Global GPU Computing News【20251206】

Amazon Releases Trainium 3 AI Chip
Amazon AWS launched its new-generation, self-developed AI training chip, Trainium 3, on December 2. Manufactured using a 3nm process, the chip delivers 2.52 PFLOPs of FP8 compute performance per unit. The Trn3 UltraServer equipped with these chips can achieve a total compute power of 362 PFLOPs, with a 40% improvement in energy efficiency compared to the previous generation.

NVIDIA GB200 NVL72 System Shows 10x Performance Gain
NVIDIA released data on December 3 indicating that its GB200 NVL72 system, which integrates 72 Blackwell GPUs, delivers approximately a 10x improvement in inference performance for open-source MoE models like DeepSeek-R1 and Mistral Large 3 compared to the previous-generation H200 platform.

NVIDIA and Synopsys Expand Strategic Collaboration
On December 1, NVIDIA and Synopsys announced an expansion of their strategic partnership to jointly advance the application of GPU-accelerated computing in engineering fields such as chip design, physical simulation, and digital twins. NVIDIA also invested $2 billion to subscribe to Synopsys common stock.

UCloud Releases New-Generation GPU Virtualization Technology
On December 1, UCloud officially launched its new-generation GPU virtualization technology. It supports splitting a single GPU's memory and compute power along two dimensions, with a minimum granularity as fine as 10%. This allows multiple AI workloads to run concurrently, improving resource utilization.

Google TPU Be Available for External Supply, Challenging NVIDIA
Google's self-developed Tensor Processing Unit (TPU) recently announced its availability for external supply, seen as a direct challenge to NVIDIA's GPU dominance. Market analysts note that TPU performance is now comparable to GPUs, which could potentially reshape the computing landscape.

Alibaba Cloud's Aegaeon Selected for SOSP, Cuts GPU Usage by 82%
Alibaba Cloud's computing pooling solution, "Aegaeon," was selected for the prestigious academic conference SOSP 2025. In beta testing, the solution reduced the number of H20 GPUs required to serve a 72-billion-parameter large model from 1,192 to just 213, representing an 82% reduction.

Amazon Unveils New-Generation AI Chip Trainium3
According to a December 4 report, Amazon AWS introduced its self-developed Trainium3 AI chip, claiming it offers four times the computational performance of its predecessor while consuming 40% less energy. It is also said to reduce training and operational costs by up to 50% compared to systems using NVIDIA GPUs.

Moore Threads Lists on STAR Market, Becomes First Domestic GPU Stock
On December 5, Moore Threads officially listed on the Shanghai Stock Exchange's STAR Market, becoming China's first listed full-featured GPU company. Its products integrate AI computing, graphics rendering, and physical simulation, supporting 10,000-card-level AI computing clusters.

LISUAN Tech Launches First Domestically Developed 6nm GPU
On December 4, LISUAN Technology launched its first 6nm GPU series, "7G100," based on a self-developed architecture, available in 12GB and 24GB models. The chip was developed without using foreign IP licenses, achieves performance at the international mainstream level, and supports the Windows on Arm platform.

Memory Tensor and SenseTime's AI Infrastructure Achieve System-Level Breakthrough in Domestic GPGPU
On December 4, Memory Tensor, in collaboration with SenseTime's AI infrastructure, successfully ran the industry's first commercially viable "Memory-Compute-Scheduling" integrated PD-separated inference cluster on domestic GPGPUs. Under real workloads, the cluster's throughput increased by 75%, and its comprehensive inference cost-performance reached 150% of the same-generation NVIDIA A100.