After many rumors and countless days speculation and waiting AMD is finally breaking the silence and reveals their new next generation Vega GPU architecture. Let’s have a look at what AMD has to show us today shall we.
Vega GPU architecture, the next from AMD
In past few years AMD has focused on 3 areas with their GPU. Gaming, workstation workload and compute. And Vega is embracing these 3 things even more. This time AMD wanted to make something revolutionary, efficient and very fast and being able to scale at rate which other rendering immerse workloads scales in. Vega features Primitive Shaders, which are built for next generation APIs, as well as improved load ballancing, making it more efficient. AMD has also completely reworked their Compute Unit design, which they have used since 7000 series Tahiti architecture, and now they call it Next-generation Compute Unit(NCU) and is now the core of Vega architecture.
After bunch of statements such as “We want to make decisions based on exabytes of data in an instant” AMD takes us to a bunch of different graphs showing how things have spiked in recent years.
The first one shows us game install size increase over years and AMD shows Deus Ex: Mankind Divided with about 20 Gigabytes of install data as an example.
The second one shows us the increase of Pro graphics asset size increase with the highest one currently being a movie “The BFG” with insane graphics asset size of 3 Petabytes of data! (1 Petabyte equals 1 000 000 Gigabytes)
The last two ones reflect the increase in training set size for compute workloads and GPU Compute increase and how it has increased so much compared to GPU storage size increase.
Memory scalability
So what are the new things in Vega architecture? Well AMD has completely redesigned their Compute Unit design, which they have used since 7000 series Tahiti architecture, and now they call it Next-generation Compute Unit (NCU) and is now the core of Vega architecture. AMD claims to have built the “World’s most scalable GPU memory architecture” into this new architecture. As expected Vega is going to utillize HBM2 with completely new memory subsystem which AMD now calls “High-Bandwidth Cache” controlled by new High-Bandwidth Cache Controller (HBCC).
This controller dynamically stores what is needed when it is needed making memory more efficient. HBCC breaks down memory space into tiny pages/tiles which can be individually managed. It predicts what resources are needed and allocating those based on whats there, compared to typical games which just load resources into memory and just keep it there even when they are not using them. HBCC is also able to handle up to 512 Terabytes of virtual address space.
Next-generation Compute Unit, compute and more.
Vega features brand new Next-generation Compute Unit. It has configurable double precision rate. 512 8-bit operations per clock, 256 16-bit operations per clock or 128 32-bit operations per clock. It is optimised for higher clockspeeds and high IPC. It also packs a lot of new things and features.
First one is new Rapid Packed Math. AMD didn’t go into detail on this new feature however what it basically means is that while every NCU is 32-bit it can do 16-bit operations at effectively double the peak Flops performance and not reduced.
NCU also features next generation Pixel Engine and Draw Stream Binning Rasterizer designed to improve performance and save power. The render back-ends of NCU are now clients of L2 cache which supposedly with applications which use deferred shading.
Geometry engine also got an improvement. Now it features new Programmable Geometry Pipeline with more than 2x peak throughput per clock. However we do have to note that it is over 2x(2.6x to be precise) over Fury X, and not Polaris.
Conclusion
So this is all AMD has to show today. A lot of exciting new things to come which on paper do sound really great but in the end the final performance matters. So how fast is it going to be? Who knows, but it’s safe to say that it will be interesting.