Skip to main content

Alibaba Hanguang 800 (Internal AI Inference, 2021)

Product Overview

Alibaba Hanguang 800 is Alibaba subsidiary T-Head Semiconductor's first-generation AI inference chip, announced at the 2019-09 Hangzhou Apsara Conference, mass production in 2021. Based on TSMC 12nm, 820 INT8 TOPS compute (the leading inference chip of its era), 700 GB/s memory bandwidth, 168W TDP. Paired with the HALO (Hanguang Accelerated Linear Operator) software stack.

Strategic significance: Hanguang 800 marks the beginning of self-developed AI chips at Chinese internet companies. Alibaba's Taobao search/recommendation, Alibaba Cloud PAI platform, and Cainiao logistics fully deployed Hanguang 800, replacing NVIDIA T4/L4 inference. Alibaba Cloud 2021 revenue exceeded ¥100B, with AI inference compute 70% from self-developed chips.

Core Specs

ItemParameter
ArchitectureHanguang NPU (proprietary)
ProcessTSMC 12nm
Compute Cores170,000 NPU Cores (proprietary ISA)
On-chip SRAM32MB
HBM32GB HBM2 (4 stacks)
Memory Bandwidth700 GB/s
INT8820 TOPS
BF16128 TFLOPS (among the top inference BF16 of its era)
FP16128 TFLOPS
TDP168W
Form FactorPCIe Gen3 x16
InterconnectPCIe + proprietary RLLink
Mass Production2021-Q3
Unit Pricenot disclosed (internal Alibaba)

Why 12nm Process

DimensionHanguang 800 (12nm)Contemporary NVIDIA T4 (12nm)Contemporary Huawei Ascend 310 (12nm)
Process12nm12nm12nm
INT8820 TOPS130 TOPS16 TOPS (310) / 22 TOPS
BF16128 TFLOPS--
TDP168W70W8W (310)
Volumeinternal Alibaba 100K+NVIDIA generalHuawei Cloud + Edge

12nm choice rationale: Mature process + stable mass production (vs 7nm early yield issues), sufficient performance + cost advantage. Alibaba 2021 overall strategy: proprietary + cost-effective + no reliance on US advanced process.

Hanguang vs Contemporaries

MetricAlibaba Hanguang 800NVIDIA T4Huawei Ascend 310Google Edge TPU
Process12nm12nm12nm28nm
INT8820 TOPS130 TOPS16-22 TOPS4 TOPS
TDP168W70W8W2W
Efficiency4.88 TOPS/W1.86 TOPS/W2-2.75 TOPS/W2 TOPS/W
Memory32GB HBM216GB GDDR68GB LPDDR48MB SRAM
Pricenot disclosed~$2Knot disclosed~$60

Hanguang 800 advantage: INT8 compute 6.3x T4, efficiency 2.6x T4, the leading inference chip of 2021.

Internal Alibaba Deployment Scenarios

ScenarioBusinessSavings
Taobao search & recommendation1 billion product search ranking50% GPU cost savings (vs V100)
Alibaba Cloud PAI PlatformML inference service5x T4 compute per card
Cainiao Logistics Scheduling10M+ package route optimizationscheduling latency 50ms
Tmall Image Recognitionproduct images 1M+/dayreplaced T4 90% workload
Alipay Risk Controlreal-time fraud detectionmillisecond-level decisions
DingTalk AI AssistantLLM inferencecost reduced by 70%

Hanguang 800 commercialization significance: Alibaba's first self-developed AI chip 100% replacing NVIDIA internally. Single chip 5x T4 compute + 70% cost savings.

HALO Software Stack

LayerToolDescription
AI frameworkHALO (Hanguang Accelerated Linear Operator)unified programming platform
PyTorch (HALO backend)automatic NPU mapping
TensorFlow (HALO backend)compatible
HALO-MLAMulti-Level API (high-level)
HALO-Litelightweight API (inference)
CompilerHALO CompilerPyTorch / TF -> NPU
RuntimeHALO Runtimescheduling
QuantizationHALO QuantINT8 automatic
Model ZooHALO ModelZoo100+ inference models

HALO advantage: internal Alibaba 100+ models pre-optimized (Taobao search, recommendation, risk control, image, NLP), ready to use out of the box.

