World of Decline

Analog Extension of Select Existing Digital CPU Math Operations

Posted in Uncategorized by isochroma on April 13, 2009

Analog Extension of Select Existing Digital CPU Math Operations

The idea is to take any 32-bit or 64-bit x86 CPU, and make it capable of something interesting: the capacity to transmit, store, and perform basic or complete math operations on analog values of certain numeric types, or use the normal.

The analog values are transmitted using existing electronic fabrication or more advanced fabrication, such as spin-torque propagation, optical, etc. Existing analog fabrication is likely sufficient.
Analog values are stored in a full-size parallel analog cache of size sufficient to hold the store in the included mathtypes only present on the main cache.

Registers that are capable of storing the affected mathtypes are duplicated in analog.

Digital logic-gates control the use of one circuit or the other. It should again be mentinoed that logic operations and instruction execution are using standard digital circuits.

The chip may best be switched between routing behaviours by Windows custom driver, or better yet system BIOS. Leaving the feature disabled will result in identical output as its commercial brothers. The initial design can be licensed from Intel, or a project proposed to them.

The chip may switch between behaviours by one or more of: a reboot, BIOS change or driver call.

The first advantage of analog computation is it provides solutions to various math processes that can conserve more precision between quantizations (quantized operator {some cases, not all}, conversion to n-bit native formats, etc.) than digital can. The highest advantage is for long-term simulations, where values stream through cache memory for quite some time.

The second advantage to analog computation is it puts the first steps in toward more complex and fuller (and likely new) architectures that make use of more interesting effects in the quantum and nanoscalar realms.

The third advantage is that analog math operations may be more energy and space-efficient than their digital counterparts.

A disadvantage is the limit of CPU cache size. Unless new main memory is implemented, analog values can only exist in the on-die CPU cache. This is a pretty severe limit on the number of storable values. With advanced fabrication, this problem is mitigated. Also, using standard fabrication, a second or duplicate set of the standard DRAM memory plane(s), but implemented to function storing analog values. This shadow cache will hold far more values than the standard cache can, because the digital cache must be split between instruction and data segments, while the analog cache is the same size but completely allocated to analog data.

It must be said that at the final step, the value must be quantized to leave the CPU. Also, this scheme, as mentioned above, only affects certain numeric types during math operations. It does not alter the x86 instruction set, other than to respond to certain currently unused call values. This method is used by the BIOS to set the mode on the fly, and a Windows driver can too.

Finally, the analog alternative offers much lower nominal precision than the digital, since the lowest dynamic range of the worst sender and receiver are the reasonable limit. This limits the range of math operations to perhaps a few thousand, including negative values.

This class of limits may be removed through advanced fabrication and design. There are several appparent techniques to accomplish this end. The first one might think of: using spacially separate or integrated multilayer sensors and merging the results. Each sensor is narrowband with some overlap between their ends. Thus combined, the values of all sensors added together creates a full spectrum in the area of capability.

Another more sophisticated approach stores the number in ‘scientific notation’ using analog equipment. In a yet more ambitious implementation, either of the base or exponent may be digital or analog, or both analog or digital.

The best of various techniques may yield 10,000 but probably no more.

%d bloggers like this: