Revolutionizing AI Computing: Liqid’s UltraStack and Meta’s VR Project
Meta (formerly Facebook), known for its dedication to adopting cutting-edge technologies, has found a solution to their virtual reality research project’s need for greater GPU-driven compute power. The project faced limitations due to the IT industry’s common roadblocks of limited GPU server density and availability. However, Meta turned to Liqid, a pioneer in composable disaggregated infrastructure, to overcome these challenges.
Liqid introduced the UltraStack server solution, featuring 2U Dell PowerEdge servers equipped with up to 20 NVIDIA L40S GPUs. This innovative solution not only meets the demand for high GPU density but also significantly improves performance. UltraStack has the potential to accelerate next-generation workloads, delivering a remarkable 35 percent increase in performance. Additionally, it offers cost savings with a 35 percent reduction in power consumption and a staggering 75 percent reduction in software licensing costs compared to servers with lower GPU capacities.
Vasu Agrawal, a staff research engineer at Meta, sheds light on the impact of Liqid UltraStack with NVIDIA L40S GPUs and Dell PowerEdge servers in a video interview. He emphasizes how this infrastructure has transformed their VR project by enhancing the compute capabilities and driving significant advancements.
The key advantage of UltraStack lies in its ability to seamlessly connect a large number of GPUs to a standard server. By doing so, it effectively addresses three critical challenges in AI computing: cost and complexity, time, and performance.
Firstly, UltraStack eliminates the cost and complexity associated with inadequate GPU density. Many AI deployments struggle with a limited number of GPUs per server, resulting in increased power, cooling, and space costs, as well as system management complexities. However, UltraStack solves this problem by incorporating up to 20 GPUs per server.
Secondly, AI infrastructure implementation often faces delays due to the limited availability of high-performance GPUs. UltraStack’s integration of readily available NVIDIA L40S GPUs ensures timely and efficient deployment.
Lastly, UltraStack delivers exceptional performance by consolidating up to 20 GPUs per server. Test results using MLPerf 3.1 Inference LLM (Large Language Model) demonstrated remarkable query-per-second rates across various domains. For instance, with 16x L40S GPUs, UltraStack achieved an impressive 94 queries/sec at ~7,000W for MLPerf 3.1 Inference LLM. Moreover, UltraStack showcased its prowess in Object Detection, Medical Imaging, and Natural Language Processing, achieving 10,104 queries/sec, 62.6 queries/sec, and 44,730 queries/sec respectively.
Liqid offers UltraStack in both stand-alone and cluster-ready server options. Stand-alone servers house up to 20 GPUs, while cluster-ready servers incorporate up to 16 L40S GPUs, along with Liqid IO Accelerator NVMe SSDs, NVIDIA ConnectX-7 NICs, and NVIDIA BlueField-3 DPUs for enhanced connectivity and efficiency in complex AI environments.
Liqid’s UltraStack, coupled with Meta’s commitment to AI technology adoption, signals a significant leap forward in the domain of AI computing. By addressing crucial challenges and providing exceptional performance, UltraStack paves the way for accelerated AI advancements and impactful virtual reality experiences.
Additional facts relevant to the topic include:
– Liqid’s UltraStack server solution is based on composable disaggregated infrastructure (CDI), which allows for the dynamic composition and reconfiguration of compute, storage, and networking resources. CDI enables organizations to optimize resource allocation, improve flexibility, and streamline infrastructure management.
– Meta’s virtual reality research project aims to create immersive experiences for users by leveraging advanced technologies such as AI and computer vision. The increased GPU compute power provided by UltraStack enhances the project’s ability to render realistic and interactive virtual environments.
– The partnership between Meta and Liqid demonstrates the importance of collaboration between technology companies in pushing the boundaries of AI computing and virtual reality. By working together, they were able to overcome the limitations of existing infrastructure and drive innovation in their respective fields.
The most important questions related to the topic are:
1. How does UltraStack’s GPU density compare to traditional server configurations?
2. What are the specific performance improvements achieved with UltraStack in different AI workloads?
3. How does UltraStack address the challenges of cost and complexity in AI computing?
4. What are the potential applications of the enhanced GPU compute power in Meta’s virtual reality research project?
5. How does the integration of Liqid IO Accelerator NVMe SSDs, NVIDIA ConnectX-7 NICs, and NVIDIA BlueField-3 DPUs contribute to the efficiency of cluster-ready servers?
Key challenges or controversies associated with the topic include:
1. Scalability: While UltraStack addresses the limitations of GPU density in traditional server configurations, scaling the infrastructure to accommodate larger AI workloads may still pose challenges. Organizations may need to carefully plan their infrastructure deployment and consider factors like power and cooling requirements.
2. Cost: While UltraStack offers cost savings in terms of power consumption and software licensing, the upfront investment in high-density GPU servers can still be significant. Organizations need to assess the long-term cost-effectiveness of such solutions and the specific benefits they provide for their AI workloads.
Advantages of Liqid’s UltraStack and Meta’s VR Project:
1. Increased compute power: The integration of UltraStack’s high-density GPU servers significantly enhances the compute capabilities of Meta’s virtual reality research project. This increased power allows for the rendering of more complex and realistic virtual environments, leading to more immersive VR experiences.
2. Cost savings: UltraStack offers cost savings through reduced power consumption and software licensing costs. This can be particularly beneficial for organizations with large-scale AI deployments that require a high number of GPUs.
Disadvantages:
1. Upfront investment: Deploying UltraStack and similar high-density GPU solutions requires a significant upfront investment. Organizations need to carefully evaluate the financial implications and potential return on investment before adopting such infrastructure.
2. Scalability limitations: While UltraStack addresses the limitations of GPU server density, there may still be scalability challenges when it comes to accommodating larger AI workloads. Organizations need to consider the future growth and scalability requirements of their AI infrastructure.
