AMD's Mantle finally emerges: Turns out it's actually for boosting low-end CPUs, not GPUs

This morning, AMD planned to launch its Mantle driver, with support for Battlefield 4 (that patch, from EA, should still be scheduled to go live at 4 AM EST) and with one technology demo: Star Swarm, by Oxide. Unfortunately, a last-minute delay by AMD due to driver issues has caused the company to hold back the shipping driver a little longer. So this is a preview of the performance data AMD has released thus far, with our own testing to follow tomorrow, featuring both BF4 and the Star Swarm demo.

Mantle miscommunication

There are some key elements of Mantle that haven't been communicated as clearly as they might have been. Mantle was announced at a GPU launch as an alternative to existing GPU APIs like OpenGL and DirectX. Most of the discussion of the API has therefore focused on how much Mantle would improve the performance of AMD's various graphics cards by allowing for a vastly higher number of draw calls to be submitted per frame.

The problem with this focus is that it implies that Mantle makes a game run faster by making the GPU more efficient. In reality, Mantle is aimed at improving the performance and reducing the workload on the CPU side of the equation. That means Mantle's performance improvements vis-à-vis DirectX 11 will depend on whether or not a game is CPU-bound or not.

According to AMD, these scenarios are actually quite common in gaming, as existing APIs have huge amounts of overhead and difficulty scaling out to multiple CPU cores. AMD expects Mantle to offer a principle benefit "for the majority of PC gamers that have entry-level and mid-range processors." (emphasis added). This means Mantle's greatest performance increases should be visible on APUs, particularly chips like the A10-7850K and even Kabini-based parts.

• Low validation and processing of API commands
• Explicit command buffer control
• Close to linear performance scaling from recording command buffers onto multiple CPU cores
• Reduced runtime shader compilation overhead

Where Mantle matters

AMD explicitly states that Mantle isn't necessarily going to improve performance much at the very top of the CPU/GPU performance stack. According to the company, "Mantle makes less of an impact in cases where high resolutions and 'maximum detail' settings are used, as these settings are likely to be maximally taxing GPU resources in a manner that is more difficult to improve at the API level (so-called GPU-bound scenarios)."

Mantle can still boost performance over DirectX 11 in such instances thanks to reduced command buffer submissions, asynchronous DMA queuing for data uploads, asynchronous compute capabilities, and advanced anti-aliasing techniques, but the company isn't making big predictions for the category yet.

Next page: AMD's performance claims...

Profiling Star Swarm

Star Swarm is a technology demo from Oxide Games based on the Nitrous engine. Nitrous is designed around a technology Oxide calls the Simultaneous Work and Rendering Model (SWARM). This engine is specifically designed to send draw calls to whatever CPU core is most available, meaning it's almost tailor-made to showcase the advantage of an API like Mantle. AMD's own documents note that Star Swarm is "a beautiful use case for Mantle."

AMD's performance claims for Star Swarm and BF4:

• AMD A10-7700K + Radeon R9 290X: +319% performance in "Extreme" mode (CPU-limited, 1920x1080)
• Intel Core i7-4960X + Radeon R9 260X: +5.1% performance in "Extreme" mode (GPU-limited, 1920x1080).

Dismantling Mantle

We're going to hold off on performance analysis until we have our own data to work with, but I don't mind telling you what I expect to see. Based on the results AMD has shared, Mantle is going to make its biggest difference with low-end CPUs. That means AMD APUs (both Kabini and Kaveri), and, to a lesser extent, the AMD FX-class of chips will see the biggest improvements. Remember that lifting workloads off those chips is probably most important.

We're going to be testing the same Core i3-4330 that we used for the Kaveri review, because that chip is a dual-core Intel with Hyper-Threading and it's possible that Mantle will also shine on this kind of configuration. Freeing up more CPU resources when a chip only has about 2.4 cores worth of execution capability may still give Intel users a reason to cheer as well. If that's the case, we expect Mantle to matter particularly in mobile, where Intel still ships a fair number of dual-core systems.

The alternative is that Intel chips are simply good enough in single-threaded code that Mantle doesn't prove a major benefit, in more than the lowest-end chips. That's a use-case we'll investigate as well, but it's not the outcome we'd like to see. Standards like Mantle have the greatest chance of achieving market penetration if they grant a benefit to a large number of customer systems -- and that' means it'll be important to demonstrate strong improvements across both Intel and AMD hardware.