impact on throughput (tokens/sec) by adjusting AI Inference configuration (vLLM) on ND-H100-v5. Results comparing Dynamo vs traditional vLLM Fig 7: Dynamo vs traditional vLLM throughput (tokens/sec) comparison on 1 ND_H100_v5 (8 GPU’s). The best traditional vLLM configuration (8 x vLLM tensor_parallel=1) throughput performance is compared with various Dynamo configurations (i.e different combinations of GPU’s assigned for decode and prefill). Note: vLLM does have an experimental disaggregated prefill capability with various connector types. I attempted to run vLLM experimental disaggregated prefill using the kv-connector = LMCacheConnectorV1 (with Nixl enabled). I got mixed results, eventually running into the following issues (and deciding to switch to Nvidia Dynamo instead). Limited control over allocating GPU’s to decode vs prefill (used tensor parallel option, but limited by specific ratio with number of attention multi-heads). Memory management problems, got OOM errors even though there was plenty of HBM memory available on GPU’s (HBM was not distributed evenly amongst GPU’s). Analysis The performance of the inference prefill is determined by length and number of input prompts, this phase is compute bound (effectively a Matrix-Matrix operation) and we see that the code-generation profile does best primarily because it had the largest number of input tokens. The decode phase is a memory bound operation (effectively a Vector-Matrix operation) and performance in this phase is heavily dependent on the KV cache hit percentage. The code-generation profile only had ~1.7% KV cache hit percentage (there was plenty of HBM capacity, only ~10% of available KV Cache capacity was used), which resulted in slow decode performance impacting its overall throughput and latency especially at higher QPS (the code-generation benchmark was the only one which had requests being backed-up and waiting). The classification profile did well in the decode phase primarily due to the high KV cache hit percentage (~100%), it did struggle in overall throughput due to the small length of the output tokens. Adjusting the size of MAX_NUM_BATCHED_TOKENS had very little impact on the chat and classification benchmarks probably because they had high KV Cache hit percentages, but it did impact the performance of the code-generation benchmark (a ~3% improvement in tokens/sec using a larger value). Quantization of AI model can free up HBM memory to allow a large model to load or improve the AI inference performance by providing more memory for KV caching, it can also improve performance by performing computations with a higher FLOPS lower precision (e.g. FP8) format. In this case there is plenty of HBM and all three profiles did not use all the available KV cache space available, so FP8 quantization does not improve KV caching efficiency. Improvements in compute performance are observed with quantization, especially for the code generation profile which had a low KV cache hit percentage. The code generation tokens/sec on 1 GPU improved by ~38%. Since the Llama 3.1 8B model can easily fit on 1 H100, you can get significantly better total throughput (tokens/sec) on ND-H100-v5 if a complete model is loaded into each separate GPU instead of splitting the model across multiple GPU’s (via tensor parallel). The chat, classification and code generation inference throughput improved 4.2, 5.2, 1.9 times respectively. Newer inferencing server architectures feature disaggregated prefill, which allows you to decouple prefill from decode and assign resources (GPU’s, CPU’s) to each type of worker (prefill or decode). This is especially suited for large reasoning models with large context windows, running on large GPU inference clusters, significant performance gains have been reported. In this case we have a modest size (8B parameters) NLP model running on a single ND_H100_v5 node (only 8 GPU’s), so we were not expecting any significant performance improvements. The traditional aggregated vLLM was much faster than the best Dynamo configuration, running this inference benchmark on llama 3 8B model on ND_H100_v5. In this case the model can fit in a single GPU and the overhead of disaggregation might outweigh any parallelism gains when one GPU can already handle both phases efficiently. Conclusions When analyzing the performance of AI inference, it’s important not only to focus on the number of input and output tokens but also on the type of AI inference application, which can have a big impact on KV Cache effectiveness. Smaller AI models provide opportunities and more options to configure your environment in an optimal way to maximize token/sec throughput. References Welcome to vLLM — vLLM Mastering LLM Techniques: Inference Optimization | NVIDIA Technical Blog huggingface/inference-benchmarker: Inference server benchmarking tool meta-llama/llama-models: Utilities intended for use with Llama models. [2309.06180] Efficient Memory Management for Large Language Model Serving with PagedAttention ND-H100-v5 size series - Azure Virtual Machines | Microsoft Learn Hugging Face – The AI community building the future. Introducing NVIDIA Dynamo, A Low-Latency Distributed Inference Framework for Scaling Reasoning AI Models | NVIDIA Technical Blog Appendix Installation of hugging face inference benchmarker Install Rust curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh Build and install inference benchmarker cargo install --git https://github.com/huggingface/inference-benchmarker/ Installation of vLLM Install uv curl -LsSf https://astral.sh/uv/install.sh | sh Set-up python workspace uv venv --python 3.10 --seed source .venv/bin/activate Install vLLM uv pip install vllm --torch-backend=auto Install FlashInfer git clone https://github.com/flashinfer-ai/flashinfer.git --recursive pip install ninja cd flashinfer pip install --no-build-isolation --verbose . Start vLLM server #!