The GPU and the Graphics Card

The "Vega 20" ASIC isn't laminated into the acrylic; it comes out, so we can take a close look at the GPU before moving on to the card itself.


The "Vega 20" is a multi-chip module GPU, much like "Vega 10" and "Fiji". AMD has for the past three generations used stacked HBM (high bandwidth memory), and these stacks need to sit in close proximity to the GPU. AMD has hence mastered an ingenious (and expensive) multi-chip-module design. The GPU die and HBM2 memory stacks of the "Vega 20" sit atop a silicon interposer. The interposer is a silicon die that enables microscopic wiring between the GPU die and memory stacks. Since the density of wiring is too high to be passed through the fiberglass package substrate, a combination of TSVs (through silicon vias) and microscopic bumps bind the GPU and memory to the interposer and the package.

The main die on the "Vega 20" MCM is the centrally located GPU die built on TSMC's 7 nanometer silicon fabrication process. Measuring 331 mm², this silicon packs 13.26 billion transistors, which is a similar transistor count to NVIDIA's TU104 graphics processor. The "Vega 20" GPU die features a 4096-bit wide HBM2 memory interface, which has double the bus width of "Vega 10". This means the GPU has to be surrounded by four memory stacks instead of two. On the Radeon VII, there are four 32 Gbit (4 GB) stacks built on a 10 nm-class node supplied by either SK Hynix or Samsung. We don't know the process node of the interposer, but given it was 65 nm for 4096-bit "Fiji", we can't imagine why it would change for "Vega 20". The interposer has no logic of its own and draws no power for itself. Its only purpose is to provide miniaturized wiring between GPU and memory stacks, and as such, there's nothing to be gained by changing its fabrication process if the desired wiring density has been achieved. A metal-reinforcement brace runs the periphery of the fiberglass package substrate to distribute mounting pressure from the cooler.


And now, the card itself! The Radeon VII is clad in aluminium with diamond-cut edges. Aluminium makes up both the cooler shroud and the backplate, giving the card a solid industrial look that feels great in hand. This is AMD's first reference-design card to feature a triple-fan design that rivals custom-design cards in the market these days. AMD picked up cues from rival NVIDIA, which built its "Turing" family of Founders Edition cards using a dual-fan cooling solution that competes both in looks and performance with custom-design cards. There are also some unique design callbacks to the Radeon "Vega" family, such as the illuminated Radeon Technologies Group logo "cornerstone" ornament and illuminated Radeon logo along the top face, next to the power connectors. The cooler features an aluminium fin-stack heatsink with a vapor-chamber base plate and flattened heat pipes ventilated by a trio of 100 mm fans. The fin stack is arranged such that hot air from the cooler is exhausted through the top and bottom sides (i.e., the side with the power connectors and the side facing the PCIe slot). The shroud is wide open along these two sides, with no slits obstructing airflow.


The Radeon VII graphics card draws power from two 8-pin PCIe power connectors. This setup is capable of 375 watts of power draw (although this does not necessarily mean that the Radeon VII pulls that much power). Display outputs are a combination of one HDMI 2.0b and three full-size DisplayPort 1.4 connectors, and all ports are capable of driving AMD FreeSync 2 HDR monitors. The backplate is made out of aluminium and has a central cutout near the GPU area and its mount-holes, and some vents surrounding it. There is a polyethylene sheet and rubber spacers at random locations between the backplate and the PCB to ensure the bare metal doesn't accidentally short anything.