Home / Shop / IBM / Networking (IBM / Lenovo)


- Description
- Specifications
Spare part 02YG118 is the ThinkSystem NVIDIA A100 40GB PCIe Gen4 Passive GPU, which Lenovo identifies as option 4X67A13135 with feature code BEL5. The option consists of one NVIDIA A100 accelerator with a full-height 3U adapter bracket already fitted, plus its documentation.
The accelerator uses the NVIDIA Ampere architecture and carries 512 third-generation Tensor Cores together with 8192 FP32 CUDA cores. Its 40 GB of HBM2 memory runs with ECC and reaches a bandwidth of 1,555 GB/s. Peak rates are 9.7 TFLOPS for FP64, 19.5 TFLOPS for FP32 and 312 TFLOPS for half precision, and sparse models can double the tensor figures. Multi-Instance GPU can divide the card into as many as seven isolated instances of 5 GB each.
The adapter connects through a PCIe Gen 4 x16 interface rated at 64 GB/s and fills a full-height, full-length slot that is two slots wide. Maximum power consumption is 250 W for this 40 GB PCIe model. Cooling is passive, and three NVLink bridges join a pair of cards at 600 GB/s. Operating temperature runs from 0 to 50 degrees C at 5 to 85 percent relative humidity. CUDA, DirectCompute, OpenCL and OpenACC are supported as compute interfaces.
Servers that take 02YG118 include ThinkSystem SR650 V2 (7Z72, 7Z73), SR655 (7Y00, 7Z01), SR665 (7D2V, 7D2W), SR670 V2 (7Z22, 7Z23) and SR860 V2 (7Z59, 7Z60). Earlier SR650 (7X05, 7X06) and SR670 (7Y36, 7Y37) models take it too. The 80GB sibling is option 4X67A76715, supplied without a CEC chip.
| Spare part number | 02YG118 |
| Manufacturer | Lenovo |
| Option number | 4X67A13135 |
| Feature code | BEL5 |
| GPU architecture | NVIDIA Ampere |
| Memory | 40 GB HBM2 with ECC |
| Memory bandwidth | 1,555 GB/s |
| Tensor cores | 512 third-generation |
| CUDA cores | 8192 FP32 |
| Host interface | PCIe Gen 4 x16 |
| Form factor | Full height, full length, double width |
| Maximum power | 250 W |
| Cooling | Passive |
