Looking Forward to AMD’s Vega GPU
AT AMD’S TECH SUMMIT in December, we got some concrete details and demos for Zen (now Ryzen) and Vega. Ryzen should be available around the time you’re reading this, but Vega is further out, with a release target of “first half, 2017”—probably June/July,
Along with working silicon, AMD provided the first official details on Vega, and it’s shaping up to be
(ahem) out of this world. It builds on everything that makes Polaris great, but it’s not simply a bigger chip with more cores. AMD didn’t provide Vega’s core count or clock speed, but it will likely be 4,096 cores clocked at around 1.5–1.6GHz. The reason we can be so specific is that AMD also announced a new line of machine intelligence accelerators, called Radeon Instinct MI6, MI8, and MI25. The MI25 uses Vega, and will provide up to 25 TFLOPS (with FP16—half that for FP32), which means Vega should be 45 percent faster in gaming performance than the Fury X.
But it’s not just about TFLOPS. AMD has reworked several elements of its GCN architecture, a major one being the memory subsystem. Vega includes 8GB (possibly 16GB) of HBM2 memory in two stacks. These deliver the same 512GB/s bandwidth as the four stacks of HBM1 in Fiji, but with two stacks, the silicon interposer doesn’t need to be as large. AMD isn’t just calling this HBM or VRAM; it’s now a High-Bandwidth Cache (HBC), and there’s also a new High-Bandwidth Cache Controller (HBCC). The distinction is important, because the HBCC plays a much more prominent role in memory accesses. AMD profiled VRAM use for TheWitcher3 and
Fallout4. In both, the amount of VRAM allocated is two to three times larger than the amount of VRAM accessed. The HBCC takes this into account, allowing the GPU to work with far larger data sets, with a 512TB virtual address space. AMD also demonstrated a real-time physically rendered image of a house using more than 600GB of data, running on what I assume is an 8GB Vega card. If the HBCC works properly, even a 4GB card could behave more like an 8–12GB VRAM card, while an 8GB card would equal a 16–24GB one.
Vega also has a new geometry pipeline. AMD notes that there can be a 100x difference between polygons in a scene and those that are visible. The new geometry engine will have over twice the throughput per clock compared to AMD’s previous architecture. The compute unit is also improved, with native support for packed FP16 operations, which should be useful for machine learning applications. AMD’s Mike Mantor also stated, “We’ve spent a lot of time tuning and tweaking to get our frequencies up significantly and power down,” though the Radeon Instinct MI25 still has a “<300W” TDP.
Finally, AMD improved the pixel engine, with a new Draw Stream Binning Rasterizer that helps cull pixels that aren’t visible in the final scene. All the render backends are also clients of the cache now, reducing the number of memory accesses. This should provide significant performance improvements with deferred rendering engines, which is what many modern games are using.
Based on performance numbers, Vega would be impressive, but factor in the other changes and AMD’s DX12/ Vulkan performance, and we’re looking at another exciting year in graphics cards. The GTX 1080 leads the Fury X by around 35 percent (less at 4K), so a 45 percent boost would put Vega well ahead, and if the architecture improvements can add another 10– 15 percent, Vega might even exceed Titan X. For graphics junkies like me, June can’t come soon enough.
If the HBCC works properly, a 4GB card could behave more like an 8–12GB VRAM card.
Jarred Walton has been a PC and gaming enthusiast for over 30 years.