Nvidia H200

Headlines · 5

1. 5d
   Velox: spatial memory router speeds up video world models

   A paper on arXiv introduces Velox, a training-free attention router that selects context at the token-block level using camera frustum slab correspondences from streaming geometry, aiming to cut the growing KV-cache cost of chunked video diffusion world models as rollouts lengthen. The authors report a 4.36x attention speedup and 2.43x denoising speedup at chunk 48, and a gap of within 0.15 average points versus the dense LingBot-Fast-Cam baseline across the full 158-case WBench navigation split. The work is by Qiuhong Shen, Shizun Wang and Xinchao Wang; all figures are author-reported and have not been independently verified.

   Hugging Face · Papers · 🔥 0
2. 6d
   Mixture-of-Experts kernel engineering on Amazon SageMaker AI

   Mixture-of-Experts (MoE) is a sparsely-activated architecture that scales large language model (LLM) capacity while keeping per-token compute low. Because its operations account for a large share of total compute time, inefficiencies there directly raise request latency and cost. Therefore, optimizing MoE operations through kernel engineering offers the largest opportunity for performance gains. Fused MoE is a key optimization technique for the MoE architecture. A fused MoE system merges token routing, data shuffling, and specialized network calculations into a single batched operation on the

   AWS Machine Learning Blog · 🔥 0
3. 6d
   Agnes open-sources 28.7B-parameter MoE video model

   Agnes has open-sourced Agnes-Video-2.0-MoE, an audio-video generation model built by converting its roughly 15.3B-parameter dense pretrained model DaVinci-MagiHuman into a mixture-of-experts design: most of the structure is kept, while parts of 32 of its 40 layers are each expanded into three expert networks, with routing selecting one expert per step so activated parameters stay near 15.3B while total parameters reach about 28.7B. Agnes says the model generates a 5-second video in roughly 30 seconds on a single H200 GPU — about 20% faster inference than the dense version and more than twice as fast as comparable MoE video models. The team also applies EENP elastic expert neuron pruning, first scoring expert computation channels separately for video, audio and text and then merging the scores to cut channels; at 60% pruning, average activated parameters fall to about 11.7B and denoising time drops from 26.83s to 20.92s (about 22% less), though the VBench-2.0 score is about 92.4% of the unpruned MoE, with details degrading more in high-motion scenes as pruning deepens.

   品玩 实时要闻 · 🔥 0
4. Oct 11
   Biren plans HK$4.02B placement, total raise hits HK$17.26B

   Biren Technology announced on Oct. 8 a placement of 130 million H shares at HK$31.8 apiece to no fewer than six placees, about 4.76% of post-issue share capital, for estimated net proceeds of roughly HK$4.02 billion. The price was a 9.76% discount to the previous close, and the stock fell 11.85% to HK$30.36 that day, closing at HK$31 on Oct. 9 for a market value of about HK$80.6 billion. Including a roughly HK$6.204 billion net raise at its January listing and a HK$7.038 billion placement in July, Biren has raised about HK$17.261 billion (about RMB 14.7 billion) since listing, the most among listed Chinese GPU makers. The company says about 70% of this round goes to strategic supply-chain procurement, production preparation and commercialization for its next-generation flagship BR20X, which it benchmarks against NVIDIA's H200 and expects to enter post-tape-out validation and customer testing in Q4, with large-scale deployment in 2027.

   钛媒体 · 🔥 0
5. Oct 10
   Huawei says Ascend tops Nvidia in China share as supply stays tight

   According to a media briefing summary published by Jiemian, Huawei rotating chairman Xu Zhijun said Ascend is still far from meeting domestic demand and that Huawei has no plan to expand broadly overseas, urging the supply chain to "go faster, produce more"; he said Huawei's own data suggests Ascend's China market share has surpassed Nvidia's, though the note gave no figure or methodology. Nvidia said in its quarterly report that H200 shipments to some Chinese customers permitted since February accounted for less than 1% of data center revenue in the quarter, and it excluded China data center compute revenue from next-quarter guidance. The article also reports intent prices for Ascend 950DT accelerator cards up 20%-50% in two months to over 250,000 yuan and 950PR up from about 60,000 yuan early this year to over 80,000 yuan (intent quotes, not final prices), and that DeepSeek published Ascend-targeted software components including TileKernels and DeepGEMM-Ascend on September 30.

   虎嗅 AI · 🔥 0

Experience and discussion from the community

Share my Nvidia H200 experienceAsk about Nvidia H200