Iluvatar CoreX Product Research Collection

Status: research in progress; official pages will be created after ByteDance's procurement is confirmed Last updated: 2026-06-28 Sources: official site, Moark, Zhihu, industry reports

---

Company Overview

Item	Content
Company name	Shanghai Iluvatar CoreX Semiconductor Co., Ltd.
English name	Iluvatar CoreX
Founded	2015
Founder	Diao Shijing
Headquarters	Shanghai
Positioning	Domestic general-purpose GPU chip design company
Official site	https://www.iluvatar.com

---

Product Line Overview

1. Zhikai series (inference GPUs)

Zhikai 100 (MR-V100)

Item	Specification
Release	Lit up in May 2022, released at the end of 2022
Product positioning	Cloud inference general-purpose GPU
Architecture	Second-generation general-purpose GPU architecture (in-house)
Process	7nm
Packaging	2.5D COWOS
Memory	32 GB HBM2E
Peak compute
- FP32	24 TFLOPS
- FP16	96 TFLOPS
- INT8	192-384 TOPS (sources differ)
TDP	150W (board-level power)
Interface	PCIe Gen4.0 x16
Video decoding	128 channels of 1080P@30fps (H.264/H.265/VP9/AVS2)
Software stack	IXUCA (CUDA-ecosystem compatible)
Use cases	AI inference, video analytics, security, internet, finance, healthcare

Performance benchmark: NVIDIA RTX 4090 (inference scenarios)

Product features:

• Fully in-house architecture, core IP, and software stack
• CUDA-ecosystem compatible, cutting migration time by more than 50%
• Supports 800+ general-purpose instructions
• Supports mixed-precision inference with FP32, FP16, and INT8
• FP8 not supported

Zhikai 50 (MR-V50)

Item	Specification
Memory	16 GB HBM2E
Peak compute
- FP32	16 TFLOPS
- FP16	64 TFLOPS
- INT8	256 TOPS
TDP	75W (board-level power)
Dimensions	Half-length, half-height, single-slot PCIe card
Cooling	Passive cooling

---

2. Tiangai series (training GPUs)

Tiangai 100 (BI-V100)

Item	Specification
Release	March 2021 (China's first cloud training chip with a fully in-house GPU architecture)
Product positioning	Cloud training general-purpose GPU
Architecture	First-generation general-purpose GPU architecture (in-house)
Process	7nm
Memory	32 GB HBM2E
TDP	250W (estimated)
Use cases	AI training, high-performance computing

Strategic significance: marks the breakthrough of the domestic general-purpose GPU from zero to one

Tiangai 150 (BI-V150)

Item	Specification
Release	2023 (estimated)
Product positioning	Cloud training general-purpose GPU (upgraded version)
Architecture	ivcore11 in-house general-purpose GPU architecture
Process	7nm
Memory	64 GB HBM2e (a 32GB variant also exists)
TDP	350W (peak power)
Supported precisions	FP32, FP16, INT8, FP8 (requires the ixTE library)
Use cases	Large AI model training, general-purpose computing
Performance benchmark	NVIDIA A100 (claimed)

Product features:

• Compatible with mainstream international GPU general-purpose computing models
• Supports mainstream domestic and international AI ecosystems and deep learning frameworks
• Supports native operators of TensorFlow, PyTorch, etc.

---

3. Tongyang series (new product line)

Tongyang TY1000/TY1100/TY1200

Note: new product line with little available material; under continued monitoring.

---

Software Stack: IXUCA

IXUCA (Iluvatar Unified Computing Architecture) is the unified computing architecture software stack independently developed by Iluvatar CoreX.

Core components

Component	Name	Function	Counterpart
Deep learning frameworks	PyTorch-Cambricon, TensorFlow-Cambricon	Adapted deep learning frameworks	PyTorch, TensorFlow
Inference framework	IGIE	High-performance inference framework	TensorRT
Inference engine	IxRT	Dedicated inference acceleration engine	TensorRT
LLM inference framework	IxFormer	LLM inference and training optimization	vLLM
Compiler	IXUCA Compiler	Compiler	nvcc
Math libraries	ixDNN, ixBLAS	Fundamental deep learning operators	cuDNN, cuBLAS
Communication library	ixCCL	Multi-card communication library	NCCL
Management tool	ixsmi	GPU management tool	nvidia-smi

Ecosystem compatibility

• ✅ CUDA-ecosystem compatible: supports CUDA C++ programming with low migration cost
• ✅ Mainstream framework support: TensorFlow, PyTorch, PaddlePaddle
• ✅ Mainstream inference framework support: vLLM, TGI, LMDeploy
• ✅ 200+ models adapted: covering LLM, CV, NLP, speech, and other domains

---

Market Dynamics

ByteDance procurement rumors

Item	Content
Date	June 15, 2026 (Reuters report)
Status	In negotiation, not yet confirmed
Procurement scale	At least 50,000 units (planned)
Primary use	Inference workloads
Impact	If the deal closes, Iluvatar CoreX would become the third domestic AI chip vendor to supply ByteDance at scale, after Huawei and Cambricon

Market reaction:

• On June 18, 2026, the SSE STAR Market chip design theme index rose 3.85%
• Cambricon rose 13.31%
• Huatai Securities forecast: Iluvatar CoreX's 2026 revenue could exceed 3.04 billion CNY

Shipment forecast

Year	Chip Shipments	Revenue Forecast
2025	~42,000 units	-
2026 (forecast)	>100,000 units	>3.04 billion CNY

---

Information Pending Confirmation

urgently needed

1. ✅ Zhikai 100 INT8 compute: sources differ between 192 TOPS and 384 TOPS; needs confirmation
2. ❌ Tiangai 100/150 detailed compute: official FP32/FP16/INT8 figures not published
3. ❌ Tiangai 150 memory bandwidth: not published
4. ❌ Tongyang series specs: very little material
5. ❌ ByteDance procurement confirmation: awaiting official announcement

Requiring continued monitoring

1. Iluvatar CoreX official developer center: https://support.iluvatar.com/
2. DeepSpark open-source community: https://www.deepspark.org.cn/
3. ByteDance procurement progress

---

Page Creation Plan

Conditions

• ByteDance procurement confirmed (primary trigger)
• Tiangai 150 detailed compute data obtained
• Zhikai 100 final specs confirmed (INT8 compute)

Pages to create

1. docs/cards/iluvatar/mr-v100.md - Zhikai 100 (inference card)
2. docs/cards/iluvatar/bi-v150.md - Tiangai 150 (training card)
3. docs/cards/iluvatar/bi-v100.md - Tiangai 100 (optional, little material)

References

• Zhikai 100 - Moark
• Tiangai 150 - Moark
• Product specs - test page
• ByteDance to purchase 50,000 domestic chips - Tencent News
• Zhikai 100 - Baidu Baike

---

Next steps:

1. Continue following ByteDance procurement developments
2. Try contacting Iluvatar CoreX for detailed spec sheets
3. Follow actual Tiangai 150 deployment cases to obtain performance data
4. Create the pages immediately once procurement is confirmed