ARM already has a lot going for it. Today it has added another piece to its architectural chessboard, the Mali-T604 embedded GPU.
For those of you in the back who were so wrapped up in the ATI-Nvidia desktop duels, you should know that a few years ago ARM sought out a small Norwegian company called Falanx who was developing GPUs for handheld devices.
Skip forward a few years and Falanx is fully integrated into the ARM business and thriving at the opportunity to inject some Nitrous Oxide into the current mobile/handheld/low-power device market, with a new graphics architecture dubbed "Midgard".
The Mali-T604 GPU is the first in a series of products from ARM's new Midgard architecture. It shares some common traits with its predecessor the Mali-400MP, but it design more closely resembles that of the Cortex-A15. What we're sure about is the fact that it's a solid kick in the pants for embedded graphics processing.
Straight up, ARM promises to increase performance by a factor of 5 over its Mali-400MP predecessor, with support for 16x FSAA (4x without any performance hits), wide API support (OpenGL ES 1.1, OpenVG 1.1, OpenCL 1.1, 1.2 and DirectX) and scalable core design which lets you build a SoC with anything from 1 to 4 shader cores with a single shared 32KB L2 cache. The GPU itself features a tri-pipe design that has ARM engineers all hot and bothered.
Mali-T604 is also designed to integrate perfectly with the Cortex-A9 and future company CPU designs such as the Cortex-A15 'Eagle', by sharing needed resources with the SoC across the AMBA4 interface without getting in the way of the CPU. This allows ARM to reduce bandwidth consumption within the SoC.
GPGPU done small
While ARM doesn't spell it out as clearly as the other achievements, one of the most interesting features on the new GPU is its ability to perform GPGPU computation. Through a full OpenCL 1.1 implementation, the new GPU can share computational tasks with the CPU with little effort. DirectX support isn't described, but is merely mentioned in the information provided. We are aware that this includes texture compression support and possibly DirectCompute. It's only a matter of time until someone develops a killer-app as some would put it.
So far there are no specifics on the manufacturing process, but having adopted the Cortex approach to chip design, ARM licensees should be able to shrink the design to whichever fabbing process suits them best.
With the Mali-T604 GPU, ARM has a single architecture that will take on anything from a 2.8-inch interface to a wide-screen TV or LCD. Need more power? Pile on the scalable graphics cores. If ARM is at fault of anything, it's the lack of snappy codenames for their graphics parts.