ENRIGIN D20
ENRIGIN's new-generation fully domestic AI inference accelerator, launched November 2025. Uses a dual-chip-per-card architecture (two proprietary AI chips on a single PCIe card, connected via PCIe 5.0 bifurcation at 2×x8) with chip design, fabrication, and packaging all completed domestically — the first China-designed, China-fabricated, China-packaged cloud AI inference card to reach volume production. The D20 Max carries 256GB of memory, and appliance partners (Ping Gao's AI Station) run DeepSeek 671B at full capacity.
Core Specifications
| Spec | Value |
|---|---|
| Architecture | 2x ENRIGIN proprietary AI chips (dual-chip card, direct bifurcation interconnect) |
| Process | Undisclosed |
| INT8 | 320 TOPS (whole-card, dual-chip combined) |
| Precision support | FP32 / TF32 / BF16 / FP16 / INT8 |
| Memory | 128 GB LPDDR5X |
| Higher variants | D20 Pro 128GB / D20 Max 256GB per card (4-card aggregate up to 1TB) |
| Memory bandwidth | Undisclosed |
| TDP | 145 W (whole card) |
| Interface | PCIe Gen5 x16 (2x x8 bifurcation, single slot) |
| Video codec | 256-decode / 40-encode channels (1080p@30fps) |
| Cooling / form | Full-height full-length single-slot PCIe card; active/passive cooling (FH3/4L); ECC |
| Launch | 2025-11-11 (chip in production since 2025-09) |
| Price | Undisclosed |
📌 Sourcing note (2026-09 cross-validation): INT8 320 TOPS / 128GB LPDDR5X / 145W / PCIe 5.0 figures come from ENRIGIN's official product page and launch coverage (Elecfans, Securities Times, China Daily) — multiple sources in agreement. FP16/BF16 peak compute is not officially published and is left unlisted per the no-estimates policy. INT8 320 TOPS is the whole-card dual-chip figure (per-chip ≈160 TOPS is an inference, not an official figure, for reference only).
Highlights
- Dual chips per card: two chips in a single slot with direct interconnect — no PCIe Switch needed — doubling compute and memory density while cutting cost and power
- Massive memory: 128-256GB LPDDR5X designed for hundred-billion-parameter model inference (memory-first strategy compensating for process-node gaps)
- Full-stack software: proprietary software stack with 200+ mainstream models; migrating from GPU to D10 requires a one-line change, and D10→D20 requires none
- Liquid-cooled appliance ecosystem: Ping Gao AI Station (single/dual/quad-card) and a 4U 16-card AI server (5 POPS INT8, 4TB memory per machine)
Positioning
ENRIGIN was founded in November 2022: D10 (volume production May 2025 — first fully domestic end-to-end AI accelerator) → D20 (November 2025, dual-chip) → T800 (flagship training chip, production 2026). It targets government, finance, healthcare, and rail-transit private deployments, with a deep appliance partnership with Ping Gao (688227); Qwen 235B and DeepSeek 685B are adapted.
Use Cases
- Private large-model inference (32B on a single card; full 671B on appliance clusters)
- High-concurrency inference for AI data centers (4TB-memory server configurations)
- Search/ads/recommendation, computer vision, and speech-to-text workloads
- Xinchuang environments (Hygon CPU + Kylin/openEuler OS certifications)
- ❌ Large-model training (D series is inference-only; training awaits T800)
Related Cards
- Enflame S60 — same-generation domestic inference card comparison
- Iluvatar Zhikai 100 — domestic inference card comparison
- Kunlun P800 — domestic inference/training comparison
- NVIDIA L4 — inference-tier comparison