Chips

NVIDIA optimizes DGX Spark to speed up MoE inference

NVIDIA demonstrated three stacked optimizations to accelerate MoE inference on DGX Spark. On Ant Group's Ling-3.0-flash model, GSM8K decode throughput increased by 42.5% with no detectable change in accuracy. This helps improve single-unit AI inference efficiency.

Summarized by AI from one outlet

Source

Latest

  1. Chips

    NVIDIA Funds ASE to Expand CoWoS Advanced Packaging as TSMC Capacity Tightens

    Advanced packaging has become a critical bottleneck for AI chip production. NVIDIA is providing billions of dollars to support ASE's subsidiary Silan Precision to expand CoWoS capacity amid tightening supply from TSMC.

    IT之家

  2. Chips

    Intel rumored to halt Arc B580 for AI pro cards

    Industry reports suggest Intel is considering halting production of the Arc B580 gaming GPU to shift capacity to Arc Pro workstation cards. The move reflects the squeeze from AI edge inference hardware, with Intel aiming to capture higher margins in the enterprise market.

    T客邦

  3. Chips

    MediaTek AI ASIC revenue hits $2.5B amid investment recovery

    MediaTek's Q4 AI ASIC revenue is projected at $2.5 billion. CEO Rick Tsai noted that a transition period remains for investment recovery amid rising AI capex.

    科技新報

  4. Chips

    PNNL uses agentic AI to speed up radiation-resistant chip development

    The US Department of Energy's Pacific Northwest National Laboratory (PNNL) has implemented agentic AI to accelerate the development of radiation-resistant chips for satellites and nuclear reactors. This initiative aims to mitigate damage from high-energy particles in space and improve hardware reliability.

    科技新報

  5. Chips

    TSMC reportedly considering 3D DRAM production for AI hardware

    According to The Information, TSMC is planning to produce '3D DRAM' to meet the high-bandwidth needs of AI compute customers like NVIDIA. If materialized, this expansion into the memory market, currently led by Samsung and SK Hynix, would help alleviate supply pressures for AI accelerators.

    IT之家科技新報2 outlets

All news → · All rights belong to the original creators