New High End CPU
Arm Cortex-A77 – faster even without Moore's Law
Fortsetzung des Artikels von Teil 2
Conclusion
Compared to its predecessor Cortex-A76, the Cortex-A77 enabled Arm to increase computing power by more than 20% without compromising energy efficiency. This is all the more remarkable as the figures refer to an identical manufacturing process and an identical clock frequency. Since Moore's law could not help this year and new manufacturing processes will not generate shrinking geometries to the same extent in the future as was originally the case, one will have to prepare oneself to no longer be able to optimize IPC and absolute power consumption simultaneously with every new CPU generation. Co-processors such as GPUs or NPUs for AI applications are therefore becoming increasingly important in order to relieve CPUs of such specific workloads.
With his ML processor, Arm transferred his 2018 energy efficiency assumptions. Instead of 3 TOP/W, 5 TOP/W should now be possible. Today, 85% of all AI workloads are still executed on CPUs or CPUs/GPUs - for lack of an alternative. Simple use cases such as the recognition of keywords (e.g. "Hello, Siri", around 400 MOP/s are required) may still run well on CPUs, but the activation of devices based on facial recognition already requires around 30 GOP/s - the ML processor would thus be working at around 85 % capacity.
The decisive statement is that different applications require different hardware - CPU, GPU and NPU. Arms NN-Framework offers optimal support here - and can even integrate third-party IP from third-party manufacturers.
