Aiserver Manufacturer
1 / 5
| Product Introduction: | The Kunlun Chip P800 96GB is the third-generation cloud high-performance AI accelerator card launched by Kunlun Chip Technology under Baidu. It is built on the self-developed XPU-P architecture, focusing on large-scale AI model training and inference scenarios. It has outstanding advantages in memory specifications, computing performance, and ecosystem compatibility, and has been widely applied to intelligent computing demand scenarios in key industries such as finance, internet, and telecommunications. |
| Name: | Kunlun Chip P800 96GB |
| FP16: | 345 TFLOPS |
| INT8: | 512 TOPS |
| VRAM type: | HBM2E (High Bandwidth Memory, with approximately 30% higher bandwidth compared to HBM2, suitable for the high bandwidth demands of large model inference) |
| Chip fabrication process: | TSMC 7nm process, paired with 2.5D CoWoS advanced packaging (ensuring core heat dissipation and signal stability) |
| DirectX support: | Does not support DirectX or OpenGL for non-graphics rendering positioning, no game/graphics related API adaptation |
| Video memory frequency: | Effective frequency 9.375Gbps. By combining the HBM2 memory type with a 1.2TB/s bandwidth, the typical frequency of HBM2 memory in the same industry level is inferred. |
| Memory Bus Width: | 1024bit Based on a memory bandwidth of 1.2TB/s and an HBM2 effective frequency of approximately 9.375Gbps (bandwidth = bus width × frequency ÷ 8), it complies with the HBM2 specification characteristics. |
| Video memory speed: | 7ns (based on the physical characteristics of HBM2E memory chips, corresponding to an equivalent frequency of 4000 MHz, with latency better than GDDR6) |
| Output Interface: | No traditional video output interfaces (DP/HDMI/VGA), focused on data center inference/computing scenarios |
| Output interface type: | The core interface is PCIe Gen4.0 x16 (bidirectional bandwidth 64GB/s), supporting x86/ARM host interconnection. |
| Form Factor: | FHFL dual - slot |
| Weight: | 1064g |
