Zhonghao Xinying XuYu TPU (AI Training/Inference)
Product Overview
Zhonghao Xinying is an emerging Chinese TPU-architecture AI chip startup. On June 30, 2026, it officially released the XuYu TPU (unified training and inference), becoming one of the few companies globally to master TPU architecture after Google. Per official launch figures: 896 TFLOPS mixed-precision floating point, 1792 TOPS INT8, TDP 600W, with up to 2048 chips connected via all-optical interconnect in a single supernode; the company claims about 50% lower power consumption than traditional chips at the same compute level. The Tianjin Mobile TPU AI Computing Center is already operational, marking the first benchmark case of domestic TPU commercialization.
Core design philosophy: Abandon GPU's graphics rendering modules; a pure ASIC design focused on AI computation (fully in-house IP, instruction set, and operator library), delivering significantly better energy efficiency than traditional GPUs at the same process node. The software stack is compatible with PyTorch / vLLM / SGLang / DeepSpeed / Megatron.
Core Specifications
| Item | Parameter |
|---|
| Release | 2026-06-30 (XuYu official launch) |
| Architecture | Self-developed TPU (pure ASIC, no graphics rendering; in-house IP/instruction set/operator library) |
| Process | Not disclosed |
| FP16/BF16 Compute | 896 TFLOPS (mixed-precision floating point, official launch figures) |
| INT8 Compute | 1792 TOPS |
| FP32 Compute | Not disclosed |
| TDP | 600 W |
| Interconnect | All-optical interconnect, up to 2048 chips per supernode |
| Software Compatibility | PyTorch / vLLM / SGLang / DeepSpeed / Megatron |
| Production Status | In mass production and delivery |
| Unit Price | Not disclosed |
📌 Data correction (2026-09 cross-validation): This page previously recorded "INT8 512 TOPS / FP16 256 TFLOPS (estimated) / TDP 400W / released 2026-05" based on early media reports; it has now been updated to the official XuYu launch figures of 2026-06-30: 896 TFLOPS mixed precision, 1792 TOPS INT8, 600W, up to 2048 chips all-optical interconnect per supernode.
Efficiency Comparison
| Chip | Power | Compute | Positioning |
|---|
| Zhonghao Xinying XuYu TPU | 600 W | 896 TFLOPS mixed precision / 1792 TOPS INT8 | Official claim: ~50% lower power than traditional chips at the same compute level |
| NVIDIA H100 | 700 W | 3959 TOPS INT8 | Baseline |
| Cambricon MLU590 | 350 W | 512 TOPS INT8 | Domestic counterpart |
ℹ️ Source of the efficiency advantage: Pure ASIC design with no graphics overhead; a dedicated Matrix Multiplication Unit (MXU) architecture analogous to Google's TPU delivers significantly lower power consumption and cooling costs than GPUs in inference scenarios.
Commercial Deployment
| Item | Details |
|---|
| First Customer | Tianjin Mobile |
| Deployment | Tianjin Mobile TPU AI Computing Center |
| Status | Operational |
| Industry Significance | Among the first benchmarks of domestic TPU commercialization |
Architecture Differences vs GPU
| Dimension | Zhonghao Xinying TPU | Traditional GPU (e.g. H100) |
|---|
| Design Philosophy | Pure AI ASIC | General-purpose GPU (graphics+AI) |
| Energy Efficiency | High (no graphics overhead) | Lower |
| Programming Flexibility | Lower (fixed dataflow) | High (CUDA general-purpose computing) |
| Ecosystem Compatibility | Self-developed (no CUDA compatibility; compatible with mainstream framework interfaces) | CUDA ecosystem |
| Use Cases | AI inference + training | General-purpose computing |
Use Cases
- ✅ AI inference (high-efficiency scenarios)
- ✅ AI computing center construction (domestic compliance)
- ✅ Training of hundred-billion-parameter large models (supernode-scale all-optical interconnect)
- ✅ Low power / low cooling cost scenarios
- ❌ Complex dataflow models (less flexible than GPU)
- ❌ Graphics rendering / general-purpose computing
Manufacturer Info
| Item | Content |
|---|
| Company | Zhonghao Xinying (Hangzhou) Technology Co., Ltd. |
| Positioning | Emerging domestic TPU-architecture AI chip player |
| Core Product | XuYu TPU (released 2026-06-30) |
| First Customer | Tianjin Mobile |
| Launch Date | June 30, 2026 |
| Funding | Multiple rounds |