6

For both the x86 and SPARC based machines, it was possible to get not only FPUs installed as a separate, discrete IC, but also third party FPUs.

In particular, I've read about Weitek on Wikipedia.

Question:

Did Weitek produce the fastest FPUs available for the x86 and SPARC during the era when it was still possible to replace the FPU as a discrete IC? Specific examples would be nice: e.g. what is the absolute fastest FPU available for the 286? 386?

Bonus:

Is there anything else to this story of third party FPUs that is worthwhile to mention? What other companies beyond Weitek did this?

Stephen Kitt
  • 121,835
  • 17
  • 505
  • 462
AlphaCentauri
  • 1,533
  • 1
  • 11
  • 22
  • Related: https://retrocomputing.stackexchange.com/questions/2421/were-there-8086-coprocessors-other-than-the-8087 – snips-n-snails Feb 14 '18 at 05:21
  • IIRC I got a Russian x87 in Keramic casing with glass window buried somewhere under serious layer of dust but hard to say if it was 1:1 clone or just compatible with different performance... so yes there where other "companies" that produced FPUs but hard to say if it can be considered 3th party ... and assuming performance without measuring is non sense There where old benchmarks (even under MS-DOS like SYSINFO or something like that) may be you can dig some dbs ... – Spektre Feb 14 '18 at 09:12

1 Answers1

11

I don’t recall SPARC systems having separate sockets for discrete FPUs; in particular, the Weitek SPARC POWER µP was a replacement SPARC CPU which derived much of its speed benefit from doubling its internal clock.

On the x86 FPU side of things, quite a few different companies produced FPUs: Intel of course, AMD, IIT, Cyrix, Chips & Technologies... Many of them were marketed as being faster than Intel’s 80287 and 80387 designs; Cyrix’s FPUs even had that in their name — FastMath 83D87! BYTE volume 15, issue 12 (November 1990) has a detailed FPU comparison, with benchmarks and explanations of the differences between the FPUs. Most FPUs used the same interface as the Intel FPUs, with a direct CPU-FPU connection and special opcodes; the Weitek FPUs were the only exception, using a memory-mapped interface. The IIT and Cyrix FPUs were supposedly faster as-is, but they both added some features beyond speed: IIT FPUs had more registers, Cyrix FPUs used more accurate transcendental calculations. Weitek FPUs used a completely different programming model, so they required specific support from applications, and lacked support for 80-bit extended-precision values; they also required specific motherboard support (and it was possible to build systems with both a 387 and a 3167, or a 486 and a 4167).

As far as benchmarks go, the results from the BYTE issue above match what other benchmarks found: the fastest 286-class FPU was the IIT 2C87 (but the benchmark doesn’t cover Weitek’s 1067), and the fastest 386-class FPU was the Weitek 3167, followed by the Cyrix 83D87. Weitek’s 4167 was faster than the 486’s built-in FPU in some operations but not all.

BYTE volume 13 issue 3 (March 1988) contains a number of articles on x86-compatible FPUs, including the Weitek 1167, and explains how to program them.

Stephen Kitt
  • 121,835
  • 17
  • 505
  • 462
  • 1
    For an overview (including performance data) of all 80387 competitors, one might also want to have a look at my 1993 write-up "Everything you always wanted to know about math coprocessors", a nicely formatted copy of which one can find here – njuffa Apr 01 '20 at 23:06