Kunlun R300
Kunlun R300 is an OAM accelerator module (OCP-OAI standard) built on the 2nd-generation Kunlun XPU-R chip by Kunlunxin (a Baidu company). Same silicon as the R200 full-height PCIe card but in module form: R300 modules populate the R480-X8 server baseboard (8 modules per board) for large-scale data-center training and inference clusters.
Core Specifications
| Spec | Value |
|---|---|
| Architecture | Kunlun Gen-2 XPU-R (proprietary architecture, on-chip video codec) |
| Process | 7nm |
| FP16 | 128 TFLOPS |
| INT8 | 256 TOPS |
| FP32 | 32 TFLOPS |
| Memory | 32 GB GDDR6 |
| Memory bandwidth | 512 GB/s |
| TDP | 150 W |
| Form factor | OAM accelerator module (OCP-OAI standard, UBB baseboard) |
| Chip-to-chip | 200 GB/s (dual-ring topology inside R480-X8) |
| Launch | 2021-08 (with Kunlun Gen-2) |
| Price | Undisclosed |
📌 Form-factor clarification (2026-09 cross-validation): The R300 is not a new chip — it is the OAM module variant of Kunlun Gen 2. Kunlunxin's product matrix lists the R200 as the PCIe card, the R300 as the OAM module line, and the R480-X8 as the 8-module server baseboard; compute specifications are identical across all three (FP16 128 / INT8 256 / FP32 32 / 32GB GDDR6 / 512GB/s / 150W).
Highlights
- OCP-OAI standard: UBB server baseboard carrying 8 OAM modules (R480-X8) delivers roughly 1 PFLOPS FP16 per server
- Integrated codec: on-chip video encode/decode enables decode-plus-AI-inference pipelines without host data shuttling
- Hardware virtualization: per-chip virtualization improves cloud resource utilization
- 10K-card cluster building block: a foundational compute unit of Baidu Cloud's 10K-card Kunlun cluster (lit up in 2024 — China's first at-scale cluster on proprietary silicon)
Positioning
Kunlun product lineage: Kunlun Gen 1 (K100/K200 PCIe cards) → Kunlun Gen 2 (R200 PCIe card / R300 OAM module / R480-X8 server) → Kunlun Gen 3 → P800 (2024, third-gen training card). The R300 serves Baidu Cloud and China Mobile-class carrier AI data centers (Kunlun-based servers took the top share in China Mobile's 2025-2026 inference-type AI server tender, CUDA-ecosystem segment).
Use Cases
- Data-center-scale training and inference (R480-X8 8-module servers)
- Search, recommendation, and video-analytics inference for Baidu's core businesses
- Carrier and xinchuang AI data-center tenders
- ❌ Single-card developer setups (choose the R200 PCIe form factor)
Related Cards
- Kunlun P800 — Kunlun Gen-3 training card
- Kunlun 1 — first-generation chip
- Baidu Kunlun R200 — same silicon, PCIe form factor
- NVIDIA T4 — inference-tier comparison
- Huawei Ascend 910B — domestic training-card rival