Source量子位•Stalecollected in 64m
SenseTime launches Galaxy Plan for domestic compute clusters
#compute-cluster#domestic-chips#ai-infrastructuresensetime-large-model-computing-platformsensetimegalaxy-plan
Massive domestic compute expansion to support large-scale AI model training in China.
30-Second TL;DR
What Changed
Launch of the 'Galaxy Plan' with 20+ partners
Why It Matters
This initiative significantly boosts the domestic AI training capacity in China, providing a viable alternative to international GPU clusters for large-scale model training.
What To Do Next
If you are training large models in China, assess the compatibility of your training stack with the new SenseTime domestic compute clusters.
Who should care:Enterprise & Security Teams
Key Points
- •Launch of the 'Galaxy Plan' with 20+ partners
- •Construction of five 10,000-card domestic compute clusters
- •Focus on scaling domestic AI infrastructure profitability
Deep Insight
AI-generated analysis for this event — not the original article.
Enhanced Key Takeaways
- •The Galaxy Plan specifically targets the optimization of the 'SenseCore' large model training stack to ensure compatibility with heterogeneous domestic GPU architectures.
- •SenseTime is implementing a unified scheduling layer designed to reduce inter-node communication latency, a critical bottleneck for 10,000-card clusters using domestic hardware.
- •The initiative includes a 'compute-as-a-service' model that provides partners with pre-configured software environments to lower the barrier for deploying large language models (LLMs) on non-NVIDIA infrastructure.
- •The plan addresses the 'fragmentation' issue in the domestic chip market by creating a standardized abstraction layer that allows models to run across different domestic chip brands without extensive code refactoring.
- •SenseTime is integrating its proprietary 'SenseNova' model series as the primary workload for these clusters to demonstrate performance parity with international standards.
Competitor Analysis
Huawei (Ascend)
- Focus Area
- Full-stack AI
- Key Advantage
- Vertical integration (Chip + Framework)
- Domestic Compute Strategy
- Proprietary ecosystem (CANN/MindSpore)
Baidu (PaddlePaddle)
- Focus Area
- Cloud/Model
- Key Advantage
- Massive scale/Experience
- Domestic Compute Strategy
- Integrated AI Cloud/Kunlun chips
Alibaba (Tongyi)
- Focus Area
- Model/Cloud
- Key Advantage
- Infrastructure scale
- Domestic Compute Strategy
- Hybrid cloud/Open-source model focus
| Competitor | Focus Area | Key Advantage | Domestic Compute Strategy |
|---|---|---|---|
| Huawei (Ascend) | Full-stack AI | Vertical integration (Chip + Framework) | Proprietary ecosystem (CANN/MindSpore) |
| Baidu (PaddlePaddle) | Cloud/Model | Massive scale/Experience | Integrated AI Cloud/Kunlun chips |
| Alibaba (Tongyi) | Model/Cloud | Infrastructure scale | Hybrid cloud/Open-source model focus |
Technical Deep Dive
- Cluster Architecture: Utilizes a high-speed interconnect fabric designed to mitigate the bandwidth limitations often found in domestic GPU clusters compared to InfiniBand-based systems.
- Software Stack: Employs a customized version of the SenseCore infrastructure, featuring an optimized collective communication library (CCL) tailored for domestic chip interconnects.
- Scheduling: Implements a multi-tenant, hierarchical scheduler that dynamically balances workloads across heterogeneous domestic GPU nodes to maximize utilization rates.
- Optimization: Focuses on kernel-level optimizations for domestic architectures to improve FP16/BF16 training throughput, aiming to close the performance gap with mainstream international GPUs.
Future ImplicationsAI analysis grounded in cited sources
Domestic AI training costs will drop by at least 30% within 18 months.
The standardization of the software stack across heterogeneous domestic chips will significantly reduce the engineering overhead currently required for model porting and optimization.
SenseTime will shift its revenue model from model-licensing to infrastructure-as-a-service.
By controlling the compute clusters and the underlying software stack, the company is positioning itself as a primary provider of domestic AI infrastructure rather than just a model developer.
Timeline
2023-04
SenseTime officially releases the SenseNova foundation model series.
2024-01
SenseTime announces strategic focus on domestic AI infrastructure and chip adaptation.
2025-06
SenseTime achieves initial success in training large models on domestic GPU clusters.
2026-07
Launch of the Galaxy Plan to scale domestic compute clusters.
- 2023-04SenseTime officially releases the SenseNova foundation model series.
- 2024-01SenseTime announces strategic focus on domestic AI infrastructure and chip adaptation.
- 2025-06SenseTime achieves initial success in training large models on domestic GPU clusters.
- 2026-07Launch of the Galaxy Plan to scale domestic compute clusters.
Weekly AI Recap
Read this week's curated digest of top AI events →
AI-curated news aggregator. All content rights belong to original publishers.
Original source: 量子位 ↗
This is a summary, not the original. Read the source, or get the weekly briefing.
The weekly digest
One email a week. Unsubscribe anytime.