Inside Tegra 4’s GPU
From the announcement of its first Tegra SoC back in 2008, Nvidia’s greatest advantages in the mobile segment were its background with GPUs and platforms. The company’s close relationships with game developers was destined to be a boon too, since most mobile titles are easily characterized as mainstream. Increasingly, though, ISVs are utilizing powerful architectures to rival current-gen consoles, benefiting from more mature tools to fully exploit potent graphics engines.
Although we saw companies like Intel go after the power consumption of Nvidia's Cortex-A9-based Tegra 3, the company is eager to show its Tegra 4 as a solution designed with performance per square millimeter and, in turn, performance per watt in mind. In fact, Nvidia already has a reference phone design based on its Tegra 4i SoC called Phoenix. The two boards below fit into the 5" device and host different implementations of Tegra 4.
We already know that Tegra 4's GPU isn’t a unified shader design. Nvidia claims the time is simply not yet right to make that transition. And so, we still have separate programmable pixel and vertex shaders. The company also isn’t able to declare OpenGL ES 3.0 compatibility, though it’s emphatic this doesn’t adversely affect what developers are able to do with Tegra 4.
And so, the GPU in its newest SoC looks a lot like an evolution of Tegra 3, plus a number of improvements.
| Header Cell - Column 0 | Tegra 4 | Tegra 4i | Tegra 3 |
|---|---|---|---|
| Vertex Processing Engines | 6 | 3 | 1 |
| Pixel Pipes | 4 | 2 | 2 |
| MADs | 72 | 60 | 12 |
| Clock Rate | 672 MHz | 660 MHz | 416 /520 MHz |
| Fill Rate | 2.68 Gpix/s | 1.32 Gpix/s | 1.04 Gpix/s |
| Memory Interface | 2 x 32-bit | 1 x 32-bit | 1 x 32-bit |
| Memory Support | DDR3L-1866, LPDDR3-1866 | DDR3L-1866, LPDDR3-2133 | DDR3-1600, LPDDR2-1066 |
| Manufacturing | 28 nm | 28 nm | 40 nm |
Tegra 3 employs a single vertex shading unit with four FP32-capable cores. It also includes two fragment pipes, each with four cores capable of FP20 precision. The four vertex and eight pixel shaders is how we come to call Tegra 3’s GPU a 12-core design.
In contrast, Tegra 4 has six vertex processing engines with four “cores” each. Factor in clock rate differences (using 672 MHz for Tegra 4 and 520 MHz for Tegra 3), and that adds up to about 7.75x more vertex shading performance this generation.
Its four pixel pipes contain 12 shader “cores” each (that’s three ALUs per pipe, and four multiply-add units per ALU), adding up to 48. Assuming the same clocks, you’re again looking at 7.75x more fragment shader performance.