/bin/bash NUM_GPUS=8 #NUM_GPUS=4 #NUM_GPUS=2 #NUM_GPUS=1 MODEL=meta-llama/Llama-3.1-8B-Instruct #MODEL=nvidia/Llama-3.1-8B-Instruct-FP8 #MODEL=neuralmagic/Meta-Llama-3.1-8B-Instruct-FP8 PORT=5000 export VLLM_LOGGING_LEVEL=INFO python -m vllm.entrypoints.openai.api_server --host 0.0.0.0 --port $PORT --model $MODEL --tensor-parallel-size $NUM_GPUS --dtype auto Run Inference benchmark TOKENIZER_NAME="meta-llama/Llama-3.1-8B-Instruct" #TOKENIZER_NAME="neuralmagic/Meta-Llama-3.1-8B-Instruct-FP8" #TOKENIZER_NAME="nvidia/Llama-3.1-8B-Instruct-FP8" inference-benchmarker \ --tokenizer-name $TOKENIZER_NAME \ --url http://localhost:5000 \ --profile code-generation -n Nvidia Dynamo installation Install python virtual environments sudo apt install python3.10-venv Create dynamo virtual workspace python3 -m venv mydynamo Activate virtual environment source /home/azureuser/mydynamo/bin/activate Check out dynamo github code. git clone https://github.com/ai-dynamo/dynamo.git cd dynamo git checkout $(git describe --tags $(git rev-list --tags --max-count=1)) Install dynamo prerequisites pip install "ai-dynamo[all]" Install some additional python modules and packages pip install tensorboardX pip install pynvml sudo apt install etcd systemctl status nats Restart etcd systemctl restart etcd Set-up nats-server wget https://github.com/nats-io/nats-server/releases/download/v2.10.22/nats-server-v2.10.22-linux-amd64.zip unzip nats-server-v2.10.22-linux-amd64.zip sudo mv nats-server-v2.10.22-linux-amd64/nats-server /usr/local/bin/ sudo chmod +x /usr/local/bin/nats-server Create/edit /etc/systemd/system/nats.service [Unit] Description=NATS Server After=network.target [Service] Type=simple ExecStart=/usr/local/bin/nats-server -js Restart=always RestartSec=10s LimitNOFILE=40000 [Install] WantedBy=multi-user.target export NATS_SERVER="nats://localhost:4222" Start Dynamo server/service cd $DYNAMO_HOME/examples/llm edit disagg.yaml file to modify parameters. dynamo serve graphs.disagg:Frontend -f ./configs/disagg.yaml Set Dynamo environmental variables export DYNAMO_HOME=$(pwd) cd $DYNAMO_HOME/examples/llmOptimizing Large-Scale AI Performance with Pretraining Validation on a Single Azure ND GB200 v6
Small performance gaps on a single virtual machine lead to large and costly performance losses at scale. Running small-scale pretraining jobs enables single-VM validation and allows for fine-grained identification of issues such as performance degradation, hardware bottlenecks, or software inefficiencies ahead of large-scale runs. In this post, we present a practical methodology for benchmarking the ND GB200 v6 virtual machines (VMs). A single ND GB200 v6 VM on Azure is powered by two NVIDIA Grace CPUs and four NVIDIA Blackwell GPUs. To ensure reliability in production workloads we used automated pretraining of lightweight Llama models with the NVIDIA NeMo framework. By systematically exploring and tuning key performance parameters and rigorously cross-checking results with detailed telemetry, we identify conditions that most significantly stress the GPUs. You can reproduce and reuse these pretraining workloads from our fully automated Azure AI Benchmarking guide.
Creating a Slurm Job Submission App in Open OnDemand with Copilot Agent
High Performance Computing (HPC) environments are essential for research, engineering, and data-intensive workloads. To efficiently manage compute resources and job submissions, organizations rely on robust scheduling and orchestration tools. In this blog post, we'll explore how to use Copilot Agent in Visual Studio Code (VSCode) to build an Open OnDemand application that submits Slurm jobs in a CycleCloud Workspace for Slurm (CCWS) environment. We'll start with a brief overview of CCWS and Open OnDemand, then dive into the integration workflow.Performance at Scale: The Role of Interconnects in Azure HPC & AI Infrastructure
Microsoft Azure’s high-performance computing (HPC) & AI infrastructure is designed from the ground up to support the world’s most demanding workloads. High-performance AI workloads are bandwidth-hungry and latency-sensitive. As models scale in size and complexity, the efficiency of the interconnect fabric—how CPUs, GPUs, and storage communicate—becomes a critical factor in overall system performance. Even with the fastest GPUs, poor interconnect design can lead to bottlenecks, underutilized hardware, and extended time-to-results. In this blog post, we will highlight one of the key enabling features for running large-scale distributed workloads on Azure: a highly tuned HPC-class interconnect. Azure has invested years of system-level engineering of the InfiniBand interconnect, into ready-to-use configurations for customers available on Azure’s HB-series and N-series virtual machine (VMs).
Introducing NVads V710 v5 series VMs
Cost-optimized AI inference, virtual workstations, and cloud gaming. AI inferencing and graphics-intensive applications continue to demand cost-effective, low power, high performance GPUs with more GPU memory and faster CPUs. Today we are thrilled to announce the General Availability of NVads V710 v5-series virtual machines (VMs) to expand our NV VM lineup to meet these very needs. As we mentioned in our previous preview announcement, customers running small-to-medium AI/ML inferencing workloads, Virtual Desktop Infrastructure (VDI), visualization, and cloud gaming workloads need NVads V710 v5-series VMs. Each VM is powered by an AMD Radeon™ Pro V710 GPU with up to 28 GB of GPU memory, which easily serves popular small open-source language models and handles the most demanding visualization scenarios. On top of this, the vCPU cores are backed by high-frequency 4th Generation AMD EPYC™ CPUs (3.9 GHz base and 4.3 GHz max frequency) for compute-intensive workloads that demand both CPU and GPU performance. Right size the VM for your workload needs. With NVads V710 v5 VMs, you only pay for the GPU/CPU compute you need. GPU partitioning capabilities enable customers to allocate fractions of the GPU – as small as 1/6 th of a V710 GPU – according to their workload requirements. This flexibility is ideal for customers that need to support a variety of inferencing and graphical workloads efficiently without requiring a full GPU for each application. The series provides several options ranging from 1/6 of a GPU with 4 GB of memory, perfect for lightweight virtual desktop experiences, to a full V710 GPU with a massive 28 GB for graphics intensive engineering applications or AI. Out of the box performance with ROCm and Radeon PRO Graphics. AMD Radeon PRO Graphics provides a seamless and reliable experience for visualization-focused workloads. The GPU and drivers are optimized or certified for all the major ISV solutions from vendors such as Adobe and Autodesk. They also support the latest ROCm releases and are designed to seamlessly integrate with popular machine learning frameworks like PyTorch, Triton, ONNX, and vLLM to serve small to medium language models. Upgrade to get up to 2.5x boost. Our own internal benchmarks of popular VDI, rendering, and visualization tests show NVads V710 v5 VMs are up to 2.5x faster when compared to NV v4 VMs. This means you can get more work done, faster, and have an overall better experience. Customer Momentum At Azure, we work closely with our partners and customers so they can take full advantage of these new VMs and accelerate their applications. Listen to what our partners at Dizzion, Cognata, and NAIO had to say. "The new Azure NVads V710 instances, powered by AMD Radeon Pro V710 GPUs, offer exceptional performance and flexibility at competitive prices. Dizzion Desktop as a Service customers delivering CAD, BIM, edge AI, and other high-performance workloads have eagerly awaited this addition to the market.” – Ruben Spruijt, Field CTO, Dizzion “In our experience, the V710 delivers excellent performance across both CPU and GPU workloads, making it a highly capable platform for a wide range of use cases. It offers a robust and reliable software stack, particularly well-suited for OpenGL and machine learning applications.” – Danny Atsmon, CEO, Cognata “We’ve tested the V710 thoroughly across a range of AI workloads, and the performance has really impressed us. It’s fast, stable, and scales well across different scenarios. It’s become a reliable, cost-effective part of our stack, and we’ll keep building on top of it as we expand our projects.” – Dr.-Ing. Timo Sämann, Chief AI Scientist, NAIO Product Details vCPUs 4th Generation AMD EPYC™ CPU. Configurations from 4 to 28 vCPUs (3.95 GHz base, 4.3 GHz max). Memory 16 GB to 160 GB GPU AMD Radeon PRO V710 GPU with 28 GB GDDR6 memory. 1/6, 1/3, 1/2, or full GPU. Storage Up to 1 TB temporary disk Networking Up to 80 Gbps Azure Accelerated Networking NVads V710 v5 VMs, now available in 5 Azure regions. We are happy to announce NVads V710 v5 VMs are now available in East US, North Central US, South Central US, West US, and West Europe. To learn more about our VMs and how to get started, please visit our documentation page.
