Tencent Zixiao
Product Overview
Zixiao is Tencent's first cloud AI inference chip, developed by its Penglai Laboratory. On 2021-11-03, at the 2021 Tencent Digital Ecosystem Summit, Dowson Tang, Tencent Senior Executive Vice President and CEO of the Cloud and Smart Industries Group, publicly revealed Tencent's chip-making plan for the first time, along with three in-house chips: "Zixiao" for AI computing, "Canghai" for video processing, and "Xuanling" for high-performance networking. Tencent Cloud President Qiu Yuepeng stated at the event that Zixiao's performance improves 100% over the industry, and that it had already taped out successfully and been powered on.
The core of Zixiao's architecture is breaking the "memory wall" of AI chips: the chip uses 2.5D packaging technology to co-package HBM2e memory with the AI core, overcoming the memory bandwidth bottleneck that constrains compute performance; meanwhile, computer vision (CV) accelerators and video codec accelerators are added inside the chip, with targeted optimization for Tencent's high-frequency workloads in image/video processing, natural language processing, and search and recommendation. It is an inference-dedicated NPU/ASIC, not intended for training.
Development entity and collaboration model: Zixiao's architecture and scenario definition were led by Tencent's Penglai Laboratory (founded in 2020, focused on full end-to-end chip design and verification), with Enflame (Shanghai Enflame Technology Co., Ltd.) deeply participating in the joint R&D — Tencent has invested in Enflame since 2018 and is currently its largest shareholder (holding about 20.5%). Chip IP and deployment belong to Tencent, mainly for Tencent's internal businesses and Tencent Cloud; the chip is not sold retail externally.
Mass production and deployment: as of 2023, Zixiao had entered mass production and been deployed at scale across multiple of Tencent's top business lines, delivering up to 3x compute acceleration and over 45% overall cost savings in Tencent's business scenarios, while speeding up speech-to-text by 4.7x and increasing OCR recognition throughput by 2.4x. Its landmark deployment is Tencent Meeting real-time subtitles: Zixiao has been fully launched (covering over 95% of total subtitle traffic in meetings), with per-card load of a Zixiao machine reaching 4x that of the NVIDIA T4, timeout rates reduced from 0.005% to 0, and overall costs cut by 75%. Other applications include advertising recommendation, OCR text recognition, image and speech denoising, content moderation, and Hunyuan large-model inference; Tencent Cloud also offers inference acceleration instances equipped with Zixiao externally. The accompanying software stack is Tencent's self-developed high-performance inference runtime, LightRuntime (supporting Auto Batch, Auto Padding, and multi-model scheduling).
Core Specifications
| Parameter | Value |
|---|---|
| Architecture | Tencent proprietary AI inference NPU/ASIC architecture (built-in CV accelerator + video codec accelerator) |
| Process Node | Not disclosed (never officially announced; TSMC 7nm circulates in the industry as unverified supply chain information) |
| FP16 / BF16 Compute | Not disclosed (third-party research reports claim the Zixiao C100 approaches the NVIDIA A10, with FP16 peak compute about 2x the comparison card) |
| INT8 Compute | Not disclosed |
| FP32 Compute | Not disclosed |
| Memory Capacity | Not disclosed |
| Memory Type | HBM2e (officially confirmed, co-packaged with the AI core via 2.5D packaging) |
| Memory Bandwidth | Not disclosed |
| TDP | Not disclosed (third-party sources claim about 70% lower power than solutions such as the NVIDIA T4) |
| Interconnect | Not disclosed |
| Interface | Not disclosed (deployed as boards in Tencent Xingxinghai custom servers) |
| Packaging | 2.5D advanced packaging (HBM2e + AI core co-packaged) |
| Relative Performance | Official framing: performance improves 100% over the industry; 3x compute acceleration in business scenarios; per-card load = 4x the NVIDIA T4 |
| Cost Efficiency | Official framing: overall cost savings of over 45% (75% cost reduction in the Tencent Meeting scenario) |
| Software Stack | LightRuntime (Tencent's self-developed high-performance inference runtime, supporting ONNX, Auto Batch, Auto Padding, multi-Session scheduling) |
| Launch | 2021-11-03 (2021 Tencent Digital Ecosystem Summit) |
| Mass Production/Availability | Mass-produced in 2023 and deployed at scale in Tencent's top businesses; internal and Tencent Cloud use only, not sold retail |
⚠️ Extremely Limited Specification Disclosure Tencent has never publicly released hard specifications for Zixiao such as compute (TFLOPS/TOPS), memory capacity, memory bandwidth, TDP, or process node; officially it has only confirmed "2.5D packaging + HBM2e + proprietary architecture" and relative performance improvement figures. The circulating "TSMC 7nm" is secondhand supply chain information, not confirmed by Tencent; "Zixiao C100 performance approaching the NVIDIA A10" comes from third-party securities research, also not an official framing. All "Not disclosed" fields on this page are left unfilled rather than speculated.
