Skip to main content

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

ItemParameter
ArchitectureVastai Unified Computing Architecture VUCA (SV100 chip)
Process7nm
INT8 Compute400 TOPS
FP16 ComputeNot disclosed (estimated ~200 TFLOPS)
TDP150 W
Form FactorPCIe (specific interface version not disclosed)
Release2022-09-03 (2022 World AI Conference)
Mass Production2022-Q4
Video CodecSupports 100-channel 1080P 30fps encoding/decoding/transcoding
Application ScenariosLive video enhancement, smart traffic, real-time semantic understanding

Vastai Unified Computing Architecture VUCA

ComponentDescription
AI EngineUltra-low latency, ultra-high throughput AI inference engine
Vector Compute EngineProgrammable vector computing
Video Processing EngineStreaming media, video encoding/decoding acceleration
VastStream APIUnified API interface, supports PyTorch/TensorFlow adaptation

Comparison with Mainstream GPU at Same Power

MetricZaiTian VA10Mainstream GPU (same power)Comparison
INT8 Compute400 TOPS~200 TOPS
Power Consumption150W150WSame
Low-latency Scenario (YoloV3 <4ms)3× performanceBaselineVA10 advantage
Video Encoding/Decoding100-channel 1080PFewerVA10 advantage

Company Information

ItemContent
CompanyVastai Tech
Founded2018-12
HeadquartersShanghai, China
FounderQian Jun (former AMD executive)
Product Pagehttps://www.vastaitech.com/product/general/va10
Software StackVastStream (CUDA-like API)
StatusMass production, second-generation SG100 released

Product Line

ProductReleasePositioningStatus
ZaiTian VA102022-09AI inference accelerator✅ Mass production
ZaiTian VE1S/VE1M2022-09Edge AI inference✅ Mass production
SG100 (QianYuan)2023-07Second-generation full-function GPU✅ Released
NanYu VG1600/VG18002023-07Data center GPU accelerator✅ Released
VA1L2023-07LLM 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