Enflame L600
Enflame's fourth-generation training+inference integrated AI chip, based on self-developed Suisi 400 chip, released on July 27, 2025 at World Artificial Intelligence Conference (WAIC), 144GB memory capacity, 3.6TB/s memory bandwidth, 800GB/s interconnect bandwidth, compute performance targeting NVIDIA H20.
Core Specifications
| Specification | Value |
|---|---|
| Architecture | Suisi 400 (DTU dataflow architecture) |
| Process | Domestic process (specific node not disclosed) |
| TDP | 400 W (estimated) |
| Memory | 144 GB HBM3 (estimated) |
| Memory Bandwidth | 3.6 TB/s |
| Interconnect Bandwidth | 800 GB/s |
| FP8 Compute | Native FP8 support (specific TFLOPS not disclosed) |
| FP32 Compute | Not disclosed (estimated 30–50 TFLOPS) |
| FP16 / BF16 Compute | Not disclosed (estimated 200–400 TFLOPS) |
| INT8 Compute | Not disclosed (estimated 400–800 TOPS) |
| Interface | PCIe 5.0 x16 (estimated) |
| Release | 2025-07-27 (WAIC 2025) |
| Price | Not disclosed (estimated ¥60,000–80,000) |
Technical Highlights
- Fourth-generation training+inference integrated: Suisi 400 chip, simultaneously supporting AI training and inference, positioned as ultra-high-performance AI computing
- Ultra-large memory capacity: 144GB HBM3, supporting 100-billion-parameter large model training and inference
- Ultra-high memory bandwidth: 3.6TB/s memory bandwidth, surpassing H20 (4.8TB/s HBM3e actual is 4.8TB/s, L600 is 3.6TB/s)
- High-speed interconnect: 800GB/s interconnect bandwidth, supporting large-scale cluster expansion
- Native FP8 support: Adapting to new-generation AI model low-precision inference trend
- Performance surpassing H20: According to Enflame co-founder and COO Zhang Yalin, L600 compute performance targets and surpasses NVIDIA H20
Product Positioning
Enflame L600 is Enflame's fourth-generation AI chip, positioned as "ultra-high-performance AI computing", mainly used for training clusters and high-performance inference. As a training+inference integrated chip, L600 simultaneously covers training and inference scenarios, targeting NVIDIA H20.
Application Scenarios
- 100-billion-parameter large model training
- High-performance AI inference (LLaMA, GPT, DeepSeek, etc.)
- Training+inference integrated AI computing center
- Domestic H20 replacement scenarios
- Scientific computing and AI for Science
Reference Price
| Channel | Price | Description |
|---|---|---|
| Official pricing | Not disclosed | Released in July 2025, estimated ¥60,000–80,000/card |
| Channel estimate | ≈ ¥70,000 | Estimated based on H20 pricing ratio |
Related Products
- CloudBlaze T20 — Second-generation training card
- CloudBlaze T21 — Second-generation training card (OAM)
- CloudBlaze i20 — Second-generation inference card
- CloudBlaze S60 — Third-generation inference card