Enflame YunSui T20 (2021)
Product Overview
The YunSui T20 is the second-generation AI training accelerator card released by Enflame Technology on July 7, 2021 at the World Artificial Intelligence Conference (WAIC). Based on the in-house Suiyuan 2.0 (DTU 2.0) chip, it adopts 2.5D advanced packaging (57.5mm × 57.5mm, integrating 9 chips), delivers 160 TFLOPS of TF32 compute (the first domestic support for TF32) and 320 TOPS of INT8 compute, and is equipped with 32GB HBM2E memory (1.6 TB/s bandwidth), with a TDP of 300W. Its GCU-LARE interconnect technology supports scaling clusters to 8192 cards (1.3 EFLOPS).
Enflame is one of the "GPU Four Little Dragons", focusing on domestic cloud AI training and inference.
Product Evolution:
- DTU 1.0 / YunSui T10 (2019): first generation, 12nm, FP32 20 TFLOPS
- DTU 2.0 / YunSui T20/T21 (2021): 2.5D packaging, TF32 160 TFLOPS — this page
- DTU 3.0 / YunSui T30 (planned): next generation
Core Specifications
DTU 2.0 Chip
| Item | Parameter |
|---|
| Architecture | In-house GCU-CARA full-domain compute architecture |
| Process | Not disclosed (industry estimate: 12nm) |
| Packaging | 2.5D advanced packaging, integrating 9 chips |
| Package Size | 57.5mm × 57.5mm (China's largest compute chip at launch) |
| FP32 | 40 TFLOPS |
| TF32 | 160 TFLOPS (first domestic support) |
| FP16 / BF16 | Supported (specific figures not disclosed) |
| INT8 | 320 TOPS |
| Memory | 32GB HBM2E (Samsung; first domestic card to support it) |
| Memory Bandwidth | 1.6 TB/s |
| Interconnect | GCU-LARE® (Enflame intelligent interconnect), 300 GB/s bidirectional |
📌 Data correction (2026-09 cross-validation): This page previously recorded "64GB HBM2E / 1.8 TB/s", an early misstatement; Enflame's IPO prospectus figures and official materials specify 32GB HBM2E, 1.6 TB/s, TDP 300W, which has now been corrected.
YunSui T20 Accelerator Card
| Item | Parameter |
|---|
| Core Chip | DTU 2.0 |
| Positioning | Data center AI training accelerator card |
| Form Factor | PCIe training accelerator card |
| Multi-Card Interconnect | In-server 4-card full interconnect / enhanced 8-card full interconnect |
| Cluster | Supports scaling from single-server multi-card to thousand-card level |
| Software Stack | TopsRider 2.0 |
| Development Interfaces | C++ / Python, multi-level open APIs |
| TDP | 300 W |
| Release | July 7, 2021 (WAIC 2021) |
GCU-LARE Interconnect and Clusters
| Specification | Parameter |
|---|
| Interconnect Technology | GCU-LARE® full-domain interconnect |
| Inter-Chip Bandwidth | 300 GB/s bidirectional |
| In-Server Interconnect | 4-card full interconnect → enhanced 8-card full interconnect |
| Cluster Solution | Enflame intelligent computing cluster CloudBlazer Matrix 2.0 |
| Maximum Cluster | 8192 YunSui training cards |
| Total Cluster Compute | Up to 1.3 EFLOPS (FP32) |
| Cooling | Liquid cooling, PUE < 1.5 |
| Rack Solution | High-density deployment in a single rack |
Record at launch: Enflame's COO stated: "No one in the world has yet achieved over 1E of single-precision compute using 8,000 cards."
TopsRider 2.0 Software Stack
| Layer | Tool | Description |
|---|
| Platform | TopsRider 2.0 | Enflame's unified programming platform |
| AI Frameworks | PyTorch | Native support |
| TensorFlow | Supported |
| PaddlePaddle | Baidu PaddlePaddle |
| Development Interfaces | C++ / Python | Multi-level APIs |
| Operator Library | In-house operator library | Covers mainstream models |
| Compiler | GCU-CARA toolchain | Automated optimization |
| Performance | TF32 precision on average 2.5x that of competitors' sub-flagship cards | On par with competitors' flagships across many model types |
| Item | Details |
|---|
| Company | Shanghai Enflame Technology Co., Ltd. |
| Founded | March 2018 |
| Founders | Zhao Lidong (former AMD China executive), Zhang Yalin (COO) |
| T20 Launch | July 7, 2021 (WAIC 2021) |
| Funding | Several billion RMB in total (Tencent, Sequoia, etc.) |
| Positioning | Domestic cloud AI training/inference chips |
| Ecosystem | One of the "GPU Four Little Dragons" (MetaX, Biren, Enflame, Moore Threads) |
| Partnerships | Collaborates with partners to build the Enflame intelligent computing cluster |
Use Cases
- ✅ Domestic large-model AI training (8192-card cluster, 1.3 EFLOPS)
- ✅ Data center training (efficient training at 160 TFLOPS TF32)
- ✅ Thousand-card cluster deployment (mature GCU-LARE interconnect solution)
- ✅ Broad model coverage (multi-precision and dynamic feature support)
- ✅ Mandatory domestic compute (independent and controllable)
- ❌ Single-card inference (not the primary positioning; the YunSui i20 inference card is a better fit)
- ❌ CUDA ecosystem (in-house TopsRider; migration requires adaptation)
- ❌ FP8 training (not supported; watch the L600/T30)
Comparison with Contemporary Domestic AI Training Cards (2021)
| Metric | Enflame T20 | Cambricon MLU370-X8 | Huawei Ascend 910 | Difference |
|---|
| Release | 2021-07 | 2021-Q4 (X8 mass-produced in 2022) | 2019 | T20 launched mid-year |
| Packaging | 2.5D advanced packaging | Standard packaging | Standard packaging | T20 is more advanced |
| TF32 | 160 TFLOPS | Not supported | Not supported | Unique to T20 |
| FP32 | 40 TFLOPS | 24 TFLOPS | 256 TFLOPS | Ascend 910 leads |
| INT8 | 320 TOPS | 256 TOPS | 512 TOPS | Ascend 910 leads |
| Memory | 32GB HBM2E | 48GB LPDDR5 | 32GB HBM2 | MLU370 has the largest capacity |
| Bandwidth | 1.6 TB/s | 614 GB/s | 1.2 TB/s | T20 is the largest |
| Interconnect | 300 GB/s | 200 GB/s | HCCS | T20 leads |
| Cluster | 8192 cards, 1.3 EFLOPS | Thousand cards | 4096 cards | T20 is the largest |
2021's leading domestic AI training card: the T20 has the largest bandwidth (1.6 TB/s), the strongest interconnect (300 GB/s), and the largest cluster (8192 cards, 1.3 EFLOPS). However, its FP32 compute (40 TFLOPS) and INT8 (320 TOPS) are lower than the Huawei Ascend 910's.
Key Timeline
| Date | Event |
|---|
| 2018-03 | Enflame Technology founded |
| 2019-12 | DTU 1.0 / YunSui T10 released (12nm) |
| 2021-07-07 | DTU 2.0 / YunSui T20 released (WAIC) |
| 2021-2023 | T20/T21 deployed at scale; Enflame intelligent computing cluster commercialized |
| Planned | DTU 3.0 / YunSui T30 |