Six weeks ago, AMD announced its Radeon Pro Duo GPU, a dual graphics solution based on the company’s 28nm Fiji graphics chip. The Radeon Pro Duo is formally launching today, at a cool $1,500, but it’s explicitly intended for VR developers and isn’t expected to be available in wide production.
Still, $1,500 GPUs don’t launch every day, so let’s take a look at what AMD has cooked up here.
The Radeon Pro Duo is a standard dual GPU design at the board level. The two GPUs are connected by a PCI Express bridge (the PLX 8747, which provides 48 lanes of PCIe 3.0 connectivity — x16 for each graphics card, and an x16 lane for the motherboard). We don’t have the exact physical dimensions of the board, but the PCB is significantly smaller than you’d expect for a dual graphics card thanks to the use of HBM memory.
Formal specs on the GPU itself are below:
As expected, the Radeon Pro Duo is a “full” Fury X implementation and doubles up on the compute units, texture units, and ROPs. Top-end clock speed is listed as “up to” 1GHz, and the card’s rated TDP is supposedly 350W.
This is a little odd, to say the least. The original Fury X had a list TDP of 275W and two eight-pin PCI Express power connectors to power the card. 350W is only 27% more power consumption than 275W, despite adding a third eight-pin PCI Express connector (each PCIe 8-pin is capable of providing up to 150W of additional power).
Our guess is that the GPU is capable of drawing substantially more than 350W over brief periods of time, but that AMD set its thermal and power guidance at that point and will throttle the GPU to keep it in that specification. This might explain why the company didn’t ship a consumer variant of the card — just as Nvidia’s Titan Z was never as fast as two Titan Black’s due to lower effective clock rates on the GPU, the Radeon Pro Duo may not hit the sustained frequency targets necessary to beat a Fury X in Crossfire.
The point behind the Radeon Pro Duo is to drive developer engagement and get more game authors thinking about using multiple GPUs for VR rendering. Dedicating one GPU to each eye is a natural fit for VR, and the multi-GPU rendering methods introduced in DX12 make it much easier to support this kind of offload.
AMD also wants to make a more serious play for the professional graphics development sphere. While VR doesn’t directly tie to the company’s FirePro division, if VR content gets hot in the next few years, it could see worthwhile dividends from staking out an early position in this market.
Consumers are going to want to want to wait-and-see what the summer (and Polaris) deliver. But if you’re a developer wanting to work in VR, and create software that can run at peak performance on multi-GPU systems, the Radeon Pro Duo is meant to appeal to you. Niche products like this can pay big dividends if they shift the conversation around an emerging market, and that’s obviously what AMD is hoping will happen with this GPU.