Chip designer ARM has announced its next generation mobile GPU, the Mali-T658. ARM boasts that the design will offer ten times the performance Mali-400 MP, found in, among other smartphones, the Samsung Galaxy S II.
The T658 is ARM's second GPU using an architecture it calls Midgard. Midgard is designed to support both 3D workloads using modern APIs—it supports both OpenGL ES and Microsoft's Direct3D 11—and computation workloads using OpenCL, Microsoft DirectCompute, and Google RenderScript. To aid with compute tasks, Midgard supports full IEEE 754 floating point.
The first Midgard design, the T604, was announced last year, with companies including Samsung and LG licensing the GPU. Midgard has a design built around three kinds of pipelines: arithmetic, load/store, and texture. Each T658 core has four arithmetic pipelines, and one each of load/store and texture, doubling the number of arithmetic pipelines found in T604. Up to eight cores can be integrated into a single GPU, again representing a doubling in performance relative to T604, which allows up to four cores to be ganged together.
The result is performance that ARM claims is comparable to that of the PlayStation 3. Mali-400 MP is specced as offering about 30 million triangles per second (a crude measure of GPU performance), and with T658 offering ten times the performance, that gives it about 300 million triangles per second. The PlayStation 3 has performance of about 250 million-300 million triangles per second.
ARM isn't the only competitor in the mobile GPU space. The current king of the hill is the PowerVR SGX 543MP2 found in the iPhone 4S, which has two cores offering about 35 million triangles per second each. Next-generation parts include PowerVR's Series 6, with up to 380 million triangles per second, and Marvell's Armada 628, with up to 200 million triangles per second.
Though these mobile GPUs are starting to rival console graphics processors in raw triangle count, these days people demand more of their GPUs than simply pumping out flat triangles. The console processors boast far more memory bandwidth than the mobile GPUs are ever likely to be coupled with, so for complex applications consoles processors will still retain their edge. High-end desktop GPUs can spit out more than 2 billion triangles per second.
T658 is designed to integrate easily with both current ARMv7 designs such as Cortex A9 and A15, and future 64-bit ARMv8 designs. T604 parts are due to hit the market next year, with Samsung's Exynos 5250 set to be the first. T658 GPUs are likely to arrive at some point in 2013.