Single page Print

AMD's Radeon R9 Fury X graphics card reviewed

The red team vents its Fury

The Fury X is here. At long last, after lots of hype, we can show you how AMD's new high-end GPU performs aboard the firm's snazzy new liquid-cooled graphics card. We've tested in a range of games using our famous frame-time-based metrics, and we have a full set of results to share with you. Let's get to it.

A brief stop in Fiji
Over the past several weeks, almost everything most folks would want to know about the new Radeon GPU has become public knowledge—except for how it performs. If you've somehow missed out on this info, let's take a moment to summarize. At the heart of the Radeon R9 Fury X are two new core technologies: the Fiji graphics processor and a new type of memory known as High Bandwidth Memory (HBM).

The Fiji GPU is AMD's first new top-end GPU in nearly two years, and it's the largest chip in a family of products based on the GCN architecture that stretches back to 2011. Even the Xbox One and PS4 are based on GCN, although Fiji is an evolved version of that technology built on a whole heck of a lot larger scale. Here are its vitals compared to the biggest PC GPUs, including the Hawaii chip from the Radeon R9 290X and the GM200 from the GeForce GTX 980 Ti.

width (bits)
Die size
GM200 96 192/192 3072 6 384 8000 601 28 nm
Hawaii 64 176/88 2816 4 512 6200 438 28 nm
Fiji 64 256/128 4096 4 4096 8900 596 28 nm

Yep, Fiji's shader array has a massive 4096 ALU lanes or "shader processors," more than any other GPU to date. To give you some context for these numbers, once you factor in clock speeds, the Radeon R9 Fury X has seven times the shader processing power of the Xbox One and over seven times the memory bandwidth. Even a block diagram of Fiji looks daunting.

A simplified block diagram of a Fiji GPU. Source: AMD.

In many respects, Fiji is just what you see above: a larger implementation of the same GCN architecture that we've known for several years. AMD has made some important improvements under the covers, though. Notably, Fiji inherits a delta-based color compression facility from last year's Tonga chip. This feature should allow the GPU to use its memory bandwidth and capacity more efficiently than older GPUs like Hawaii. Many of the other changes in Fiji are meant to reduce power consumption. A feature called voltage-adaptive operation, first used in AMD's Kaveri and Carrizo APUs, should allow the chip to run at lower voltages, reducing power draw. New methods for selecting voltage and clock speed combinations and switching between those different modes should make Fiji more efficient than older GCN graphics chips, too. (For more info on Fiji's graphics architecture, be sure to read my separate article on the subject.)

This combination of increased scale and reduced power consumption allows Fiji to cram about 45% more processing power into roughly the same power envelope as Hawaii before it. Yet even that fact isn't Fiji's most notable attribute. Instead, Fiji's signature innovation is HBM, the first new type of high-end graphics memory introduced in seven years. HBM takes advantage of a technique in chip-building technology known as stacking, in which multiple silicon dies are piled on top of one another in order to improve the bit density. We've seen stacking deployed in the flash memory used in SSDs, but HBM is perhaps even more ambitious. And Fiji is the first commercial implementation of this tech.

A simplified illustration of an HBM solution. Source: AMD.

The Fiji GPU sits atop a piece of silicon, known as an interposer, along with four stacks of HBM memory. The individual memory chips run at a relatively sedate speed of 500MHz in order to save power, but each stack has an extra-wide 1024-bit connection to the outside world in order to provide lots of bandwidth. This "wide and slow" setup works because the GPU and memory get to talk to one another over the silicon interposer, which is the next best thing to having memory integrated on-chip.

With four HBM stacks, Fiji effectively has a 4096-bit-wide path to memory. That memory transfers data at a rate of 1Gbps, yielding a heart-stopping total of 512 GB/s of bandwidth. The Fury X's closest competitor, the GeForce GTX 980 Ti, tops out at 336 GB/s, so the new Radeon represents a substantial advance.

HBM also saves power, both on the DRAM side and in the GPU's memory control logic, and it enables an entire GPU-plus-memory solution to fit into a much smaller physical footprint. Fiji with HBM requires about one-third the space of Hawaii and its GDDR5, as illustrated above.

This very first implementation of HBM does come with one potential drawback: it's limited to a total of 4GB of memory capacity. Today's high-end cards, including the R9 Fury X, are marketed heavily for use with 4K displays. That 4GB capacity limit could perhaps lead to performance issues, especially at very high resolutions. AMD doesn't seem to think it will be a problem, though, and, well, you'll see our first round of performance numbers shortly.