Gamers have endured a long wait for the Radeon RX Vega, but the wait is over, or at least nearly. Over the past couple of days, I've been learning about how AMD plans to re-enter the high-end graphics card market with its next-generation graphics architecture. The company revealed most of the details of its Ryzen Threadripper CPUs to us ahead of SIGGRAPH, as well, if you'd rather catch up with that news first.
|GTX 970||1050||1178||56||104||1664||224+32||224 GB/s||3.5+0.5GB||145W|
|GTX 980||1126||1216||64||128||2048||256||224 GB/s||4 GB||165W|
|GTX 980 Ti||1002||1075||96||176||2816||384||336 GB/s||6 GB||250W|
|Titan X (Maxwell)||1002||1075||96||192||3072||384||336 GB/s||12 GB||250W|
|GTX 1080||1607||1733||64||160||2560||256||320 GB/s||8GB||180W|
|GTX 1080 Ti||1480||1582||88||224||3584||352||484 GB/s||11GB||250W|
|Titan Xp||1480?||1582||96||240||3840||384||547 GB/s||12GB||250W|
|R9 Fury X||---||1050||64||256||4096||1024||512 GB/s||4GB||275W|
|Radeon RX Vega 64
|Radeon RX Vega 64
|Radeon RX Vega 56||1156||1471||64||224||3584||2048||410 GB/s||8GB||210W|
The high-level details of the Vega architecture have been known to us for some time, but the implementation of that architecture on Radeon RX gaming cards has remained a mystery until now.
AMD will be releasing the Radeon RX Vega with two different GPU configurations across three products. The fully-enabled Vega 10 GPU will find a home in the Radeon RX Vega 64 Liquid-Cooled Edition and the Radeon RX Vega 64. Both cards will get a GPU with 4096 stream processors, 256 texturing units, 8GB of HBM2 RAM running at a transfer rate of 484 GB/s, and 64 ROPs.
The RX Vega 64 Liquid-Cooled Edition will be the highest-performance Vega card at launch. This card will offer a typical boost range of 1677 MHz, a base clock of 1406 MHz, and a board power of 345W. It'll offer peak single-precision performance of 13.7 TFLOPS and peak half-precision performance of 27.5 TFLOPS.
The air-cooled RX Vega 64 will offer a typical boost range of 1546 MHz, a base clock of 1247 MHz, and a board power of 295W. Those figures are good for 12.66 TFLOPS of peak single-precision performance and 25.3 TFLOPS of half-precision throughput. Both of these RX Vega 64 cards are positioned to compete with Nvidia's GeForce GTX 1080.
The swanky aluminum-bedecked cards you see above are both limited editions, and AMD claims that label is genuine. At least for the air-cooled card, once the stock is sold through, the only way to get a reference air-cooled Vega will be with the black shroud you'll see below.
The most interesting RX Vega graphics card may be the previously-unknown RX Vega 56. As its name implies, the Vega 10 GPU on this card has 56 of its 64 compute units enabled, for 3584 stream processors in total. Interestingly, it'll still have all 64 of its ROPs, but it'll ship with only 224 texturing units enabled. This card will have a typical boost range of 1471 MHz and base clocks of 1156 MHz, and somewhat lower memory clocks resulting in a peak transfer rate of 410 GB/s. AMD claims it'll be good for 10.5 TFLOPS of peak single-precision throughput and 21 TFLOPS of half-precision throughput. This card will have a board power of 210W, and it's positioned to compete with Nvidia's GeForce GTX 1070.
We learned a wealth of new architecture details regarding the Vega GPU at this event, although the short window between the presentation of that information and the NDA lift for this article means that I'll be holding off on a deep-dive until our full Vega review. The Cliff's Notes is that some of the performance potential of the Vega architecture, like double-rate packed math, the draw-stream binning rasterizer, support for primitive shaders, and the High Bandwidth Cache Controller, are going to require driver optimizations or developer targeting (or both) to eventually run at their best. Early performance numbers for Vega from Frontier Edition cards didn't include any gains from the DSBR, for example, and that feature will be enabled for the first time with the Radeon RX Vega release drivers.