High-resolution PCB PicturesThese pictures are for the convenience of volt-modders and those who would like to see all the finer details on the PCB. Feel free to link back to us and use them in your articles or forum posts.
High-res versions are also available (front, back).
Circuit Board (PCB) Analysis
The GPU VRM is 10-phase, controlled by an International Rectifier IR35217 controller, which is among the best controllers available on the market. Do note the two empty solder pads for additional phases, which could be populated on AMD's Radeon Instinct MI60 with the full 4096 shaders of the chip active.
A second IR35217 is used to control two additional voltages, each with a two-phase design. I did check, and the four phases are not interconnected, which makes this a 10+2+2 VRM design.
Unlike the first-generation Vega 10, AMD has chosen to use four HBM2 memory stacks on Radeon VII, which doubles memory bus width to 4096 bit, while doubling total available memory bandwidth for over a terabyte per second as well! This design choice is also the reason why Radeon VII comes with 16 GB of memory instead of less, like 8 GB or 12 GB. At this time, only 4 GB and 8 GB HBM2 memory stacks exist; smaller or odd capacities are theoretically available through custom designs, but at much higher cost. Having four stacks is required to achieve the desired 4096 bit bus width, so the only memory size options are 16 GB and 32 GB. In theory, it could be possible to create a 3-stack 12 GB version with 25% lower memory bandwidth, resulting in a performance hit; another caveat is that we don't know if the Vega memory controller can even run with a non-power-of-two number of stacks. Our review sample uses Hynix HBM2 memory (H5VR32ESA4H-N0C)
AMD's Vega 20 graphics processor is the first GPU in the world that's produced on a 7 nanometer production process. This not only reduces die size, but also enables higher clock frequencies with reduced power usage. Vega 20 is produced at TSMC Taiwan, using 13.2 billion transistors on a die size of 331 mm².
The pictured GPU chip actually consists of three different kinds of silicon. The large GPU chip you see in the middle, the four HBM2 memory stacks and, below both, covering the whole inner area, an interposer, which is a special chip design that only serves to connect the microscopic memory interface pins with their counterparts on the GPU die.