RUST, Germany, March 19, 2025 /PRNewswire/ -- Compal Electronics (Compal; Stock Ticker: 2324.TW) makes a grand debut at CloudFest 2025 (March 17–20, 2025) with the launch of its groundbreaking SG720-2A/OG720-2A AI server—a milestone that marks a major breakthrough in the convergence of high-performance GPU Accelerator and two-phase direct-to-chip (DTC) liquid cooling technology.
The SG720-2A/OG720-2A 7U AI server not only supports 8x AMD Instinctâ„¢ MI325X GPUs —built on the advanced CDNA 3 architecture up to 896GB/s Infinity Fabricâ„¢ bandwidth, which achieves a theoretical FP16 performance of approximately 20.9 PFLOPs with sparsity and exceeds 41.8 PFLOPs with sparsity in FP8 computations—but also incorporates total 2TB of high-speed HBM3E memory in the platform with up to 48TB/s of memory bandwidth, delivering unparalleled compute power and ultra-fast data access for large-scale AI inference and data processing. Additionally, it is equipped with the ZutaCore® HyperCool® 2-Phase DLC liquid cooling solution, and the synergistic integration of these technologies elevates the SG720-2A/OG720-2A's performance and operational stability to unprecedented levels.
With an optimized 850mm chassis depth, the SG720-2A/OG720-2A seamlessly integrates into a standard EIA 19" 1-meter-deep rack or Open Rack v3-compatible setup. It enhances power efficiency and simplifies management with front-facing, hot-pluggable I/Os, making it a scalable and high-performance solution for data centers.
Its modular architecture facilitates straightforward maintenance and upgrades, with independent component replacement that minimizes downtime, while hot-swappable PCIe slots offer the crucial advantage of minimizing downtime and maximizing efficiency by allowing for network card replacement or upgrades without shutting down the entire system.
Compal recently conducted an experiment to evaluate the efficiency of direct-to-chip liquid cooling methods, revealing that ZutaCore's HyperCool technology achieved an impressively low partial Power Usage Effectiveness (pPUE) of 1.01. In comparative testing against single-phase liquid cooling, HyperCool consistently outperformed across all test conditions with 5% lower pPUE, demonstrating its superior efficiency and effectiveness in optimizing thermal management. This benchmark cements HyperCool as a game-changing solution for sustainable, high-performance data centers.
At CloudFest 2025, Compal, in partnership with ZutaCore®, showcases cutting-edge core technologies and solutions for the data centers of the future. In addition to showcasing the all-new SG720-2A/OG720-2A, other leading-edge server technologies are also on display. Interested visitors are warmly invited to visit the exhibit and explore the latest advancements in AI and HPC technologies.
CloudFest 2025 – Compal Booth
Dates: March 17–20, 2025
Location: Europa-Park, Germany
Booth Number: R07
About Compal
Founded in 1984, Compal is a leading manufacturer in the notebook and smart device industry, creating brand value in collaboration with various sectors. Its groundbreaking product designs have received numerous international awards. In 2024, Compal was recognized by CommonWealth Magazine as one of Taiwan's top 6 manufacturers and has consistently ranked among the Forbes Global 2000 and Fortune Global 500 companies. In recent years, Compal has actively developed emerging businesses, including cloud servers, auto electronics, and smart medical, leveraging its integrated hardware and software R&D and manufacturing capabilities to create relevant solutions. More information on Compal server, please visit https://www.compalserver.com/
** The press release content is from PR Newswire. Bastille Post is not involved in its creation. **
Compal at CloudFest 2025: SG720-2A/OG720-2A AI Server with ZutaCore Pioneers a New Era of High-Efficiency, Energy-Saving Data Centers!
SAN JOSE, Calif., March 19, 2025 /PRNewswire/ -- Manycore Tech Inc., a fast-growing spatial intelligence company, announced that it has made its multimodal spatial comprehension model, SpatialLM, open-source at the GTC 2025, significantly lowering barriers for training embodied intelligence.
Manycore Tech is the world's largest spatial design platform, with an average of 86.3 million monthly active visitors. It has served more than 45,500 enterprise customers across over 200 countries and regions.
SpatialLM serves as a powerful model that teaches robots how to understand the surrounding 3D environment. It provides a basic training framework for practitioners in the embodied intelligence sector, allowing them to fine-tune the model to fit into specific scenarios as needed.
Global developers can access the open-source SpatialLM on platforms, including Hugging Face, GitHub and ModelScope, to power their research and development of embodied intelligence.
SpatialLM overcomes the limitations of traditional large language models to enhance the machines' ability of understanding real-world geometry and spatial relationships, much like humans do. It can automatically generate structured 3D scenes from videos, based on point cloud data extracted from those videos. It is able to accurately recognize and understand these scenes and translate them into 3D structural layouts.
In the future, SpatialLM will be iterated to be able to handle a wider range of tasks, such as interacting with humans as an intelligent assistant and supporting embodied agents to perform complex tasks in challenging environments.
"We hope to create an embodied intelligence training platform that can support the entire development circle from spatial cognition and analysis to machine-environment interactions," Victor Huang, Chairman of Manycore Tech, said in an interview. "By open-sourcing SpatialLM, we aim to facilitate foundational spatial cognition training for embodied intelligent robots. Meanwhile, we hope that SpatialVerse, the spatial intelligence solution released last year, can enable action and interaction training for robots in simulated environments leveraging synthetic data solutions."
SpatialVerse, is another highlight for Manycore Tech at this year's GTC. It enables developers to train content-generation models in virtual settings and enhance cognitive capabilities of intelligent robots, AR/VR systems, and embodied AI. The platform, along with SpatialLM, provides a versatile system for digital simulations for developing embodied intelligence.
For instance, the SpatialLM transforms real-world data into an abundance of digitally-structured scenes. Such scenes can then be generalized to create as many as trillions of new scenes using the synthetic data engine of SpatialVerse. That enriches the training data while maintaining a high quality of the dataset, which is key to embodied intelligence.
"I believe we'll soon see explosive growth in embodied intelligence, driven by advancements in computing power, algorithms, engineering and training data. By open-sourcing SpatialLM, we hope to contribute to the development of foundational technologies, and help push the boundaries of AI and accelerate the arrival of the singularity." said Mr. Huang.
Manycore Tech has reached cooperation agreements regarding spatial and embodied intelligence training with a number of global embodied intelligence companies, including some leading players in Silicon Valley.
About Manycore Tech Inc.
Founded in 2011 in Hangzhou, Manycore Tech Inc. is a spatial intelligence company powered by artificial intelligence (AI) technologies and purpose-built graphics processing unit (GPU) clusters. Manycore Tech offers a full suite of products, including the spatial design software Kujiale, its international version Coohom, and SpatialVerse, the next-generation spatial intelligence solution for AI development in indoor environments. As of 2024, Manycore Tech has amassed an average of 86.3 million monthly active visitors. It has served over 414,000 individual customers and 45,500 enterprise customers across more than 200 countries and regions.
About SpatialVerse
SpatialVerse is a next-generation spatial intelligence solution for AI development in indoor environments. At its core lies our massive, physically accurate dataset library specifically designed to train sophisticated models through realistic virtual simulations. Users can conduct industrial-scale simulations with multi-sensor compatibility and achieve high-fidelity RTX rendering aligned with NVIDIA Isaac Sim's OpenUSD framework. This technology, which bridges digital simulations and physical reality, accelerates AI development while reducing real-world testing costs.
** The press release content is from PR Newswire. Bastille Post is not involved in its creation. **
Manycore Tech Makes SpatialLM Open-source to Empower Embodied Intelligence Training
