When AMD announced its new Mantle graphics API last September, it promised that we’d see support for the new standard by December, courtesy of Dice’s Battlefield 4. Then BF4 launched in November as a colossal bug-fest with game-breaking problems. Initially, Dice and AMD promised that the update would ship on-time, before finally acknowledging that the API wouldn’t be ready until January. Now that date has been pushed back further, with a launch set for “sometime next month” (February).
Update: We’ve spoken to Robert Hallock of AMD who has categorically denied that the patch will drop in February. Since the rest of this post deals with the performance expectations around Mantle and background on the demos we’ve seen to date, we’ll leave it in its current form.
While the continued delay is disappointing, our high-level view of the situation hasn’t changed — it’s more important to fix the game, improve stability, and make it fun for everyone than it is to deliver a patch that improves performance in a title that no one actually wants to play. That said, we’ve gotten a bit more information than we had previously on just how much performance uplift AMD expects from Mantle. As part of its Tech Day presentations at CES, AMD released additional performance data and some new slides on what to expect.
Keep in mind that this performance uplift is based on a demo rather than a shipping title. AMD’s overall claim is that Mantle is “up to 45% faster than DirectX.” That gain is spread between the high-end R9 290X card family and the lower-end APUs. Mantle will assist in both situations, but having seen the performance characteristics of Kaveri’s GPU, I’ve tempered my expectations somewhat as to how much performance AMD can pick up in the lower-end. The reason is simple: For all that Mantle does a lot of great things to relieve pressure on the CPU and accelerate draw calls, Mantle can’t fix the bandwidth limitation that’s throttling Kaveri, hard.
Performance testing shows a nearly linear relationship between faster DDR3 and better Kaveri APU performance, even above DDR3-2400 clock speeds. That’s been a hallmark of AMD’s APUs since Llano, but the effect is more pronounced with Kaveri — Llano realized the majority of its gains from DDR3-1600, whereas Kaveri is still hungry for more bandwidth. I think Mantle will still improve APU performance, but the gains may sit more in the 10-15% range in shipping titles rather than the 20-30% we could see in discrete graphics.
The extent of the performance gains Mantle can offer, therefore, will depend on both memory bandwidth limitations and how much bandwidth is being chewed up by the API. It’s significant that despite showing a 3x performance gain for Mantle in one specific instance, AMD’s actual claim about Mantle performance is far smaller — up to 45%, not 3x. It stands to reason that the company has set this more modest range because the Oxide demo (embedded above), while impressive, isn’t necessarily representative of the actual expected gain across a broad set of titles. EA hasn’t changed its long-term commitment to Mantle and still intends to roll out the API for multiple titles based on Frostbite 3, so the long-term benefits (however large or small they happen to be) are still on track for 2014.
