Skip to main content

AMD reveals details of new Vega GPU architecture, goes all-in on HBM2

AMD’s Radeon RX 460, 470 and 480 were well received when released in 2016, and for good reason. The cards performed well, at reasonable prices. But the red team’s focus was strictly on the budget to mid-range market, which left gamers who want a powerful card without a good option from the red team.

That option still isn’t out yet, but now we know more about the architecture it’s based on, code-named Vega. AMD calls it the “world’s most scalable GPU memory architecture,” and that claim is not hot air. The company is making a number of tweaks to try and rid video cards of bottlenecks and build a foundation that can easily be modified for workloads with different demands.

The headline feature is the long-rumored adoption of High Bandwidth Memory 2 (HBM2). The original incarnation of HBM was picked up by AMD’s Radeon Fury line, and it gave those cards excellent memory performance. HBM2 does the same.

Compared to GDDR5, HBM2 can squeeze more memory into a smaller space. AMD says that means it offers twice the bandwidth “per pin,” and a 50 percent smaller footprint. We already saw that benefit in the Fury line. Those cards were very powerful, but also small – though AMD hasn’t said anything about the size of the final Vega cards.

Adopting HBM2 is great, but good memory is only as good as the GPU architecture its connected to. To solve that problem, AMD is using a high-bandwidth cache controller with a virtual address space of 512TB.

The over-arching goal of Vega’s design is to improve memory efficiency. AMD says most game developers code games to load far more data than they need as a hedge against hesitation, in case a game asset is called for but not found in memory. This is why modern games often require four gigabytes of memory at their highest detail settings. The faster memory design of AMD should allow quicker delivery of assets when called for. That, in turn, means games won’t need to consume as much of Vega’s memory.

Memory efficiency may be Vega’s most important enhancement, but it’s not alone. There’s also a new geometry pipeline, embracement of primitive shaders, and improved load-balancing. All these changes aim to reduce the need to draw assets that won’t actually be visible once the final frame of a game is viewed.

And then there’s Vega’s next-generation compute unit. AMD’s tactic here is clever. Rather than increasing precision of math, AMD is decreasing it – or, rather, giving developers that option. Math calculated at a lower level of precision can be processed more quickly, and many calculations in games don’t require a high level of precision. For example, Vega quotes 128 32-bit operations per clock, per compute unit. But if computed at 8 bits, it can handle 512 operations per clock.

These improvements are not all the new platform includes, but they’re by far the most important. The changes to the memory architecture, and the flexible compute precision, should prove the most important. If scalability is the goal, these changes seem like a great way to go about it, and they may solve memory issues that are becoming increasingly troublesome in modern games. On the other hand, nothing here tells us how fast Vega will be in raw compute capability, and that’s been the weakness of AMD’s high-end video cards. They simply haven’t matched Nvidia’s fastest.

While we now know a lot more about the architecture, we still don’t know anything about availability and pricing. AMD refuses to say anything firm about either. If past announcements are any guide, that means we’re at least a few months away from availability. But hey –- at least AMD’s fans can sate their appetite on Ryzen.

Editors' Recommendations