SPECTRE

By Lars Wood | November 26, 2018
Lars Wood is SPECTRE architect, pioneer in AI And High Performance Computation


SPECTRE is a scalable dynamically reconfigurable analog recurrent neural network super Turing processor.

1) SPECTRE: GaAs/GaN 3 THz un-clocked 4D Multi-Chip Module on micro-cooled diamond dielectric.

2) Full custom VLSI, extremely high frequency Schottky Diode/HEMT semiconductor process.

3) SPECTRE modules temperature and radiation hardened, using only 1 Joule of energy per module.

4) SPECTRE module computational throughput for each module die is 3 Quadrillion operations p/sec.

5) SPECTRE module neuron wireless interconnect enabled by modification to the physical gate structure transforming each analog neuron into a sub millimeter wave transceiver for high neuron fan in and fan out supporting multi-chip neuron dynamic interconnect. This enables dynamic high performance reconfiguration of SPECTRE recurrent neural networks, which provides adaptation to changing scenarios and resulting in melded real-time learning.

SPECTRE MODULE COMPUTATIONAL SUMMARY:

The SPECTRE module integrates the high precision of current digital computations with the ultra high performance of analog neural net computations.

SPECTRE MODULE APPLICATION DEVELOPMENT:

SPECTRE solves hard computational challenges by taking any digital computation and breaking it into two parts. One part remains in the digital realm and the other part in the SPECTRE analog recurrent neural network realm. Our approach is that SPECTRE acts as a approximate computational oracle to the digital computation, which reduces by many orders of magnitude the scope of the computational problem that the digital calculation is required to solve. This process involves taking any current digital calculation and translating it into equivalent ordinary differential equations (ODEs) and a complementary Finite State Machine (FSM), which implements the simplified digital computation (called the SHADOW in our architecture). The resulting FSM and ODEs are simulated within a software technical computing environment. Once validated, the ODEs are next translated into a conventional analog circuit description, which is simulated within EDA development environment. Once the analog circuit is confirmed by simulation, it is translated into its recurrent analog neural network Schottky Diode/HEMT equivalent circuits, (equivalence proved in Siegelmann-Sontag thesis).

SPECTRE "SLUICE BOX" CRYPTANALYSIS APPLICATION:

Sluice Box is a SPECTRE cryptocurrency application. In Bitcoin cryptocurrency transaction validation “mining” , the "nonce" in a bitcoin block is a 32-bit (4-byte) field whose value is adjusted by miners so that the hash of the block will be less than or equal to the current target of the network. Current Bitcoin cryptocurrency CMOS ASIC miners generate mining revenues of about $3-$6 per day per miner. Our estimates indicate each Sluice Box will generate $150K per day per Sluice Box 1U PCB. Our plan is that the Sluice Box rack mountable 1U PCB will contain 64 SPECTRE application specific modules. SPECTRE modules operates at extremely low temperatures because it is implemented using a direct band gap, electron high velocity semi-conductor GaAs/GaN process. This is in contrast to current indirect band gap, low electron velocity Silicon processes (CMOS FINFET for example), which operate at very high temperatures and electrical power energy levels. Consequently, Sluice Box is a the first green cryptocurrency Proof of Work technology.

BLOCKCHAIN COMPUTATIONAL METAPHOR:

The blockchain nonce calculation is a SPECTRE The blockchain nonce calculation is a perfect metaphor for solving any cryptanalysis challenge and extends to solving many hard computational challenges in telecommunications, biotechnology, data mining, UAVs and robotics. An recent example application proposed by one of my science advisors is refinement of CRISPR gene editing technology using SPECTRE driven quantum chemistry. CRISPR suffers from inaccuracy due to lack of precision.

DEPLOYMENT

SPECTRE deployment reaches into all systems and applications, transforming computation. The logo is based on the hybrid nature of the SPECTRE module, which obtains the precision of digital and the super Turing performance of analog recurrent neural network computation. SPECTRE transcends the Siegelmann Barrier, which limits digital Artificial Intelligence to narrowly defined applications, which is why Rethink Robotics, (funded to $70 million dollars), attempt to develop adaptive robots with digital technology collapsed and failed.

SPECTRE development is in Paris France. Lars Wood is the SPECTRE program director.