Hacker News Reader

Comment by bsder | Hacker News Reader

Removing Y and the instructions AND taking advantage of the size difference means you have to do new layout.

The NRE to make the change is at least $100,000 (probably closer to $500,000--but I'm spitballing here). And that is up front cash.

Those changes are really small. So, lets assume a penny saved.

It would take you 10 million chips to make that back.

The total run was 25 million? or so--and nobody expected that.

If the NRE was $500K at a penny saved, you never make it back.

Even at a nickel (which is probably a huge savings), again it would take a 10 million chips to make it back.

Given the oddity that this is, I bet that this was a failed batch of chips that was just enough damaged to be really cheap but not so damaged as to be useless.

View on HN

It was a custom 6502 like chip, made by a third party. And they already had to have a custom mask set beyond a 6502 because it integrated a bunch of peripherals. Also, a back of the napkin look at their changes probably means it was half to two thirds the die size, so a pretty good win at those volumes.

Datasheet for the chip they used:

http://www.ic72.com/pdf_file/-/428557.pdf

Looks like a fairly generic "sound processor" that was most likely used in tons of toys and other consumer electronics. Hell, I wouldn't be surprised if the Furby line wasn't even their biggest contract.

> Also, a back of the napkin look at their changes probably means it was half to two thirds the die size, so a pretty good win at those volumes.

There is no chance that removing Y and its instructions saved 50% die area on a chip with 80K ROM and 128 bytes of RAM. The "savings" is likely due to this being in a 350nm or 500nm process rather than anything else.

Seeing as this is a cheap Chinese fabless company and these things were Chip-on-Board, they had a failure on the Y register and just datasheeted it into compliance. This is little different from the original 6502 which datasheeted the lack of Rotate Right into compliance on the first versions.

If it's cheap enough, nobody cares.

Maybe, but removing Y also means removing the instructions which do indirect addressing through it, which removes all the decoding logic for them. So it could very well save a bunch of die space.

Of course one could go look at the layout of the (NMOS) 6502, which has been fully reverse engineered at this point, to prove or disprove this.

Here's a picture with some labels: http://breaknes.com/files/6502/6502.jpg

The Y register is WAY smaller than even I expected. It's a tiny vertical strip on the lower left. And the PLA entries are on top, and a couple entries would again be tiny.

Any savings due to omitting Y would be miniscule.

The more I examine it in detail, the more it becomes clear that it's just a bug that got datasheeted.

There's no reason to assume that chip was used only in the Furby.

It's likely it has been used in all sorts of toys and the like for the last two decades. And it was probably dirt cheap. And probably designed in the early 80s, hence the design decision to go with the 6502 as a base platform.

And probably designed in the early 80s

I thought Sunplus was a bit newer than that (90s), but a quick search reveals that your guess was pretty close:

https://www.sunplus.com/about/milestones.asp

Founded in 1990.

Removing complexity (gates) means yield goes up. My indications from being on the periphery of developing custom silicon and respinning it is that NRE is not $500k for devices that are more advanced (ie multiple sparc cores + dsp + application specific logic)

Reply on news.ycombinator.com