First, the core parameters

1. Architecture and process

• Adopt 7nm process technology, based on GCN 5.1 architecture.
• The core code is Vega20, the number of transistors is 13.20 billion, and the core area is 331mm ².

1. Frequency and computing power

• Base frequency 1200MHz, acceleration frequency 1746MHz.
• FP32 single-precision floating-point performance is about 13.4 TFLOPS, FP64 double-precision is about 6.7 TFLOPS.

1. Stream Processor and Cache

• The number of stream processors is 3840, the L1 cache is 16KB, and the L2 cache is 4MB.

Second, video memory configuration

1. Memory capacity and type

• official parameters: The standard version MI50 is 16GB HBM2 video memory, with a memory width of 4096bit and a bandwidth of 1TB/s.
• Market customized version: 32GB video memory.

1. Memory frequency

• The video memory base frequency is 1000MHz (equivalent to 2Gbps), and the actual bandwidth depends on the HBM2 stack configuration.

III, Interface and Extension

1. physical interface

• Output interface: 1 mini-DisplayPort 1.4a.
• Power supply interface: dual 8-pin power supply, the whole card TDP power consumption is 300W.

1. extended support

• Supports PCIe 4.0 bus, providing higher data transmission efficiency.
• Support AMD Infinity Fabric Link technology to improve the efficiency of multi-card interconnection.

IV. Application scenarios and performance

1. Computing and AI performance

• Suitable for deep learning training, local large model inference (such as Qwen, DeepSeek, etc.), about 12-13 tokens/s in the measured Win + Vulkan environment, and up to 15 tokens/s in the Linux + ROCm environment.
• Compared with the RTX 3060, the performance of some scenarios is improved by more than 30%.

1. Game and rendering

• Brush into the custom BIOS, Master Lu runs 38- 440,000, can smoothly run "Black Myth: Wukong" and other games (high image quality 60 frames +).