AMD's Radeon Fury X architecture revealed

The Radeon R9 Fury X made its big debut earlier this week at E3, and we were there to cover it. Today, AMD has decided to release most of the rest of the public info about the Fury X card and the new GPU behind it code-named Fiji. I've been traveling back from E3 and attempting to wrap my head around the relevant details of this new video card and graphics processor at the same time. The picture is just coming together, and it tells an interesting tale. Let me see if I can walk you through what you need to know.

Fury X in the flesh

First, that's the Radeon R9 Fury X card pictured above. It's liquid cooled with an external radiator, as you can see. The rest of the card is pretty compact thanks to the external cooling apparatus and the small physical footprint of High Bandwidth Memory (HBM). (If you're not familiar with HBM, please do read my article on it, which explains exactly what it is and why it matters.) This card lists for $649—same as the GeForce GTX 980 Ti—and it is slated to go on sale in less than a week, on June 24.

The lower-end Radeon R9 Fury should follow on July 14, and those cards should offer a mix of air- and liquid-cooling options from a range of vendors. We don't yet know all of the R9 Fury's exact specs, but the Fury X card's relevant details look like so:

That's a lot of info to process, so let me draw your attention to a few specific details. The Fury X requires dual eight-pin PCIe power inputs, and the board's "typical" power draw limit is 275W. That's essentially the practical maximum without overclocking, as I understand it, and is 25W higher than the power rating for its chief rival, the GeForce GTX 980 Ti.  The similarity here is good news given that the Fury X's predecessor, the Radeon R9 290X, requires 290W to do its thing.

Of architectures and destiny
Beyond that, the numbers above outline many of the key architectural details for Fiji, and AMD's new baby is intriguing for multiple reasons. Here's a quick overview of Fiji's functional units in the form of a simplified block diagram.

Block diagram of the AMD Fiji GPU. Source: AMD.

If you know about the Hawaii chip in the R9 290X, then much of what's above should look familiar. Hawaii also has four main shader clusters, each with its own geometry processor and rasterizer. Each of Hawaii's clusters has four render back ends, and each of those has 16 pixels per clock of ROP throughput. The CUs, or compute units, are the basic building block of AMD's Graphics Core Next (GCN) microarchitecture. There are more of them here, 16 per cluster rather than 11.

AMD has made some deeper changes inside of a number of these functional units, and we'll discuss those soon. But big GPUs based on familiar architectures like this one are very much about scale. These chips have many copies of important resources, and the mix in each chip is typically a little different. The table below shows how Fiji's resources add up in terms of key graphics hardware.

Like pretty almost every major PC graphics chip since 2011, AMD's newest GPU is manufactured on a 28-nm fabrication process. Fiji is just under 600 square millimeters, only a smidgen smaller than the GM200 chip aboard the GeForce GTX 980 Ti and Titan X.

A quick glance at how Fiji compares with Hawaii shows some areas of dramatic growth alongside a few spots where Fiji is no more capable than Hawaii on a per-clock basis. Since we have final product specs, though, let's factor in clock speeds and see how key graphics rates compare, at least on paper. I've lifted this table from a recent review and added the Fury X.

Now we have a pretty good sense of things. In certain respects, Fiji has grown by roughly half-again compared to Hawaii, including peak shader arithmetic, texture filtering capacity, and memory bandwidth. That 512 GB/s of memory bandwidth comes courtesy of HBM, Fiji's signature innovation, and puts the Fury X in a class by itself in at least one department.

In other respects, including peak triangle throughput for rasterization and pixel fill rates, Fiji is simply no more capable in theory than Hawaii. As a result, Fiji offers a very different mix of resources than its predecessor. There's tons more shader and computing power on tap, and the Fury X can access memory via its texturing units and HBM interfaces at much higher rates than the R9 290X.

In situations where a game's performance is limited primarily by shader effects processing, texturing, or memory bandwidth, the Fury X should easily outpace the 290X. On the other hand, if gaming performance is gated by any sort of ROP throughput—including raw pixel-pushing power, blending rates for multisampled anti-aliasing, or effects based on depth and stencil like shadowing—the Fury X has little to offer beyond the R9 290X. The same is true for geometry throughput.

The Fury X substantially outruns the GeForce GTX 980 Ti in terms of integer texture filtering, shader math rates, and memory bandwidth, too, since the 980 Ti more or less matches the 290X in those departments. But the Fury X has only about 70% of the ROP and triangle rasterization rates of the big GeForce.

With Fiji, AMD is offering a rather different vision of how GPUs ought to be used by game developers. That's one reason I'd expect to see continuing fights between the GPU vendors over what effects folks incorporate into PC games. Nvidia will likely emphasize geometric complexity and tessellation, and AMD will probably push for prettier pixels instead of more polygons.