AMD reveals Vega architecture at CES 2017 - Here's why it matters

AMD announced a large-scale countdown several days ago that ended this morning with a preview of Vega, the new chip architecture for their line of high-end gaming and professional GPUs. Unfortunately, the preview is just that, and doesn't give us much information on the actual cards we'll be seeing when Vega officially launches.

What we do have is a heck of a lot of information about how AMD plans to implement their newest memory solution, which could change the way developers and gamers alike look at computer hardware for years to come. Sadly, we won't have our hands on a working model until sometime in H1 2017, which is closer than it sounds in the grand scheme of things.

Vega and a 512TB Memory Limit

Vega's newest punchline concerns a rework of the way AMD handles video memory for their high-end GPUs based on the HBM tech we've seen in their previous generations of cards. This rework involves replacing the old cache with a carefully engineered high-bandwidth cache and high-bandwidth cache controller, which enable Vega GPUs to access SSDs as a form of extended video memory. This means that Vega GPUs could effectively bypass the idea of maxing out of video memory by making use of the free space on the SSD as a temporary source of VRAM.

This technique is easily one of the most revolutionary ideas rolling around the video card industry right now, but it does come with more than a few caveats from the more traditional idea of GPU memory. The big one is that most SSDs are dramatically slower than any of the GPU memory on the market, with speeds often measured in Mb/s rather than the Gb/s you see for high-end GDDR5.

The key is that AMD's Vega tech likely won't be using the SSD as its only source of VRAM, but instead use it as a high capacity cache for extremely large game files. Vega will make use of the new cache and controller to intelligently load less essential assets that don't require the speed of traditional video memory to load up to a limit of 512TB of data. 512TB of data is massive, and likely overkill for the foreseeable future, but AMD is intent on making this memory architecture nearly infinitely scalable compared to todays gaming standards.

Considering your average triple-A title hovers around 40-80GB of content, a 512TB limit seems appropriately future-proof, especially as SSD technology continues to improve. For now, this limit will likely be most beneficial to professional applications targeted by Radeon's Pro GPUs, but will likely become invaluable to developers targeting 4K content and beyond.

Additionally, although AMD hasn't broadcasted any official announcements concerning Vega's memory specifications, information leaked during their DOOM 4K gaming demo utilizing Vega and Ryzen components showed an 8GB payload of their ridiculously fast HBM2. Which, considering 4GB of HBM gave AMD enough juice to compete in enthusiast-level GPUs last year, doubling that memory and combining it with a 512TB HB cache seems like a master stroke that should make this year's bout of GPU brawls more than competitive.

Teaching Vega to be Smarter and Compute Faster

As if HBM2 and a 512TB cache weren't enough to get you excited, it's important to note that AMD is keen on optimizing Vega to be more efficient than ever. The boys in red spent a great deal of time poring over the bloated rendering that most GPUs deal with on a daily basis in triple-A titles.

Traditionally, games will render hundreds of millions of polygons in any given scene that are never seen by the player, either because they're hidden behind other objects or layered within other assets. These polygons are largely extraneous and can eat up a lot of a GPU's brainpower. Which is why, with Vega, AMD is focusing on using a programmable geometry pipeline to feed information to the GPU about the player's perspective.

This allows the GPU to intelligently use vertex shaders to determine what polygons should have render priority before passing it on to the geometry engine. The result is a GPU that can identify the most important polygons in a scene and render them first and foremost, rather than wasting resources on assets the player can't even see. Vega also manages to get this job done in half the time thanks to a newly optimized Intelligent Workgroup Distributor that focuses on allocating resources more efficiently earlier in the render process.

Additionally, Vega takes the time to do another pass before actually rendering the pixels of a scene to check for pixels and additional assets that the player won't be able to see and therefore don't need to be seen. This is accomplished via a “draw stream binning rasterizer” that streamlines the process of rendering a scene by identifying the pixels that are flat-out non-essential and telling the GPU to avoid processing them at all.

The result is a GPU that can intelligently determine what assets in a scene the player can actively perceive and focus resources on processing those assets, rather than wasting time on non-essential tasks. It's a “work smarter not harder” mentality that will hopefully give AMD a serious edge over the competition, and could signal a change in the winds for the optimization issues AMD has been known for in the past.

Fueling all of these changes is a new and improved compute engine AMD claims is capable of 512 8bit operations per tick, 256 16bit operations per tick, and 128 32bit operations per clock. The former two will reportedly give Vega an edge in certain artificial intelligence task operations, while the latter gives Vega plenty of juice to bring to the party when it comes to rendering high fidelity scenes. 

Vega on the Horizon

It's going to be a big year for AMD, especially with not just Vega but Ryzen also coming to the table. It's important to remember that even if we were a little bit stiffed by actual hardware specs at this convention that AMD still gave us a pretty decent amount of information about the hardware at large, and we'll likely see a lot more roll out in the coming months leading up to a full release.

We've seen Ryzen and Vega working together at CES 2017 and the results are extremely promising. Running DOOM at 4K and easily over 60FPS is a tough task, and it's definitely something perfectly catered to AMD's biggest strengths. So, although it might not be definitive proof that AMD is ready to steal the show from the boys in green, it is an exciting concept to see in action and can only mean that gamers will benefit in the ongoing war between the two biggest names in the GPU game. 

As it stands right now, AMD has some momentum – let's just hope their hardware can live up to the hype it's been building over the last few years.