Vega has been just beyond the horizon since Polaris was first announced. We knew there was something more than just Polaris 10 and 11, something bigger and more fitting with the direction that AMD has been going in recent years. They’ve built big chips that were fast, if power hungry, contenders.

AMD Vega 10

Vega is on the horizon, and coming fast

The big die approach isn’t necessarily the smart approach given that it also means more power and higher heat dissipation requirements. But when properly implemented with a good VRAM framework, it could spell trouble for the competition. Such a massive GPU could also help give a boost in the deep learning world as well. More cores that are more efficient equals massive performance.

So what exactly is Vega, then? It’s been shrouded in mystery despite the best attempts at rumor mill sites attempting to capitalize on your hopes and dreams. What we actually, positively know is a lot less. Thankfully AMD has been kind enough to give us a small glimpse at what we can expect a fully enabled chip to be with their announcement of Radeon Instinct. With this we can see where they seem to be focusing.

That focus, on compute based workloads, has been a few years in the making. GCN was primarily built to leverage massive parallel workloads and thus was perfect for other scientific compute applications. In the right hands it also meant a very good platform to play games. Their most recent iteration of the big GPU didn’t quite wow us like we wanted it to. Perhaps we gave into the hype, as I myself did, while waiting for its release. HBM was a necessary step forward even if it ended up being a limiting factor due to the low amount of actual VRAM that was connected. Polaris reversed this by using the fastest version of GDDR5 available, and slapping on 8GB of the stuff. We’ve also all but officially confirmed the use of HBM2 through the rumor mill, meaning that the amount of VRAM won’t be nearly the limiting factor it once was. Bandwidth, too, will be appreciably increased over its predecessor, potentially climbing to 1024GB/s of memory bandwidth.

Vega 10

In the slides, we see mention of a high bandwidth cache. Just what does that mean? Is it a pool of NAND connected via PCIe, SATA or something even faster? Is it a massive amount of GDDR5X or even HBM2? And what does that mean for the consumer version. Regardless, memory is certainly not going to be an issue with the advent of less expensive means of production.

And the actual improvements to GCN? It seems that AMD is hinting at something new, or at least much changed over the previous arch. Maybe even something with a more improved front end, better scheduler etc. Time will tell about the actual architecture, of course. We’ll all get a chance to see it in all its glory.

A 300W passively cooled, long(ish) GPU with unspecified raw performance numbers sounds delicious. Given that the Fiji-based Radeon Instinct GPU already can produce 8.2 TFLOPs of single-precision power, it stands to reason that the more expensive option would be even better. That and it supposedly can compute packed math at an even greater rate. Sure, that’s a feature for the sciences, though it could point towards better means of computing raw data and the math that we see in games. And at 300W, we certainly hope that the MI25 itself is at least proportionally faster, if not even more so given the improvements that are bound to be inside.

Vega is coming, and unlike Polaris, this is designed to tackle more complex scenarios and likely provide us consumers with a very nice performance boost over previous generations. Raja and friends haven’t been sitting on their hands with the Radeon Technologies Group, they’ve been slowly improving their architecture, of which Polaris was only the beginning. Vega is looking, as the rumors and general information seem to imply, like a very good contender for the highest-end Pascal GPU. Even if NVIDIA is waiting to release the GTX 1080 Ti until AMD makes their next move, Vega 10 might be a very lucrative move indeed.