|
|
||||
|
|
|
|||
|
|
||||
|
|
Mythos (MemorY To High Operational Speed) is the result of a research started to give answers to some issues raised in the DARPA HyDDENN (Hyper-Dimensional Data Enabled Neural Networks) program. Mythos™ can work with Hyper-Dimensional Data by using less than log2(parameters) for each learning and inference operations. |
|
|
·
Hyper-Dimensional Data Enabled Neural Network |
|
· Running ML on Von Neumann CPU |
|
|
· Statistical Consistency Selective Availability |
|
|
· Constant time for Learning and Inference |
|
|
· Continuous Learning ( Plasticity + Stability ) |
|
|
· Huge Datasets Management |
|
|
· Hardware Adaptive Technology |
|
|
· t = log2 ( parameters ) |
|
|
· No MAC operations |
|
Mythos™ is a technology
designed to obtain SIMD (Single Instruction Multiple Data) parallelism on a
Von Neumann machine, for Pattern Recognition problems. Although the current
processors are "Multi-Core", they cannot be compared to SIMD / MIMD
(Multiple Instructions Multiple Data) processors, especially due to the
limited number of cores, and are still considered as Von Neumann machines.
For this reason there are GPU coprocessors that serve to have massive
parallelism of the SIMD and MIMD type. In many cases it is not possible or
not desirable to use parallel coprocessors: the algorithms of Mythos™ technology are
designed for Pattern Recognition applications with the aim of increasing the
performance on Von Neumann machines by orders of magnitude when a SIMD
parallelism would be required by the application. In particular, the Mythos™ technology is
oriented towards Pattern Recognition applications in the Aerospace sector,
where Radiation-Hardened processors / SBCs such as
the RAD750® produced by BAE
Systems must be used. RAD750® SBC is the workhorse of the space
industry, powering more than 100 satellites that carry out a variety of space
missions. Radiation-Hardened processors must operate at lower cock
frequencies than those typically used by processors operating in sectors that
do not have this criticality. The need for execution speed in algorithms seems
now forgotten due to the widespread use of parallel SIMD and / or MIMD
processors such as high clock frequency GPUs. In
reality, there are sectors, such as aerospace, in which high clock
frequencies cannot be used and given that Radiation-Hardened GPUs have not yet been implemented, the research on
algorithms that allow to speed up Pattern
Recognition problems by orders of magnitude becomes a current need. |
|
|
|
LUCA MARCHESE Aerospace_&_Defence_Machine_Learning_Company VAT:_IT0267070992 Email:_luca.marchese@synaptics.org |
|
Contacts_and_Social_Media
|
|
Copyright_©_2025_LUCA_MARCHESE |
