RX Vega Is a PowerHungry Graphical Beast
LATE LAST YEAR, AMD revealed many of the details behind its Vega architecture. It all sounded great on paper, and frankly I wanted Vega hardware right then. But it wasn’t fully ready for release, so we had to wait… and wait. The consumer versions of Vega
On paper, there’s a lot of good stuff inside Vega. New features include the High-Bandwidth Cache Controller (HBCC), reworked shader pipelines, new instructions, far more VRAM than the previous high-end Fiji design, and a Draw Stream Binning Rasterizer (DSBR). These are designed to allow Vega to outperform all other AMD GPUs, and the result is clock speeds up to 1,677MHz on the liquid-cooled RX Vega64, or 1,546MHz on the air-cooled variant.
Combined with 4,096 shader cores, that gives Vega 13.7/12.7 Tflops of raw FP32 compute performance. Nvidia’s GTX 1080 Ti delivers up to 11.8 Tflops, and the Titan Xp up to 12.2 Tflops. That means RX Vega is going to outperform the 1080 Ti, right? Not so fast….
AMD’s GCN architecture has always delivered less gaming performance per Gflops of theoretical performance. The 980 Ti is rated at only 5.6 Tflops, compared to the Fury X at 8.6 Tflops, yet in practice the 980 Ti ends up around 5 percent faster (using my benchmarks across a large suite of games). At the unveiling of the RX Vega, AMD made it clear that its competitive target is the GTX 1080, and it might not even come out ahead consistently there.
But the kicker is the power requirements. The liquid-cooled RX Vega64 card will use up to 345W, the air-cooled RX Vega64 295W, and the only “powerfriendly” variant will be the Vega56, which gets down to 210W by drastically cutting clock speeds. (There will also be a Vega Nano at some point, but AMD hasn’t given any specifics.) In contrast, the 1080 has a 180W TDP, and the 1070 is 150W.
Why isn’t Vega the 1080 Ti killer we hoped for? Fundamentally, it all goes back to the architecture. AMD’s GCN architecture was designed to be good for compute, and it remains a strong point. Low-level access via APIs such as OpenCL, DX12, and Vulkan often allow developers to tap into the full power of the GPU, while DX11/OpenGL drivers end up being far less effective. But the key to the potential of the GCN architecture lies within the schedulers that queue up work for the GPU cores.
The Asynchronous Compute Engine (ACE) has been part of AMD’s GCN architecture since the first HD 7970 rolled off the foundry floor in 2012, but it wasn’t until Windows 10/DX12 that the real potential was even realized. Having the ability to be more flexible in scheduling work for the GPU is beneficial, but more robust scheduling is a tradeoff between power (watts) and flexibility. We’ve seen this with 290X versus 780 Ti, Fury X versus 980 Ti, and now Vega64 versus 1080. AMD is close to Nvidia in performance but less efficient in power.
AMD counters the disappointing power draw and performance by going for reasonable prices. While the liquid-cooled Vega64 card carries a healthy premium at $699 (and comes in a pack that includes some games and other incentives), the standalone air-cooled Vega64 will set you back $499, and the Vega56 will cost $399. That’s right in the ballpark of GTX 1080/1070 pricing, which is good, but if Nvidia performs even a bit faster, while using quite a bit less power, this is a clear win for the green team.
After a year of waiting since the GTX 1080 launched, AMD’s answer seems to be similar performance using 50 percent more power. AMD fans must now await Navi, and hope it can buck the trend.
If Nvidia performs a bit faster, while using less power, this is a clear win for the green team.