Enflame CloudBlaze T21 (2022)
Product Overview$
CloudBlaze T21 is Enflame's second-generation cloud AI training accelerator (OAM module) released in 2022, based on self-developed Suisi 2.0 chip, using OAM (OCP Accelerator Module) standard form factor, board-level power consumption 400W, supporting HBM2E memory, peak compute up to single-precision tensor TF32 160 TFLOPS, is the second-generation main product of Enflame's "CloudBlaze" training product line.
Positioning: Standardized OAM module, targeting AI computing center large-scale training cluster deployment, supporting multi-card high-speed interconnect.
Core Specifications$
| Item | Parameter |
|---|---|
| Architecture | Suisi 2.0 (Enflame self-developed) |
| Process | 7nm (estimated TSMC) |
| TF32 | 160 TFLOPS |
| FP16 | Not disclosed (estimated ~320 TFLOPS) |
| INT8 | Not disclosed (estimated ~640 TOPS) |
| Memory Capacity | HBM2E (capacity not disclosed, estimated 32–64GB) |
| Memory Bandwidth | Not disclosed (estimated ~1.6 TB/s) |
| TDP | 400 W (board-level power consumption) |
| Form Factor | OAM module (OCP standard) |
| Interconnect | Enflame self-developed high-speed interconnect (multi-card) |
| Release | 2022 |
| Mass Production | Since 2022 |
| Software Stack | Enflame intelligence computing software stack (PyTorch compatible) |
⚠️ Specification Note: FP16/INT8 compute and memory capacity are estimated values (not fully disclosed by official sources), subject to Enflame's subsequent official data sheet.
CloudBlaze Training Product Line$
| Product | Chip | Form Factor | TDP | Release | Status |
|---|---|---|---|---|---|
| CloudBlaze T10 | Suisi 1.0 | PCIe | ~300W | 2020 | On sale |
| CloudBlaze T20 | Suisi 1.5 | PCIe | ~300W | 2021 | On sale |
| CloudBlaze T21 | Suisi 2.0 | OAM | 400W | 2022 | Current mainstream |
| CloudBlaze T30 (estimated) | Suisi 3.0 | OAM | Not disclosed | 2025+ | Next generation |
Comparison with CloudBlaze T20$
| Metric | CloudBlaze T20 | CloudBlaze T21 | Improvement |
|---|---|---|---|
| Chip | Suisi 1.5 | Suisi 2.0 | New generation |
| Form Factor | PCIe card | OAM module | Standardized |
| TDP | ~300W | 400W | +33% |
| TF32 | Not disclosed | 160 TFLOPS | New generation |
| Cluster | First generation | Second-generation CloudBlaze AI cluster | 8000+ cards |
Second-Generation CloudBlaze AI Cluster$
Composed of 8000+ CloudBlaze T21 training cards, the second-generation CloudBlaze AI computing cluster 2.0 achieves single-precision peak compute up to 1.3 EFLOPS (1300 TFLOPS), targeting ultra-large-scale AI training scenarios.
Application Scenarios$
- ✅ Ultra-large-scale AI training clusters (OAM standardized deployment)
- ✅ AI computing centers (8000+ card cluster verified)
- ✅ Large model pre-training (TF32 160 TFLOPS)
- ✅ Domestic computing power substitution (supply chain security)
- ❌ PCIe standard card requirements (T21 is OAM module, requires dedicated baseboard)
- ❌ Inference scenarios (training-specialized, recommend CloudBlaze i20)
Enflame Product Matrix$
| Series | Positioning | Representative Product |
|---|---|---|
| CloudBlaze T Series | Training (Training) | T20, T21 |
| CloudBlaze i Series | Inference (Inference) | i10, i20 |
| CloudBlaze S Series | Training+Inference | S60 (estimated) |
Related Cards$
- Enflame CloudBlaze T20 - Previous-generation PCIe training card
- Enflame CloudBlaze i20 - Same-generation inference card
- Huawei Ascend 910C - Domestic training card competitor
- Cambricon MLU590 - Domestic training card competitor