For the past 18 months, AMD has been playing a defensive game in the GPU market. The Radeon HD 7970, which launched nearly two years ago, was initially met with strong reviews, but the card’s $ 549 price tag raised eyebrows. Then, Nvidia delivered the first 600-series cards based on the GK104 GPU. Suddenly, it was AMD with the hotter, slower, more power-hungry solution, and Nvidia on the ascent. Since then, AMD has fought back with tweaked GPU speeds, the impressive “Never Settle” bundles, and more aggressive pricing. It’s focused on improving driver compatibility and solving problems related to frame latency. It’s done everything, in short, but release a new GPU.
Today, that changes.
R9 290X hardware specs
The R9 290X is a GCN-based GPU with a little English and a lot of Miracle-Gro. The card’s full specs are available below:
There’s a lot to dig into here. First, as with Nvidia’s Titan, the R9 290X is a larger version of the underlying architecture, with 44 CUs and 2,816 cores, up from 32 CUs and 2,048 cores. At first glance, the difference between the R9 280X and 290X is much smaller than the gap between the GTX 680 and the GTX Titan. When Nvidia launched Titan, it packed in 75% more GPU cores, for a total of 2,688 cores, 224 texture mapping units (TMUs) and 48 ROPs (render outputs). The GTX 680, in contrast, offers 1536 cores, 128 texture mapping units, and 32 ROPs. So with Titan, cores and TMUs both jumped 75%, ROPs grew by 50%, and memory bandwidth increased by 50%.
AMD has chosen a different approach. The R9 290X has 37.5% more cores than the R9 280X, 37.5% more TMUs, but fully twice the ROPs — from 32 to 64. It can handle four primitive operations per clock rather than two, which means its tessellation performance should be higher. The memory bus has been increased to 512 bits wide, but the RAM itself is clocked more slowly, at 1250MHz, down from 1500MHz on most HD 7970/R9 280X cards. Total memory bandwidth, therefore, is only modestly increased, to 320GB/s, up from 288GB/s. Overclocking tests suggest the GPU can handle another 10% to both core clock and RAM — in line with expectations, but nothing extraordinary.
The other major difference between the Radeon R9 290X and the R9 280X is the addition of new asynchronous command engines (ACEs). Asynchronous command engines are important for GPGPU computation workloads — each ACE can handle its own stream of parallel instructions. The old HD 7900 family and the R7 260X, R9 270X, and R9 280X all have two ACEs per GPU. The R9 290X has eight. This is almost certainly a direct result of customization work AMD did for the PS4, Xbox, or both, and it should allow the chip to execute far more threads in parallel in certain situations. We’ll examine what some of those might be in tests a little further on.
Along with the major changes, there’s a host of minor ones. The Hawaii GPU that powers the R9 290X incorporates a host of minor improvements that could arguably classify it as a GCN 1.1 chip, like the R7 260X/Radeon 7790. It’s also the first high-end GPU from AMD or Nvidia that has no need of a Crossfire external bridge. AMD has promised that the GPU can handle multi-GPU combination without one thanks to dual DMA engines that are compatible with new frame pacing technology and handle rendering without external penalty. Unfortunately, AMD wasn’t able to provide us with a second GPU for testing this new system in Crossfire mode.
Next page: Games benchmarking
- 1 of 4