Global GPU Computing News【20250906】

China's First Open AI Computing Architecture Released: This architecture, co-released by Sugon and over 20 industry chain companies (covering AI chips, hardware, large models, etc.), is a collaborative innovation system for large-scale intelligent computing scenarios, featuring highly efficient, tightly-coupled system design centered around GPUs. It aims to promote the comprehensive opening and inclusive development of the domestic intelligent computing ecosystem.

Technical Features of Sugon's AI Super Cluster System:

• Super High Performance: A single cabinet can hold 96 GPU cards, computing power reaches the hundred-Petaflop level, total memory access bandwidth exceeds 180 TB/s, and supports expansion to ultra-large clusters with millions of accelerator cards.

• Super High Efficiency: Its thousand-card cluster performance for large model training and inference reaches 2.3 times the mainstream industry level, development efficiency improves 4x, and can increase GPU computing efficiency by 55% through compute-storage-network synergy.

• Super High Reliability: Includes 121 reliability designs for equipment and links, Mean Time Between Failures (MTBF) increased by 2.1x, Mean Time To Repair (MTTR) reduced by 47%.

• Fully Open: Hardware supports multi-brand AI accelerators; software is compatible with mainstream AI computing ecosystems (e.g., CUDA).

AI Computing Open Architecture Joint Laboratory Established: The lab, leveraging the National Advanced Computing Industry Innovation Center, will focus on areas like chip and underlying basic software co-optimization, and model-hardware co-design, aiming to break down ecosystem barriers.

Sugon Opens Three Key Technologies: Initially announced the opening of three key technical capabilities: AI storage optimization, liquid cooling infrastructure design specifications, and the DeepAI basic software stack, to lower the R&D threshold for SMEs and reduce redundant development.

Outstanding Energy Efficiency of Sugon's AI Super Cluster: Utilizes advanced cold plate liquid cooling technology and 394 energy-saving designs, achieving a Power Usage Effectiveness (PUE) below 1.12.

China's Intelligent Computing Industry Seeks Open Collaboration to Break Computing Bottlenecks: The industry recognizes challenges such as insufficient supply of high-end computing power, performance gaps of domestic accelerators, high computing costs, and immature independent software/hardware ecosystems. Open collaboration is needed to build open, standardized, and highly efficient computing clusters.

Google Ironwood TPU Performance Leap: Google disclosed details of its TPU v6 (Ironwood) at Hot Chips 2025. Its peak FLOPS performance improved ~10x over TPU v5p, power efficiency improved 5.6x, equipped with 192GB HBM3E memory and 7.3 TB/s bandwidth. Its power efficiency of 4.2 TFLOPS/Watt is close to NVIDIA's B200/300 GPUs (4.5 TFLOPS/Watt), indicating AI-specific chips are rapidly closing the performance gap with leading GPUs.

Meta Expands Ultra-Large GPU Clusters: Meta plans to expand its GPU cluster scale to over 100,000 GPUs, expecting 10x growth in the next decade. It detailed the architecture of its custom NVL72 system 'Catalina' at Hot Chips, which differs from NVIDIA's standard design, reflecting the trend of customization based on specific needs.

Networking & Optical Tech Focus for AI Infrastructure:

• Networking Tech: Broadcom introduced the 51.2 Tb/s Tomahawk Ultra switch focused on low-latency Scale-Up for HPC and AI. NVIDIA also introduced "Spectrum-XGS" Ethernet technology to address "cross-scale" issues in distributed clusters across data centers.

• Optical Tech: Companies like Lightmatter and Ayar Labs showcased advanced photonic interconnects and optical I/O chip technologies to address copper interconnect limitations, rising rack power density, and challenges in optical transceiver cost and power consumption.

AMD Discloses MI350 Details, Previews MI400: AMD introduced the MI350 GPU series at Hot Chips. The MI355X has a TBP of 1.4kW, primarily for liquid-cooled data centers; the MI350X has a TBP of 1.0kW, mainly for air-cooled infrastructure. AMD reaffirmed that the MI400 series and its "Helios" rack solution are on track for 2026.