Hi Chris,
Obviously you know that watched pots never boil, so now we know why we're still waiting.
Much as we might wish otherwise, Helium is definitely not Akida.
blending-dsp-and-ml-features-into-a-low-power-general-purpose-processor.pdf (windows.net)
https://armkeil.blob.core.windows.n...nto-a-low-power-general-purpose-processor.pdf
As to AI, it seems that helium is like the curate's egg.
***
A
nother area of interest is the ML processing performance. Several machine learning algorithms have been ported to Helium, including a keyword spotting library and a CiFAR10 image classification library. From this result, we see that the Cortex-M55 processor gives nearly ten times better performance than the Cortex-M4 processor in keyword spotting, and almost six times better than the Cortex-M4 processor in CiFAR10 image classification operations.
***
T
o keep the Cortex-M55 processor energy-efficient and fit within the power budget for the majority of IoT endpoint systems, the Cortex-M55 processor internal datapath for its vector extension is 64-bit, which means it takes two clock cycles to operate on a 128-bit vector. However, the architecture behind Helium allows a processor’s implementation to overlap execution cycles to enhance performance, providing that there are no hardware resource conflicts.
***
W
hile the Cortex-M55 processor design enables significant performance uplift in signal processing and machine learning applications, not every application can gain the same level of performance boost. Since Helium technology is based on SIMD operations, it works very well when the data processing can be vectorized. However, there is a range of application codes that cannot be vectorized. The traditional VLIW approach, however, allows different operations to be scheduled at different execution slots. This potentially allows some very sequential code parts handling to be carried out quicker (e.g. variable length encoding/ decoding in audio codecs). For Arm processors, it is also possible to achieve similar parallelism by introducing superscalar in the design. The Cortex-M55 processor, however, is not a superscalar processor, and therefore, does not have this capability. Nevertheless, with limited dual-issue capability in the Cortex-M55 processor and various new features in Armv8.1-M architecture (e.g. low-overhead loops, new conditional execution instructions, 64-bit shifts), scalar performance has been improved in various areas.
Basically, Helium expands the instruction word from 32 bits to 128 bits, but, because they only have a 64 wire bus, it takes 2 clock operations to implement the instruction word on the helium SMID (single instruction multiple data) architecture.
Helium can run CNN software instructions on the voice and image models, and can perform ML. It includes hardware modifications to improve its AI capabilities, but is software driven. Helium does not employ SNN.
Because helium is software driven it would be orders of magnitude slower than Akida and would use much more electricity.
ARM do not make chips. They license IP.
None of the ARM processor IPs include Akida.
Akida is offered as an IP option which can be added to any ARM processor IP.
The block diagram includes a Coprocessor Interface. That's where the Akida IP would connect into the ARM IP.
View attachment 50443
I Listened to it 3 times, thanks Chris.
lucky I have plenty of patience and unwavering trust in Management along with the belief that it's a matter of when... no ifs
thestockexchange.com.au
