Today for AI
HOT RADAR
computeHEAT 21.5°

AI CURATED HOT STORY · 10/29/2026

Iluvatar CoreX Launches TG300, Surpassing NVIDIA H200 in Key Inference Tasks

2 reports archived2 independent sourcesupdated 10/29/2026, 17:09:43
Synthesis & Latest Updates
2 Sources Cross-Validated

Iluvatar CoreX has officially launched the TG300 cloud GPU based on its 'Tianshu' architecture and the TY1200 edge module, marking a shift from roadmap promises to actual delivery. In live tests, the TG300 achieved text-to-image inference speeds of 95.6 seconds, outperforming NVIDIA's H200 at 112 seconds. Third-party data indicates that under PD-separated deployment for large models, the TG300's single-card end-to-end throughput reaches 1.05 times that of the H200.

Technically, the TG300 excels in specific scenarios such as long-context processing, MoE models, and online recommendations. For 64K long contexts, its Attention performance and throughput hit 1.05x and 1.15x of the H200, respectively. The accompanying TY1200 edge product demonstrated faster robot task execution in offline environments compared to NVIDIA's Jetson Thor T5000, with a next-gen TY2000 (600T compute) also previewed.

This release highlights domestic computing power's competitiveness in specific inference tasks and establishes a full-chain ecosystem through a new software stack. By offering a viable alternative to NVIDIA, particularly in cloud-edge collaboration, Iluvatar aims to reshape the industry landscape and accelerate the autonomy of China's AI infrastructure.

LATEST/Iluvatar CoreX has opened live demos of the TG300, confirming its end-to-end throughput reaches 1.05x that of the H200 in typical PD-separated deployments. The company also previewed the next-gen edge product TY2000 with 600T nominal compute, though specific configurations and launch dates remain undefined, with current focus on building a full-stack cloud-edge software ecosystem.

Hourly heat curve

2 fully observed hours
Fewer than 3 fully observed hours; trend not plotted yet.

Coverage timeline

Total 2 reports · Latest first
  1. InfoQ 中文T2·68 pts

    Iluvatar CoreX Releases TianGai 300 Benchmarks: Matches H200 in Specific Inference Scenarios

    • TianGai 300 achieves 1.05x the end-to-end throughput of NVIDIA H200 per card under typical PD-disaggregated deployment.
    • In 64K long-context and MoE model tests, Attention performance and throughput reach 1.05x and 1.15x of H200 respectively.
    • Previewed next-gen edge product TongYang TY2000 with nominal 600T compute power, though specific specs and launch date remain unannounced.
  2. 量子位 (微信公众号)T2·68 pts

    Iluvatar CoreX Launches TG300 GPU, Outperforms Nvidia H200 in Live Benchmarks

    • TG300, based on the Tianshu architecture, targets Nvidia's Hopper series and demonstrated faster text-to-image inference than H200 in live tests.
    • Third-party benchmarks indicate TG300 achieves 1.05x single-card end-to-end throughput compared to H200 in PD-separated LLM deployment.
    • Edge product TY1200 outperformed Nvidia Jetson Thor T5000 in offline robotic tasks during live demonstrations.
    • Iluvatar CoreX launched an AI-for-AI software stack, emphasizing a closed loop of cloud production, software adaptation, and edge deployment.
Iluvatar CoreX Launches TG300, Surpassing NVIDIA H200 in Key Inference Tasks | Today for AI