China-based cloud and HPC system vendor, Shanghai Tianshu Intellectual Semiconductor, has unveiled its first 7nm ‘Big Island’ GPU-based graphics card. Currently, in mass production, GPUs are set to offer almost twice as high performance as regular offerings from competitors in half the matrix and with much lower power consumption.
Chinese Big Island GPU based on 7nm process node to deliver twice the performance of competitors in half of the door area, graphics card depicted with up to 37 TFLOP horsepower at 300W
During the board meeting, the company’s founder, Zheng Jinshan, unveiled his Big Island GPU, which has a monolithic design. The GPU is based on the TSMC 7nm FinFET process node and has a 2.5D CoWoS design considering that it packs the GPU and DRAM on the same door. The GPU consists of 24 billion transistors and has up to 37 TFLOPs of FP32, 147 TFLOPs of FP16 / BF16, 371.0 TOPs INT32, 147 TOPs INT16 and 295 TOPs INT8 calculations per second. The exact core specifications and clock speeds are unknown, but they should be revealed when the launch closes.
In comparison, the AMD Instinct MI100 has a top FP32 speed of 23.1 TFLOPs and a top BFLOAT16 speed of 92.3 TFLOPs. The NVIDIA A100 GPU, on the other hand, has a top FP32 speed of 19.5 TFLOPs with Tensor cores pushing it to 156 TFLOPs and 312 TFLOPs with sparsity. The top BFLOAT16 speed of the NVIDIA A100 GPU is rated at 312 TFLOPs and 624 TFLOPs with sparse.
China Big Island GPU Specification Comparison
| GPU name | AMD Instinct MI100 | NVIDIA A100 | Big Island |
|---|---|---|---|
| Process node | TSMC 7nm | TSMC 7nm | TSMC 7nm |
| Architecture | CDNA 1 | Ampere | Unknown |
| Transistors | 50 billion | 54 billion | 24 billion |
| Kernels | 7680 | 6912 | TBC |
| Memory | 32 GB HBM2 | 40 GB HBM2 | 32 GB HBM2 |
| Memory bandwidth | 1.2 TB / s | 1.6 TB / s | 1.2 TB / s |
| FP32 Compute | 23.1 TFLOP 46.1 TFLOPs (matrix) |
19.5 TFLOPs 156 TFLOPs (Tensor) 312 TFLOPs (sparsity) |
37 TFLOPs |
| BFLOAT16 Calculate | 92.3 TFLOPs | 312 TFLOPs 624 TFLOPs (sparsity) |
147 TFLOPs |
| TDP | 300W | 400W | 300W |
In terms of memory, the China Big Island GPU has 32 GB of HBM2 memory that works with up to 1.2 TB / s bandwidth. The GPU also has a PCIe Gen 4 x16 interface that has a bandwidth of up to 64 GB / s. As for the products themselves, the Big Islan GPU will come in both PCIe and OAM form factors.
The China-based company presented images of both designs. The GPGPU-based PCIe graphics card has a standard dual-slot design with a large aluminum heatsink under a silver-colored cover. The card has a passively cooled design considering that it will be integrated into servers with a lot of active cooling. The card is powered by a single 8-pin connector and has a 300 W TDP. The company also lists 60A MOSFETs features on the card. OAM offers ships with a standard mezzanine board form factor, and the 2U stand shown has the OAM card with a solid heat sink made of aluminum with several copper heating pipes passing through it.
The company has not given any exact delivery dates or prices for the Big Island GPU, but it is likely that it will be launched sometime by the end of this year or early next year, given that it has already moved to mass production. We expect more details from
