Chinese AI Firms Suspected of Using NVIDIA Blackwell Chips for Model Development
AI research group Unemployment Labs analyzed multiple Chinese AI models and detected traces of training on restricted NVIDIA Blackwell GPUs. The findings suggest Chinese AI companies continue to use high-end chips for model development despite export controls, highlighting challenges in enforcing chip regulations.
Summarized by AI from 3 outlets
Source
- GIGAZINERead original ↗
- THE DECODERRead original ↗
- TechCrunchRead original ↗
Tools mentioned
- ContinueCoding · ★ 36.5k
Chat, inline completion and an agent mode in VS Code and JetBrains, with models and rules defined in config files you control.
Latest
- Chips
Korean firms lag behind gigawatt data center plans
Korean companies like KT and SK Telecom have announced gigawatt data center plans, but current capacity is only in the megawatt range. The gap may hinder customer acquisition and delay expansion, impacting AI compute growth.
- Chips
Google Releases TPU Sync for KV Cache Transfer
Google released TPU Sync, an open-source C++ library transferring KV caches and weights across TPU and host memory. It reportedly cuts transfer latency by over 90% on large MoE models.
- Chips
Users Run 2.8-Trillion Parameter Kimi Model on Optane Memory
Users demonstrated running the 2.8-trillion parameter Kimi K3 model using discontinued Intel Optane memory and Xeon processors without a GPU. Inference speed is around 1.5 tokens per second.
- Chips
Ubitus Donates GPU Servers to National Central University to Alleviate AI Compute Shortage
Ubitus donated AI GPU server equipment to National Central University, hosted by Professor Tzong-Han Tsai in the Department of Computer Science and Information Engineering. The initiative aims to bridge the AI compute gap between academia and industry in Taiwan to support research.
All news → · All rights belong to the original creators