150 likes | 161 Views
Overcoming computer memory and processor limitations with innovative Ovonic devices. Learn how 3DXpoint technology and GPU advancements are enhancing data transmission and processing speeds significantly. Explore the neural network computing capabilities and effectiveness of Ovonic Cognitive Devices.
E N D
Ovonic Cognitive Computer, LLC formed 9/26/2002 Stan Ovshinsky Helmutt Fritsche Morel Cohen Dave Strand Guy Wicker Boil Pashmakov Pat Klersy Robert Miller
The problem with computers Memory Processor Data Transmission Processor is too slow Memory access is too slow Not enough memory
Memory Heirarchy gives unlimited Memory, but leads to “bottlenecks” Processor Cache DRAM SSD NAND HDD cloud slow 10 nanoseconds 60 nanoseconds 1 millisecond 1 second 10-100 second 128 Megabytes 8 Gigabytes 512 Gigabytes 10 terrabytes infinity Obfuscation of origin 3DXpoint bridges the gap between DRAM and SSD
3DXpoint is Ovonic Threshold and Memory switches connected ECD 1970 3D version ECD 1987
3DXpoint solves the “bottleneck”, speeding up access 10 to 100X But processors are still too slow Memory Processor Data Transmission GPU Data Transmission Graphics processors are now widely employed to speed large, data intensive operations
Graphics Processors are 100X faster than normal microprocessors But they require more data transmission and more memory Nvidia Jetson TX2 GPU $600 Connect 6 video cameras to it and it can drive a car unaided. It can do 1.2 trillion calculations per second
GPUs drive cars, recognize any human language, and analyze huge amounts of data using ~1000X1000X8 Neural Network emulation They emulate the function of a neuron in a network of neurons. It takes an enormous amount of calculation to model each neuron and sequentially calculate the entire network. The Ovonic Cognitive Computer elements have similar function to a neuron. They don’t emulate neural networks, they actually are Neural networks.
Neurosynaptic Cell Ovonic Single and Multiple Cells have the same properties as neurons and biological cells Dendrites Output fires when the threshold is reached by summing the inputs Nucleus Axon Inputs saturate with a sigmoidal response Neural Network computing Uses a nerve-like connected array of elements with saturating, weighted inputs and outputs when a threshold is reached
Ovonic Threshold Device Ovonic Memory Device Ovonic switches can implement a Physical Neural Network • Switching in chalcogenide materials based on lone-pair excitation: • Threshold --- noncrystallizing --- OTS • Memory --- phase change --- OMS The voltage and current characteristics can be tailored for the requirements of the application This is the only practical device that works. RRAMs, Memristors and quantum computing devices are being researched. Ovonic devices are now commercial products.
DEVICE RESISTANCE (Ohms) PROGRAMMING VOLTAGE (V) Ovonic Memory Multi-State Data Storage For Storage: Multiple bits per cell increases storage density and reduces cost For Processing: The analog characteristic provides ideal means for synaptic weighting
Ovonic Multi-Terminal Threshold Device This simplifies control in a synaptic environment
Well over 1000 startups are pursuing cognition with >$20B invested
September 10 - 16, 2017 Aachen, Germany 2016 2011