Vendor Information

ItemContent
CompanyAlibaba / T-Head Semiconductor
Business UnitAlibaba Cloud + T-Head Semiconductor
FoundedT-Head 2018-09 (C-SKY + DAMO Academy merger)
Hanguang 800 TeamAlibaba DAMO Academy + T-Head (Shanghai + Hangzhou)
Total Investment$5B+ (2018-2023 cumulative)
First ChipHanguang 800 (2019 released, mass production in 2021)
Follow-upsXuantie C910 (RISC-V CPU), Yushan 600 (SSD controller), Zhenyue 510 (enterprise SSD)
Customersinternal Alibaba 100% + Alibaba Cloud PAI platform
EmployeesT-Head ~1500 (semiconductor)
FabTSMC 12nm + 5nm (2024 estimated)

T-Head Product Line

ProductTypeReleasedComputePurpose
Hanguang 800AI Inference2019-09 / mass production 2021820 INT8 TOPSinternal Alibaba + Alibaba Cloud
Xuantie C910RISC-V CPU20192.5 GHz 12-coreIoT / Edge
Xuantie C906RISC-V CPU20191 GHzMCU
Yushan 600SSD Controller2020PCIe Gen4data center SSD
Zhenyue 510Enterprise SSD202216TBAlibaba Cloud Pangu
Hanguang 900 (estimated)AI inference next-gen2024 estimated2-3 PF2024+

Alibaba AI Strategy

Dimension2019 Hanguang 800 Era2024+ Estimated
Businessinternal Alibaba + Alibaba CloudAlibaba Cloud + external customers
ApplicationsSearch / Recommendation / Logistics+ LLM (Qwen) / Multimodal
Compute820 INT8 TOPS2-3 PF (estimated)
Memory32GB HBM296-128GB HBM3 (estimated)
Mass Production10K+ units / year100K+ units / year
T-Head Teamsemiconductor independentAlibaba Cloud full-stack integration

Key Features

  • 820 INT8 TOPS: the leading inference chip of 2021
  • 128 BF16 TFLOPS: supports LLM inference
  • 32GB HBM2: 32GB memory for LLaMA 1 65B inference
  • HALO software stack: internal Alibaba 100+ models
  • 100% in-house replacement: Alibaba Taobao / Tmall / Alipay / Cainiao / DingTalk
  • TDP 168W: single GPU replacement
  • Drawbacks: discontinued (2023-12), never sold externally, 4-year ecosystem

vs Chinese AI Chips (2021 Era)

MetricAlibaba Hanguang 800Huawei Ascend 310Cambricon MLU 370
Process12nm12nm7nm
INT8820 TOPS22 TOPS96 TOPS
TDP168W8W35W
Memory32GB HBM28GB LPDDR448GB HBM2
Deploymentinternal AlibabaHuawei Cloudgovernment/enterprise cloud

2021 Hanguang 800 compute was 37x Ascend 310, but post-2022, Huawei Ascend 910 series + Cambricon 590 quickly caught up, Hanguang 800's advantage faded.

Use Cases

  • internal Alibaba AI inference (Taobao / Tmall / Alipay / Cainiao / DingTalk)
  • Alibaba Cloud PAI inference service (PAI-EAS)
  • LLM inference (Qwen 7B / 14B / 72B optimized)
  • Search / Recommendation / Image / NLP (100+ models pre-optimized)
  • Alibaba e-commerce (Taobao search ranking)
  • External sales (internal + Alibaba Cloud only)
  • AI training (inference only)
  • CUDA proprietary workloads (requires HALO porting)

Key Timeline

TimeEvent
2018-09Alibaba founded T-Head Semiconductor (C-SKY + DAMO Academy)
2019-09Hanguang 800 announced at Hangzhou Apsara Conference (DAMO Academy R&D)
2020T-Head internal testing + small-scale Alibaba deployment
2021-Q3Hanguang 800 mass production, internal Alibaba 100K+ units deployed
2022Hanguang 800 deployed to Alibaba Cloud PAI platform
2023internal Alibaba inference compute 70% Hanguang 800
2023-12Hanguang 800 officially discontinued (transitioning to next-gen)
2024+Next-gen Hanguang (estimated 900 series)