One of the main attractions at this years CES certainly was AMD presenting details on their upcoming Vega GPU architecture. It’s just great to see that AMD appears to have new products coming that might really be able to keep up with the competition - talking about Vega and Ryzen. Nevertheless we here at ocaholic are still sceptical since in the past AMD created so many false expectations, that we believe numbers when we tested a product ourselves.
As we already mentioned AMD was talking in more detail about the architecture of their upcoming Vega cards. Apparently the majority of people was hoping AMD would actually show or even announce Vega graphics cards, but since the chip is still in a pretty early stage, talking about the architecture was about the best things they could have done thus far.
So the questions are what has AMD done to improve the meanwhile five year old GCN (Graphics Core Next) architecture and what effect might it have on performance?
At this point we have to mention we’re not going to talk about all the different changes in depth, since that would mean this news was to become endless. Instead we will focus on what we think are the most important architectural changes. Apart from that talking about effects on performance those will all be speculations based on the benchmarks AMD has shown on their live demo systems.
Like Fiji Vega is making use of HBM memory, but this time it’s the second generation, HBM2, they’re putting right next to the main GPU, which makes this - again - a multi-chip module. In combination with a completely overhauled memory hierarchy, which should help in greatly optimizing memory accesses, this should improve how the already very high memory bandwidth is being addressed.
AMD has made changes to the entire rendering pipeline with the goal of making the chip capable of using clock cycles in a more efficient ways, thus increasing IPC performance in general. Undoubtedly one of the biggest changes comes with the added support for 8-bit and 16-bit ops. In theory the so called Rapid Packed Math feature should allow the Vega NCU (Next Compute Unit) to run four times as many operations per clock cycle than the GCN architecture was capable of.
Overall AMD allocated a lot of time to finding solutions in how to prevent wasteful usage of resources in general. In other words Vega is supposed to greatly improve the efficiency of the entire chip, when compared to the five year old GCN architecture.
Now looking at the the actual performance demos AMD had systems on display at CES running Doom 3 as well as Star War Battlefront. The resolution was set at UHD and according to other websites details levels in both cases were set to ultra. Under these conditions a system with an AMD Ryzen CPU and a Vega based graphics card was capable of pumping out something between 60 and 75 fps. This is definitely a very respective figure and massively quicker than what Pascal or Fiji are capable of. It’s actually Nvidia Titan X level, but as AMD was raising false expectations so many times in the past, we now finally believe values when we tested something ourselves. Apparently we’re hoping these are really the performance levels Vega can achieve, since that would mean AMD is finally competitive again. That in the end then implies that prices of graphics cards might be coming down to more reasonable levels.
After all it will be most crucial for AMD to release Vega as soon as possible. We’re sincerely hoping that Vega is going to be announced somewhen during the first quarter this year, since Nvidia might bring their Pascal refresh already in Q2 and those cards will raise the bar once more.