Neural Network Computing: BCM TOD? Value Vectors

by Rogier F. van Vlissingen, 2022, CC BY 4.0. BCM Co-founder & Advisor

The BCM TOD? family of neural networking computers is set to revolutionize the face of computing along with a wave of Neuron Emulating Processors from the majors, IBM, Intel, Samsung, et al., using solid-state materials. BCM calls these NEPs, Neuron Emulating Processors, to distinguish them from live neuron processors. Neural Computing technology will begin hitting the market in the next two years. As a result, the promise of neural network computing is finally set to become a reality. With it comes a three order of magnitude increase in processing speeds, and massively parallel processing capability.

The BCM TOD? - Tissue Operating Device – is based on actual live neurons. Natural neurons have innate capabilities of learning and intelligence that allow them to operate with a much-reduced instruction set, compared to processors that emulate neurons. Natural neurons need a life support system, and initial versions will not be miniaturized for portable devices, but in the data center they have all the advantage, because they don’t need line-by-line coding instructions, but instead can execute higher-level commands.

1. Processing Performance

Processing improvements with neural processing have been widely reported as three orders of magnitude, or 1000x, and that in and of itself is hard to grasp in terms of its practicality, and still not a very meaningful way of assessing the performance issue. It is one thing to say that a 1000 second (16.7 minute) task would now take one second. The whole paradigm is different, and, indeed raw speed is a feature, because within our computers, the core processing units and all the memory are within a live neuron and linked neural networks within a cluster of a million neurons, (TC Disk). That means they are both processors and memory chips combined, a no-latency network among those processors and a separate memory device, because the memory devices is the cell’s DNA, all in one neuron.

One dimension of the performance advantage can be grasped easily: in a digital computer the CPU and memory are different components that all need to communicate via wiring, which introduces constraints whereas with the live neuron-based system all of that happens in the same space with internal connections.

TOD? uses live neurons, which nature has designed differently. Live neurons perform processing, hold and access memory all within the core of the neuron. In addition to neurons having lightning-fast processing speed and massive data storage capacities, this natural all-in-one core design eliminates the time and issues related to digital computer memory calls and data transfers.

Although functional and effective live tissue processing has been performed in research facilities and laboratories since 1990, for over 30 years, the development and volume manufacturing of an effective commercial tissue computer, using live neurons is now being realized by BCM for the first time.

The actual performance is a function of both the raw speed and low energy within such a system, but just as much of the massively parallel structure provided by millions of neural networks that renders these live neuron machines extremely efficient when it comes to handing big data tasks and chaos processing applications.

2. Energy Performance

The extreme energy efficiency of such a system is also hard to grasp. The reduction in energy consumption is in the range of 3-5 orders of magnitude, which therefore makes energy a non-issue. One would need to have some experience with the difficulty of siting data centers to understand the importance of this. Even from published reports, it is evident that energy and even more so green energy is a huge issue in the data center space, for the industry has a bad rep as a prolific polluter.

At the low end, the PC tower next to my desk houses a system with Intel Core i-7 6 core, 64 GB RAM; Radeon RX480 8GB; SSD 500 GB; HD 2x 2TB/8GB cache; CD/DVD RW, Blu-ray/DVD/CD burner; Card reader. It also sports an 850 Watt power supply to drive all that. I don't know what it adds to my A/C load, but it is certainly a non-zero amount. In the same space a base-model BCM TOD? with 16 million neurons, might have the processing power equivalent to 700-1,000 similar machines and consume about 75 Watts and not be noticeable in the thermal load.

The two concepts that can perhaps best illustrate how this is possible, is the realization that the human brain consists of around 80-85 billion neurons by most estimates, and it uses about 10 Watts. And the other is to realize that in a silicon chip, which houses millions of transistors, they all need to be powered at all times in order to be able to work. The synapses in a neuron, however, need power only for one brief moment, in order to fire.

The upshot is that energy is no longer a limiting factor in terms of data center locations. The location of the customers and the presence of fiber become the operative concerns.

3. Real Estate Performance

The issue of floor space is perhaps the easiest to understand. The physical units are simply more compact, and just as important, fundamentally do not generate heat. Operating conditions are in the human range: 50 – 85 degrees Fahrenheit. By virtue of those requirements, there is also no heavy-duty HVAC involved, nor is there a need for a lot of redundant power, as it would be easy to provide even a 30-day battery backup. Solar power might be adequate.

In short, a TOD? populated data center might get more computing service performance output in 5,000 or 10,000 sq. ft. room than a traditional data center produces in a huge 100,000 sq. ft. data center facility. This means that a data center could be located in any office building as long as sufficient fiber capacity is available.

Processing Performance Considerations

Neural Network Computing

In and of itself, neural network computing is perhaps a sub-class of high-performance computing (HPC), this being the class of application areas where intensive computing is required, so that the speed of the processing is as important or more important than the speed of I/O. It is the terrain of supercomputers and HPC where TOD? resides. Examples of applications that might benefit from neural network computing directly are e.g. scientific computing, big data applications, various financial trading applications, meteorology, and complex modeling, in general.

General Computing

All computing will of course benefit from the speed gains of TOD?, but HPC is first and foremost, and from that standpoint using TOD? for general computing might not seem the most natural. However, the two other factors, the space economy, and the low power consumption could still drive the selection of the TOD?, and the higher the energy and real estate costs involved, the more likely the preference for TOD?. Even a small machine could consolidate the functions of a large number of legacy servers into a single cabinet with extremely low power consumption. Evidently, the higher power and costs, the more this issue matters. Some data centers might pay $0.10/kWh and others $0.20, or more. This will drive the equation.

Sensor Input-Based Analog Computing

There will be a special version of TOD? for processing sensor inputs. The notable feature of using live neurons in this regard is that they have the capacity of processing analog images, and thereby avoid the loss of detail that results from the digitization of analog inputs. Applications for this are in medical imaging, and seismic exploration, including for oil and gas drilling and any number of other scientific applications. Among other things the company means to have a field deployable, movable unit, perhaps the size of a 20’ container, which could include a 30-day battery backup and if need be, could be installed with a solar panel to be in effect permanently independent from the grid, using a satellite uplink where there is no other internet connection.

Data Center Operations

Another area of application that seems counter-intuitive at first, but which arrives as soon as we can construct a network of TOD? machines, is that we could connect the TOD? machines to legacy equipment via direct fiber in a data center, and certain tasks could be offloaded, such as backups – an encrypted, redundant, and geographically dispersed backup would be easy to implement. Likewise, authentications done via a TOD? backbone through the BCM sister company MCI, could be far faster than is now the case. One and the same TOD? machine could offer high value “Access TOD?” services, and simultaneously provide these mundane background tasks.

Conclusion

TOD? is a paradigm shift in computing in multiple dimensions and therefore it will drive a diversification in the computing space. The speed of adoption will always be driven by monetizing those advantages and therefore we fully expect that HPC will primarily drive the bus, with areas such as financial trading in the lead. Mineral exploration is another area and in general anywhere super computers are now in use. However, as should be obvious, depending on the location, other considerations might also favorably impact the economic value of TOD?, and all of them are equally valid. In the data center space, we foresee a mixed environment as the most likely scenario for a long time to come.