Imagination NN cores implement variable data resolution

Imagination NN cores implement variable data resolution

Technology News |
By Peter Clarke

The AX2185 and AX2145 offer inferencing capability on mobile applications and support a high degree of flexibility in terms of data resolution across neural networks, one of the more notable features of the architecture.

The PowerVR 2NX  has been designed as a core for inferencing on a mobile SoC from the ground up. It is a scalable architecture that comes with up to 8 neural network compute engines. Each engine is capable of up to 128 16bit multiply accumulates per clock (MACs/clk) and 256 8bit MACs/clk, giving at the high end 1024 16bit and 2048 8bit MACs/clk.

The AX2185 targets the high-end smartphone, smart surveillance, and automotive markets. It features eight full-width compute engines and provides 2,048 MACs/clock (4.1 Tera Operations Per Second).

The AX2145 targets mid-range smartphone, digital TV and set-top box, smart cameras and consumer security markets. The PowerVR AX2145, which offers up to 1.0TOPS has a streamlined architecture intended to deliver neural network inferencing for ultra-low bandwidth systems. Imagination declined to provide TOPS-per-watt figures for the two cores.

Imagination has taken a broad view of its approach to neural networking acceleration with series of drivers and low-level APIs, SDKs and high-level APIs, compilers, debuggers and analytical tools  that interface between the hardware and open-source and third-party neural network framework tools.

In addition these software tools can direct neural networking to a variety or resources including general purpose processors, GPUs such as Imagination’s own PowerVR range as well as the Series 2NX accelerators. Both cores support the Android Neural Networks API (NNAPI), used by developers to bring neural network capabilities to Android-based mobile devices.

Next: Vary the data resolution

The AX2185 and AX2145 both support bit depths from 16-bit to 4-bit and this can be varied from layer to layer within a neural network as appropriate resulting in higher performance at lower bandwidth and power. “We do also support floating point both internally externally for interoperability with other processors,” said Francisco Socal, product manager for vision and AI at Imagination Technologies. “And for fixed point not only can the resolution be different by layer but also for weights and data.

The precision flexibility can produce significant benefits. Moving from 8bit to 4bit can enable 60 percent more inferences per second while reducing required bandwidth by 46 percent, reducing power by 31 percent while reducing result accuracy by 1 percent.

Imagination provided some benchmarks of performance against GPU. In some areas, such as searching/sorting photos and video analysis, the NN accelerators showed improvements between about 30x and 170x improvement.

When implementing Alexnet AX2185 is 3.5 times the performance of Nvidia’s Maxwell Titan X card.

The cores are being targeted at the 16nm/12nm manufacturing processes although some customers are looking at 45nm and even 55nm, said Socal.

Imagination has two licensees, one in China. Socal said there are no tape-outs as yet but he expected to see products in 1H19. The AX2185 core has already been delivered to lead customers and both cores are available for licensing now.

Related links and articles:

News articles:

Imagination launches flexible neural network IP

Intel to launch “commercial” Nervana NN processor in 2019

ARM launches two machine learning processors

Graphcore’s two-chip ‘Colossus’ close to launch

If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :    eeNews on Google News


Linked Articles