AMD might be having a tough time on the CPU front, but when it comes to graphics cards it’s taken a leap ahead of the competition. Its new single-GPU flagship – the Radeon HD 7970 – is the first to use a 28nm manufacturing process, and beats rival Nvidia to the punch. The impact is significant: the HD 7970 die is 378mm2 compared to the 389mm2 of its predecessor, the Radeon HD 6970, and AMD crams in 4.3 billion transistors – a huge jump over the 2.6 billion included in last year’s top-end single GPU. It compares favourably to Nvidia’s GeForce GTX 580, too, which includes 3 billion transistors in a 520mm2 package.
AMD has also given its Very Long Instruction Word 4 (VLIW4) architecture the boot, deeming its lacklustre parallel performance a hindrance. While VLIW4 cores and their schedulers proved adept at handling groups of identical operations concurrently, they struggled with varied groups of tasks required by more complex applications. Some tasks were scheduled and processed promptly, but often the scheduler couldn’t keep up, with bottlenecks created by instructions being left behind.
It’s a big change. VLIW-based architectures, including VLIW4, have been used in AMD graphics cards since the Radeon 9700 in 2002. Instead, the HD 7970 uses multiple instruction multiple data cores (MIMD). These are constructed from several single instruction multiple data cores (SIMD) grouped together, and are capable of efficiently handling a more diverse range of tasks, as well as making dynamic changes to the compute schedule – something VLIW4 couldn’t do.
Each MIMD package is made from 64 SIMD cores, and each package has its own L1 cache, with L2 cache and memory controllers shared between several packages. AMD has given these cores their own name, too, with the marketing department swooping into action to dub each unit a Graphics Core Next.
The HD 7970 includes 2048 SIMD cores inside 32 MIMD clusters, which is more than the 1536 stream processors used in the HD 6970. The core clock is 925MHz, there’s 3GB of GDDR5 RAM running at 1,375MHz, and the memory bus is 384 bits wide. The latter is an improvement on the 256-bit bus of last year’s cards, and on a par with the GTX 580.
In our 1920 x 1080 Very High quality Crysis test, the HD 7970 scored 65fps: 11fps faster than the single-GPU GTX 580, 17fps quicker than the HD 6970, and only 2fps short of the dual-GPU Radeon HD 6990. Adding 4x anti-aliasing saw the HD 7970 score 59fps – still 11fps better than the Nvidia – and running it at 2560 x 1600 returned a 43fps result, 15fps more than the GTX 580.
An average of 76fps in Just Cause 2’s highest settings, at 1920 x 1080 and with 8x anti-aliasing, is 2fps quicker than the GTX 580 and more than twice the speed of the HD 6970. It’s an adept overclocker, too: we boosted the core and memory clocks to 1100MHz and 1500MHz and found the 43fps score in our toughest Crysis test rose to 49fps.
The HD 7970 hit a reasonable 79˚C during our tests, but its peak power draw of 394W in our test rig is more than the 292W pull of the GTX 580 and the 377W draw of the HD 6970. It’s louder than the Nvidia card, too, although we anticipate board partners will release cards with quieter coolers.
You’ll also have to make sure your PC can handle the card. It’s dual height and 281mm long, and needs both eight-pin and a six-pin power adapters. The backplate has only one DVI-I output alongside HDMI and two mini- DisplayPort sockets, with more room given over to pumping out hot air.
There’s an elephant in the room, though: even the cheapest cards cost over $700 – which makes it the most expensive single-GPU card on the market. Nvidia’s GTX 580 can be bought for less than $550, and the Radeon HD 6970 is cheaper still at less than $400.
That puts the HD 7970 beyond all but the wealthiest gamers, and its performance means you’ll need multiple monitors to truly exploit the horsepower. However, the revamped technology will trickle down to affordable models, and when that happens we’ll be first in line.