AMD Vega: Everything you need to know about the next GPU architecture
AMD is busy preparing to blow the doors off CES 2017 later today. While Polaris impressed us with its mainstream focus, the anticipation surround the enthusiast tier Vega chip is palpable. We’ve now been given our first ever glimpse of the Vega GPU, AMD’s top-end graphics chip which will power the next generation of Radeon graphics cards. Just like Polaris, Vega is manufactured on the 14nm process, although AMD Vega is about more than just specs. This isn’t a simple graphics card refresh with some clock speed bumps. Everything about the Radeon architecture is being redesigned from the ground up.
AMD Vega will officially be the first graphics chip to support HBM2 (second generation High Bandwidth Memory). This allows for up to 8x the capacity per stack and twice the bandwidth per pin, all while having a much smaller footprint than GDDR5. This is the start of AMD’s efforts to massively increase the amount of video memory available to us. AMD believes the growth in processing power is outpacing growth in memory capacity by a significant margin, and the games of the future will benefit from Radeon graphics cards not having to sacrifice their capacity.
In total we’re looking at at least two 4GB stacks of HBM2, providing a total of 8GB VRAM, although it’s likely we’ll also see a 16GB HBM2 premium configuration on day one. This has a total memory bandwidth of 512GB/s, with memory clock speeds up to 2Gb/s, up from 1Gb/s on the original HBM.
Another key difference is the introduction of High-Bandwidth Cache. That’s the new name AMD has coined for the VRAM, or the framebuffer. In the case of Vega 10 this is comprised of HBM2 memory, but this can in fact be other memory systems for cheaper costs. This will be managed by the HBC controller, handling GPU data, communicating with the rest of the system and ensuring a smooth flow. Its potential is huge, with a maximum addressable memory of 512TB. Radeon boss Raja Koduri calls it “the world’s most scalable GPU memory architecture.”
In essence what you are looking at is a more efficient manner of managing memory resources and ensuring you’re getting peak performance. AMD uses the example of The Witcher 3 and Fallout 4 at Ultra 4K, which are both currently allocated more than 8GB VRAM and yet actually consume less than half of this.
Taking efficiency a step further is Vega’s new programmable geometry pipeline. This is a nifty little tool which helps the GPU identify what geometry the player as actually able to see at any one time, only rendering the necessary parts. Less resources are wasted on rendering parts of a scene you can’t even see. The example Koduri gave is Deus Ex: Mankind Divided. Within the market scene there are around 220 million polygons rendered, yet only 2 million are visible to the player.
Also at CES, AMD showed Vega 10 along with a Zen CPU running both DOOM and Star Wars Battlefront at 4K/ Ultra. In the case of DOOM, Vega 10 is capable of hitting around 70 frames per second average, albeit with one heavy dip down to around 30 fps during a takedown. For point of reference, a GeForce GTX 1080 will struggle to hit 60 FPS average. Star Wars Battlefront meanwhile was bizarrely shown on a 4K monitor locked to 60Hz via V-Sync. Nevertheless, it was pulling in a solid 60 frames per second. This AMD Radeon finally delivering a competitive single-GPU 4K solution. From the looks of early benchmarks Vega 10 looks to be marginally more powerful than a top-tier GTX 1080.
Now for the downside. We didn’t get any graphics card reveals at CES 2017 from AMD, as some were perhaps expecting. Instead it’s looking as if we’re going to have to wait until the summer to see what Vega is really about. This will give AMD plenty more time to improve performance and efficiency even more. Unfortunately for AMD this means Nvidia is surely going to have its next generation of GPUs ready to launch soon after. The hype for Vega is very much real then, but even the most ardent of Radeon fans are going to struggle to wait yet another six months for the latest hardware considering what's available on store shelves right now.