Hacker News Reader

Comment by jgw | Hacker News Reader

This is reminiscent of the Intel 486 processor and 487 co-processor.

Intel sold 486SX and 486DX parts - the latter had a built-in floating point unit, whereas the former did not. Or at least that's how they were sold. If you had the SX, you could get a 487 "math co-processor" as an add-on to get your floating point unit.

As it turns out, the 486SX, 486DX, and 487 were actually the same chip, with different functionality enabled (although I think the 486SXs were actually ones with defective FP units).

Anyway, if you upgraded your 486SX with the 487, really what happened was that your 486SX got disabled completely, and the 487 became your one and only processor, with a FP unit.

[EDIT: typo]

View on HN

This is pretty common practice to increase yield by using only partially working parts that are still usable in limited way. While 486DX/SX is prime example, there are lots of other:

Original Xbox used marginal DRAM chips and detected by software on each boot which regions of memory are actually usable.

AMD's current CPUs with "weird" number of cores (3, 6, 12..) are produced by disabling cores that does not pass tests.

Cell BE as used in Playstation 3 has one SPU disabled for same reason.

While in most cases such disabling is done by blowing some fuses in hardware I can imagine that in embedded systems software only solution (like the Xbox one) is actually more convenient.

Adittionally, probably the most famous example is the Celeron range.

Celerons were originally Pentium IIs with large segments of the cache fused off, unusable due to defects. However they turned out to be so popular that Intel started to disable the cache of otherwise good cores, and then later began specifically fabbing them, to meet the demand.

For me, the most famous example is DRAM chips in Sinclair ZX Spectrum. Sir Clive got them cheap because they were ones rejected at the production test. Spectrum issue 1 48 K model used 64 kbit chips where one half was broken, and the motherboard was then wired to only use the other half.

That was a time when hardware, particularly memory, was expensive. This was in 1982.

Not exactly. Celerons began as straight up Pentium IIs designed to operate without any L2 cache. They didn't have great performance but were crazy cheap to make and super easy to overclock. Then they started cramming in just a little bit of on-die L2 cache (about 1/4 of the minimum for a regular Pentium II) to help out performance. This turned out to be wildly successful, and eventually this formulation (one chip with on-die L2 cache) became the predominant design for the mainline Intel micro-processors. Afterwards the Celeron family of processors were identical to retail Pentium/Core cpus but with sections of cache disabled, reduced clock rates, etc. (either due to binning by defects or intentional deactivation).

Interestingly, Intel now offers "upgrade cards" for low-cost CPUs whereby components are deactivated via firmware or suchlike and special codes can be entered into the bios screen to turn them on (for a price, of course).

Ah, it seems you're right - thanks for correcting me.

And of course GPU vendors also employ similar techniques. It is not uncommon to see the same chip in multiple models, with lower-end models with parts of the chip disabled either due to defects or for product differentiation.

A more direct and recent example would be the Intel Core Solo, which was an identical situation. New chip, high bin rate while working out the kinks, so they disabled dead cores on Core Duos and sold them as Core Solos.

I had one of these (The SX), it was AWFUL! I couldn't even play games that required the math co-processor. I ended up upgrading the processor to one of those Evergreen ones. Worked like a dream. I wish I can remember which game wouldn't run. I believe it was the original Demolition Derby?

Really? I also had 486SX and the only time I needed the co-processor was when I wanted to run some 3D software (mostly 3D Studio). And for it to work the only thing I needed was coprocessor emulator. In case of games it wouldn't probably be a viable solution as it was obviously slower than the real thing.

Quake just about ran on a 486DX, but failed completely on a 486SX.

Well that's to be expected, 3D graphics usually uses floating-point vector math.

Quake was, IIRC, the first mass market game which used the floating point unit. 486SX owners just couldn't run it. It also suffered more than most games on 486es and non-Intel chips, since the Pentium FPU was streets ahead of any other x86 chip at the time.

http://www.team5150.com/~andrew/carmack/johnc_interview_1997... (search for "float")

I had a 486DX-50 at the time, the fastest chip ever produced in the series with a 1:1 clock ratio, and it couldn't really run Quake. The FPU wasn't designed to be exercised much, probably intended for light-duty use. In practice it behaved like an "embedded co-processor".

A Pentium 75, perhaps the slowest Pentium ever made, was at least ten times faster because of the floating point gains.

Also used the coprocessor emulator for 3D Studio, nearly had forgot how that went till I read this post. Now I work on two quad core workstations and some stuff needs to go to a render farm in the states. What a flashback. :)

Reply on news.ycombinator.com