Nvidia announces the Tegra X1, 256 CUDA cores in tow



Nvidia has unveiled its next Tegra SoC, a significantly beefier CPU/GPU combo dubbed Tegra X1. The SoC is built upon the foundations of its predecessor but swaps the 32-bit ARM Cortex-A15 cores for brand ARM Cortex-A57s baked into a quad core arrangement backed by another quad core island with ARM Cortex-A53 cores which effectively brings the core count to eight.

The graphics departments also gets a major overhaul, the single Kepler SMX has been replaced with the newer and more power efficient desktop-class Maxwell 2 architecture. This time however, Nvidia opted for 2 SMMs inside a single cluster (GPC), each of which sports 128 unified shaders or stream processors (advertised as CUDAcores’) giving it a grand total of 256 shaders. Interestingly enough, Nvidia is still using the heavily abused ‘core’ term as a fancy PR/marketing name to sell the product.


The SoC is built on TSMC’s 20nm process but the cores won’t quite work the usual big.LITTLE, Nvidia has opted for cluster migration which was previously seen on early Exynos 5 Octa SoCs instead of averaging global task scheduling, this essentially means that only a set of cores will be active at any given time but Nvidia nevertheless asserts that most of performance impacts of cluster migration will be negligible as their SoC is fully cache coherent.



API-wise the Nvidia Tegra X1 will support OpenGL 4.5 along with the scaled down OpenGL ES3.1 (and AEP) and DirectX 12 (via FL11_1), 60FPS 4K video content is also supported. Clock speeds for the CPU cores were not disclosed but the GPU clocks will probably be sitting around the 1GHz mark while still being power efficient, thus allowing the chip to have better at half and double precision performance. Nvidia also claims that the move to the Maxwell 2 architecture brought a 2X performance per watt improvement over the Kepler architecture used in the Tegra K1.

Sampling availability hasn’t been announced but we’ll surely see the Tegra X1 in action in a few months time.

via Nvidia, PR