Both Intel and AMD unveiled details on their upcoming graphics technologies at GDC last week, and we’ve rounded up the major announcements from both manufacturers and summarised the findings. Intel is adding additional capabilities to Haswell rather than simply boosting the number of GPU cores, while AMD is taking a major bet on cloud gaming, its “Never Settle” bundle, console sales, and an official Radeon 7990.
Intel: Slowly catching up
Intel’s GDC presentations focused on Haswell’s new graphics capabilities. The new core adds support for two DirectX extensions – PixelSync and InstantAccess. PixelSync is (according to Intel) a superior method for handling order independent transparency (OIT).
There’s some confusion about whether or not this is a new feature – AMD started talking up DX11 and OIT support four years ago with the HD 5000 family. DirectX 11 added support for doing OIT, but it doesn’t mandate a particular approach.
The problem with many OIT implementations is that they’re extremely expensive. If you play games with anti-aliasing enabled (especially if you’re like me, and you prefer super-sampled AA), you’ve almost certainly seen this problem, even if you didn’t realise what it was.
A game will run smoothly until you encounter a fog bank or smoke cloud – and when that happens, performance craters. Order-dependent transparency is less expensive, but prone to breaking – in the below screenshot on the left, objects are being rendered improperly.
What Intel has added with PixelSync is a new method of performing OIT calculations that it claims is far less expensive. PixelSync allows a game to force shader programs to run serially when working on a particular set of pixels. Forcing serial calculation ensures that shader operations are applied in the proper order.
PixelSync is an evolved version of an algorithm Intel first discussed several years ago. This new approach will require game engines to specifically support it, but the shift towards mobile gaming and Haswell’s improving GPU performance make it likely that developers will add support for this method of OIT. Intel commands too great a share of the total graphics market not to take it seriously.
InstantAccess is an Intel-specific implementation of what’s generally called “zero-copy.” The term refers to a scenario in which data is copied from one location to another without requiring the CPU to intervene. InstantAccess allows both the CPU and GPU to read and write data to the same locations in memory. Again, this is a technology Nvidia and AMD have supported for years, but Intel didn’t release enough data regarding its implementation to determine whether or not it offers any advantages over its competitors.
Finally, Intel announced that it’s working with HandBrake’s developers to bring QuickSync acceleration to the popular encoding program. Current GPU-assisted encode implementations are pretty terrible – hopefully HandBrake’s QS flavour will accelerate the task without compromising quality.
AMD takes to the sky, teases Radeon HD 7990
AMD’s GDC presentations focused on console wins, cloud gaming, and the upcoming Radeon 7990. The PS4, Wii U, and Xbox 720 all got explicit mentions. The company also announced its new Radeon Sky cards – GCN-based GPUs that are “cloud-optimised.”
Team Red also announced a partnership with CiiNow, a cloud gaming infrastructure provider. Its nascent partner is convinced that cloud gaming is the future (obviously) and that it can offer a superior experience to Gaikai or OnLive. The company offered an interesting breakdown of precisely where latency is added to game streaming, including a downright odd claim that it was actually faster to stream FEAR 3 than to play it on the Xbox 360.
Despite all the drum-banging, we remain dubious of cloud gaming’s long-term potential. As a consumer, the last thing I’m interested in doing is signing up for two or three services in order to play the games I want to access (the chances of every game coming to a single service is infinitesimal).
The problem with latency is that it’s variable by definition. AMD and Nvidia tease the idea of streaming high-end PC graphics to tablets and smartphones, but ignore how brittle Internet service can be. Actual performance depends on the time of day, whether anyone else is using the router in your home, how they’re using it, the type of Internet service you have, the general quality of the lines and equipment used to deliver it to you, and whether Mercury is in retrograde.
AMD’s big unveil (and worst-kept secret) was the Radeon 7990. Manufacturers like PowerColor and Asus have actually been shipping Radeon 7990s for over a year, but these cards weren’t sanctioned by AMD. Once the S10000 showed up (that’s AMD’s dual-GPU virtualisation solution), it was only a matter of time before the company built a consumer version of the card.
Close-up examination of the 7990 comes courtesy of Anandtech. This concludes that AMD’s official flavour is different from the current unofficial versions in at least one respect – it uses two eight-pin connectors with a maximum power draw of 375 Watts rather than the 500 Watts plus and three eight-pin connectors on some of the other cards.
Whether or not this will impact overall performance is unclear; the PowerColor and Asus cards are both designed for overclocking and are likely tuned with substantial margins of error where power consumption is concerned. The S10000, however, only offers 1792 GPU cores rather than the full 2048 that vendor versions of the HD 7990 have shipped with. Honestly, it’s a bit surprising that AMD is shipping a dual GPU this late in the HD 7000 family’s life cycle. It’s likely a belated response to the GTX Titan from Nvidia and an attempt to polish the Graphics Core Next family through the end of the year.
AMD’s public disclosures were very short on details concerning whatever graphics core it is that will replace GCN in late 2013 or early 2014; the company hasn’t released an updated HSA roadmap slide in over a year and GCN 2.0 architectural details are nearly non-existent.