Conclusion

AMD's Vega GPU architecture indeed heralds some of the biggest changes to GPU design from the company's stable in a while. With Vega, AMD has taken on some of the fundamental shortcomings of its GPU architectures in the area of memory management while building on its strengths in the area of compute.

Memory management has historically been AMD's architectural bane. Rival NVIDIA has been able to make do with comparatively narrower memory buses, while making use of the available bandwidth better. At least the perception that memory is bottlenecking performance was never allowed to come up for NVIDIA. AMD, on the other hand, has been throwing in brute memory bandwidth to overcome its fundamental memory management issues, and was beginning to hit a wall where throwing physical bandwidth at the problem was either getting to costly or too energy inefficient. A case in point is its 512-bit GDDR5 memory interface with "Hawaii" and the industry's first HBM implementation to take on GPUs with relatively narrower 384-bit GDDR5 memory interfaces. With Vega, AMD is addressing many of the memory management issues that have been intrinsic to its Graphics CoreNext architecture, with the introduction of its High Bandwidth Cache technology.

The company has, simultaneously, built on its architectural strengths over NVIDIA, in the areas of its compute machinery. The industry has taken a liking for the powerful and open Compute Unit design of AMD Graphics CoreNext, and with Vega, the company improved these compute units with even better instructions per clock (IPC) thanks to Rapid Packed Math, support for ISA FP16 micro-ops, which will prove big as game developers handle several of their shaders and effects as low-bandwidth FP16 ops. AMD has also revamped its rendering pipeline to be more efficient, with conservative rasterization and an improved geometry processing pipeline.

As with the "Fiji" silicon from 2015, which powered the pioneering Radeon R9 Fury series, the "Vega 10" silicon implements exotic technology like high-bandwidth memory on-package, sitting on a silicon interposer. AMD, however, does not want to be caught unprepared for a price-war with NVIDIA where it couldn't sell the R9 Fury series below a certain price. The "Vega 10" silicon has been more cleverly built. It has just two memory stacks instead of four, while retaining memory bandwidth and doubling memory capacity. The package itself is smaller.

AMD has carved out three consumer SKUs from the "Vega 10" silicon, which are priced to compete more with NVIDIA's GeForce GTX 1080 and GTX 1070 based on the company's third-biggest GP104 silicon (after the GP102 and GP100), and so for AMD marketing, it's back to the alt-left marketing strategy of giving the most for the least money and bringing the most in features to the performance segment. We only hope AMD doesn't restrict itself to this segment and builds bigger GPUs based on this otherwise technically advanced architecture.