Subsequent Evolution
| Model | Time | Positioning and Public Information |
|---|---|---|
| Zixiao (V1) | Released 2021-11 / mass-produced 2023 | Pure AI inference NPU, 2.5D + HBM2e, covering CV, NLP, recommendation, advertising, speech transcription, OCR, and Hunyuan large-model inference |
| Zixiao 2.0 | Released 2026-07 (WAIC 2026) | For large-model training and inference (unified training and inference); per Tencent Cloud, training performance improves 4x and inference throughput 3x over the previous generation |
Zixiao 2.0 Information Note: Zixiao 2.0 was released in July 2026 at WAIC 2026; according to public reports, Tencent plans for in-house Zixiao chips to account for 30% of new additions in 2026, with deployment scale expanding to the 100,000-card level by year-end. Some sources call Zixiao 2.0 Tencent's "sixth-generation in-house AI chip"; that claim comes from UGC platforms and its reliability awaits official confirmation. At the highest-end training compute, Zixiao 2.0 still falls short of NVIDIA H100/B200-class hardware; Tencent's training-compute self-sufficiency route advances on two tracks: "in-house Zixiao + controlling stake in Enflame".
Key Features
- 2.5D packaging + HBM2e breaking the memory wall: co-packaging HBM2e with the AI core specifically addresses the memory bandwidth bottleneck of inference scenarios — Zixiao's most core architectural innovation
- Built-in CV accelerator and video codec accelerator: dedicated hardware offload for Tencent's image/video-heavy business workloads, rather than a purely general matrix engine
- Inference-dedicated positioning: V1 explicitly does not do training; the architecture is highly bound to Tencent's own business scenarios
- Official performance framing: performance improves 100% over the industry; 3x compute acceleration in business scenarios; over 45% overall cost savings
- Landmark Tencent Meeting real-time subtitle deployment: fully launched, covering 95%+ of total subtitle traffic, per-card load of 4x T4, timeout rate reduced to 0, costs down 75%
- Major speech and OCR speedups: speech-to-text 4.7x faster, OCR throughput up 2.4x
- LightRuntime software stack: Auto Batch improving throughput by about 20%, Auto Padding automatic bucketing, multi-Session priority scheduling; already refined in Tencent Meeting, WeChat, and other businesses
- Deep Hunyuan large-model optimization: co-tuned with Tencent's Hunyuan large model
- Tencent–Enflame collaborative R&D model: Penglai Laboratory leads + Enflame co-develops, with Tencent holding about 20.5% of Enflame
- Full-stack closed loop: Zixiao chip + Xingmai high-performance network + Xingxinghai custom servers + TI-One platform + Hunyuan large model
- Limitations: not sold externally; V1 does not support training; hard specifications long opaque, preventing like-for-like compute comparisons
Vendor Information
| Parameter | Value |
|---|---|
| Company | Tencent Holdings Ltd. (HKEX 00700.HK) |
| Headquarters | Nanshan District, Shenzhen, Guangdong Province |
| Founded | November 1998 |
| Chip Development Entity | Penglai Laboratory (founded in 2020, responsible for end-to-end chip design and verification); Shannon Laboratory (participated in the Canghai video chip) |
| Joint R&D Partner | Shanghai Enflame Technology Co., Ltd. (Tencent holds about 20.5%, largest shareholder) |
| Business Affiliation | Cloud and Smart Industries Group (CSIG) / Tencent Cloud |
| In-House Chip Matrix | Zixiao (AI inference), Canghai (video transcoding, started 2019 / mass-produced 2022, tens of thousands of units deployed), Xuanling (smart NIC, 4x performance improvement) |
| Other Chip Investments | Enflame (AI chips), Yunbao Intelligence (DPU), Lightelligence (photonic chips), Barefoot Networks (programmable network chips) |
| Sales Model | No retail chip sales; available only to Tencent's internal businesses and Tencent Cloud inference acceleration instances |
| Official Website | https://www.tencent.com |
Use Cases
- ✅ Tencent Meeting real-time subtitles / streaming speech recognition (landmark deployment)
- ✅ WeChat / Channels content recommendation and advertising inference
- ✅ OCR text recognition, image and speech denoising
- ✅ Content safety moderation (image/video CV inference)
- ✅ Tencent Hunyuan large-model inference services
- ✅ Tencent Cloud external inference acceleration instances (delivered as PaaS)
- ✅ Inference workloads such as financial risk control and medical imaging
- ❌ External chip/card procurement (Tencent does not sell retail; only indirectly usable via Tencent Cloud)
- ❌ Large-model AI training (V1 is a pure inference architecture; training requires Zixiao 2.0 or external GPUs)
- ❌ Graphics rendering and general GPGPU computing (dedicated AI ASIC, no graphics capability)
- ❌ Direct migration from the CUDA ecosystem (requires adaptation via the LightRuntime / ONNX path)
- ❌ Situations requiring public specifications for comparison (hard metrics not disclosed)
Related Cards
- Alibaba Hanguang 800 — Direct benchmark from a major Chinese internet company's in-house AI inference chip (Alibaba T-Head, also an internal NVIDIA T4 replacement route)
- Enflame YunSui i20 — Enflame's cloud inference product, the joint R&D partner's own product line
- Enflame YunSui i10 — Enflame's first-generation cloud inference card
- Moore Threads MTT S5000 — Domestic full-function GPU route comparison (general-purpose GPU vs dedicated ASIC)
- MetaX C600 — Domestic high-end general-purpose GPU comparison
- Cambricon MLU 370 — Domestic dedicated AI acceleration chip comparison
References
- Tencent reveals three in-house chips, some entering tape-out (China Electronics News, 2021-11-12, official framing of the Zixiao/Canghai/Xuanling debut)
- Latest progress on Tencent's in-house chip "Canghai": 8 global firsts, tens of thousands of units mass-produced and deployed (Electronic Product World, including a recap of the November 2021 three-chip launch)
- Computer industry April 2024 research views: a review of major internet companies' in-house chips (Haitong Securities, Zixiao mass production and Tencent Meeting deployment data)