Gregory's ServerProbably the strangest chip that you'll see today: the Intel 2920, a digital signal processor (DSP) from 1979. It was the "first microprocessor capable of translating analog signals into digital data in real time." Chips are usually 16-bit or 32-bit, but this was a 25-bit processor. It didn't have any jump instructions, instead running code in a loop from the 192-word EPROM. Each instruction combined an ALU operation, a shift, and an analog I/O operation. 1/7
 18 comments
 @kenshirriff There's something I never realised I didn't know: how does an EPROM actually work at the component level? (And is it related to the thing about not being able to take flash photos of an uncased Raspberry Pi without it crashing?) @darkling The quick explanation is that an EPROM has a transistor with a"floating" gate for each bit. You program a bit by forcing electrons through the insulating oxide with high voltage, and the electrons are trapped there. Ultraviolet causes ionization in the oxide, releasing the electrons and erasing the chip. @kenshirriff Thanks! So the programming voltage is applied across the gate and drain, and is sufficient to temporarily break down the resistance of the oxide? Then the surplus electrons sit in the depletion layer, making it not depleted any more, and allow current to flow between source and drain? @darkling Yes, there's a diagram at Wikipedia: https://en.wikipedia.org/wiki/EPROM#Operation  @neffo @darkling Yes, although gamma rays will mess up most chips. The Intel 2920 was designed for a wide range of analog applications from filters to speech synthesis to missile guidance. It was created by Ted Hoff (one of the inventors of the microprocessor) and Marshall Townsend. Although the chip was called "Product of the Year" by the magazine Electronic Products, it was buggy and not a market success. 3/7 Interesting chip art on the Intel 2920: a cactus on the die, maybe because Intel opened a site in Arizona that year. 4/7 The 2920 had four analog inputs and eight analog outputs. The digital-to-analog converter (DAC) was unusual and large: a chain of 256 tiny resistors arranged in a grid. Energizing two digital switches (i.e. transistors) provided a tap for the desired voltage. Analog-to-digital conversion was built on top of the DAC, "guessing" one bit at a time. 5/7 The chip didn't have multiplication or division in hardware, but it had a barrel shifter to quickly multiply or divide by a power of 2. The compiler for the chip is very unusual. Instead of normal code, you define poles and zeros and the compiler builds the code to implement the desired filter. 6/7 Thanks to @tubetime for providing the chip.  @kenshirriff "And you try and tell the young people of today that ... they won't believe you."  More than 40 years ago, I used the UV facial tanner lamp of my mom to erase the UV EPROMs I put into my ZX Spectrum. Fun times!  Am I understanding it correctly that it takes analog signals as input, digitizes them, processes them, and then outputs them as analog again? Strange indeed. I've always thought of DSPs as specialized chips that process an already digital signal that may come from an ADC but also from memory for example. @grishka Yes, that's how the chip works. It has 4 analog inputs and 8 analog outputs and the analog/digital conversion is done on the chip. There are no digital inputs or outputs (unless you use an analog I/O as digital).  @kenshirriff reminds me of a MAXQ DSP I read about described as "opcode-less" Every part of the chip would do something every clock cycle, so the "instructions" were nothing but a set of operands for all the units Seems like quite a rare design now -- probably because scheduling, pipelining, and speculative execution have beaten it in total throughput Amazing how silicon design has changed @kenshirriff I remember these! When I was getting my EE degree at Arizona State University back in the mid-80s we had a few development systems for these chips in the lab and I remember reading up on them at the time. They were already obsolete and unused AFAIK but it definitely piqued my curiosity. Sad that they didn't fix the bugs and expand on the concept as it definitely would have been useful. Microchip's dsPIC products came along years later and were very successful in the same space. |
Why is the die photo distorted at the edges? Because I took the photo through the chip's round quartz window. In the olden days before flash, you could store code in an EPROM, memory that could be erased by exposing it to strong ultraviolet light. The quartz window lets the light in. 2/7