See how the HPE HPC & AI performance engineering team helped the HPE Cray XD670 achieve #1 rankings in 30 scenarios, showcasing AI innovation with outstanding MLPerf benchmark results.
Weโre pleased to share new HPE leading results in the recently published MLPerf Inference: Datacenter v5.0 benchmark tests, providing third party validation of our strong leadership in AI innovation. Alongside a multitude of world records achieved by HPE ProLiant Gen12 offerings, we also obtained a remarkable #1 in 30 different scenarios with HPE Cray XD670, including in LLMs and Computer Vision1,2.
At the forefront of these achievements is the HPE HPC & AI performance engineering team, and their ability to optimize systems for peak performance. By leveraging HPE technology together with the expertise of our engineers, customers can achieve maximum efficiency and benchmarking excellence in diverse AI inferencing scenarios.
The power of people: HPE HPC & AI performance engineering
MLPerf benchmarks, developed by the MLCommons engineering consortium, provide an objective way to compare performance across multiple platforms and diverse AI use cases. Considering the technology foundation is the same across comparable submissions, the superior benchmark results achieved by HPE are a testament to the HPE HPC & AI Performance Engineering team. These experts demonstrated a detailed understanding of system capabilities, architectural and storage nuances, and fine-tuning resulting in maximum performance in the systems customers need to develop and run AI models.
In addition to leading HPEโs MLCommons performance strategy and the MLPerf benchmark process, this team of expert engineers assists customers with sizing AI workloads, and tuning individual applications and workloads, with the goal of optimizing performance throughout the life of the systems. This enables customers to extract maximum value from their environments, which is particularly important in the AI space today, considering the rising costs of GPUs and their ever-shorter lifecycles.
Leading benchmark results
In the latest MLPerf Inference: Datacenter v5.0 benchmark, which delivers machine learning (ML) system performance benchmarking in an architecture-neutral, representative, and reproducible manner3, HPE Cray XD670 featuring 8 NVIDIA H200 SXM GPUs delivered the highest throughput per node in the HPE portfolio submitted for testing, and was a top performer compared to other systems featuring the NVIDIA H200-SXM GPUs including:
- #1 in Computer vision including image classification, object detection, and 3D medical imaging with models Resnet501 , Retinanet1, 3D-Unet2 respectively.
- #1 Recommender delivering 8% better results than comparable platforms with models DLRMv2 99.0, 99.9 - Server & Offline2. These models show how personalized results on social media or ecommerce sites are provided by analyzing interactions between users and items such as products or ads.
- #1 in Large Language Models delivering 7% better results than comparable platforms with GPT-J 6B1 and Llama2-70B1 -Interactive. These represent deep learning algorithms trained on extensive datasets capable of recognizing, summarizing, Q&A, translating, predicting, and generating content for various applications.
Notably, the HPE Cray XD670 results more accurately reflect real-world storage performance compared to other vendors, who conducted their tests using local storage disks. Unlike these, the HPE Cray XD670 benchmarks were performed with external storage solutions โ including Cray ClusterStor E1000 and HPE GreenLake for File Storage โ which are more representative of the storage environments typically found in large, clustered implementations.
Advancements in AI-optimized servers
In addition to HPE Cray XD670, HPE offers a range of other purpose-built, 8-way GPU servers designed and optimized to accelerate large AI model training and inferencing. These are ideal for service providers and large enterprises building and training their own large AI models and come with cooling options including air-to-liquid and HPEโs industry-leading direct liquid cooling. For example, we recently introduced the HPE ProLiant Compute XD685, featuring 8 NVIDIA Blackwell (B200) or 8 NVIDIA H200 SXM GPUs, and we look forward to submitting MLPerf benchmark results with this platform in future tests.
For more detailed information about the MLPerf Inference: Datacenter v5.0 benchmark results, explore the MLCommons results interactive table.
If you want to learn more about HPE ProLiant Compute XD685 follow this link, or contact your HPE representative.
Diana has spent the past 27 years working with the technologies that power the worldโs most demanding IT environments and is interested in how solutions based on those technologies impact the business and the world. A native from Colombia, Diana holds an MBA from Georgetown University and has held a variety of regional and global roles with HPE in the US, the UK and Sweden. She is based in Stockholm, Sweden. Connect with Diana on LinkedIn.
[1] Source: โMLPerf Inference: Datacenter v5.0 resultsโ MLCommons, April 2025 โ Top performance for Resnet50-Offline, Retinanet-Offline, GPTJ 99.0-Offline, GPTJ 99.9-Offline, Llama2-70B-Interactive 99.0-Server, Llama2-70B-Interactive 99.9-Server compared to other H100-SXM systems (submission IDs: 5.0-0039, 5.0-0040), and 7% better in Llama2-70B-Interactive 99.0/99.9 Server scenarios compared to other H100-SXM results
[2] Source: โMLPerf Inference: Datacenter v5.0 resultsโ MLCommons, April 2025 โ HPE Cray XD670 is #1 top-performing server for 3D medical imaging (3D-Unet 99 and 99.9), Recommenders (DLRMv2 99.0 & 99.9 Offline/Server) for all H200-SXM and H100-SXM systems (submission IDs: 5.0-0039, 5.0-0040, 5.0-0041), and 8% better in DLRMv2 99.0-Server scenario compared to other H100-SXM results
[3] Source: MLCommons Releases New MLPerf Inference v5.0 Benchmark Results, MLCommons, April 2025
ComputeExperts
Our team of Hewlett Packard Enterprise server experts helps you to dive deep into relevant infrastructure topics.