Vastai Tech ZaiTian VA10 (2022)
Product Overview
ZaiTian VA10 is Vastai Tech's first data center AI inference accelerator released on September 3, 2022 at the 2022 World Artificial Intelligence Conference, equipped with self-developed SV100 chip, using 7nm process, INT8 peak compute up to 400 TOPS, power consumption only 150W, overall highest inference performance reaching 2× above mainstream GPU at same power consumption, is an important breakthrough for domestic AI inference cards.
Positioning: Data center AI inference acceleration (low latency, high throughput), supporting live video enhancement, smart traffic management, real-time semantic understanding and other high-real-time cloud AI application deployment scenarios.
Core Specifications
| Item | Parameter |
|---|
| Architecture | Vastai Unified Computing Architecture VUCA (SV100 chip) |
| Process | 7nm |
| INT8 Compute | 400 TOPS |
| FP16 Compute | Not disclosed (estimated ~200 TFLOPS) |
| TDP | 150 W |
| Form Factor | PCIe (specific interface version not disclosed) |
| Release | 2022-09-03 (2022 World AI Conference) |
| Mass Production | 2022-Q4 |
| Video Codec | Supports 100-channel 1080P 30fps encoding/decoding/transcoding |
| Application Scenarios | Live video enhancement, smart traffic, real-time semantic understanding |
Vastai Unified Computing Architecture VUCA
| Component | Description |
|---|
| AI Engine | Ultra-low latency, ultra-high throughput AI inference engine |
| Vector Compute Engine | Programmable vector computing |
| Video Processing Engine | Streaming media, video encoding/decoding acceleration |
| VastStream API | Unified API interface, supports PyTorch/TensorFlow adaptation |
Comparison with Mainstream GPU at Same Power
| Metric | ZaiTian VA10 | Mainstream GPU (same power) | Comparison |
|---|
| INT8 Compute | 400 TOPS | ~200 TOPS | 2× |
| Power Consumption | 150W | 150W | Same |
| Low-latency Scenario (YoloV3 <4ms) | 3× performance | Baseline | VA10 advantage |
| Video Encoding/Decoding | 100-channel 1080P | Fewer | VA10 advantage |
| Item | Content |
|---|
| Company | Vastai Tech |
| Founded | 2018-12 |
| Headquarters | Shanghai, China |
| Founder | Qian Jun (former AMD executive) |
| Product Page | https://www.vastaitech.com/product/general/va10 |
| Software Stack | VastStream (CUDA-like API) |
| Status | Mass production, second-generation SG100 released |
Product Line
| Product | Release | Positioning | Status |
|---|
| ZaiTian VA10 | 2022-09 | AI inference accelerator | ✅ Mass production |
| ZaiTian VE1S/VE1M | 2022-09 | Edge AI inference | ✅ Mass production |
| SG100 (QianYuan) | 2023-07 | Second-generation full-function GPU | ✅ Released |
| NanYu VG1600/VG1800 | 2023-07 | Data center GPU accelerator | ✅ Released |
| VA1L | 2023-07 | LLM large model AI accelerator | ✅ Released |
Application Scenarios
- ✅ Live video enhancement (100-channel 1080P real-time processing)
- ✅ Smart traffic management (low-latency YoloV3 detection)
- ✅ Real-time semantic understanding (high-real-time AI inference)
- ✅ Data center inference deployment (150W low power consumption, high cost-performance)
- ❌ AI training (inference-specialized)
- ❌ Graphics rendering (no rendering engine)
References