Hi,

This is a little different to the usual post, but I was wondering if anyone could help, I am trying to figure out a vague time when FPU's became commonplace as built in architecture rather than as a seperate purchaseable installable for CPU's. From what I can see, Intel were the early adopters, but does anyone recall when they were added to things like PowerPC and AMD CPU's?

Also, has anyone seen processors today that don't include an FPU (or more than one) as standard and thus have to have them programmed on/installed as extras?

As a little background, I am trying to begin analysing an algorithm's run time and as part of that want to look at the numerical computations completed, thus drawing a conclusion about how the number of these scale with size and the effect this has, especially when constrasting a FPU vs non-FPU CPU. Was hoping to understand the context of exactly when inclusion became commonplace.

Though another interesting thing would be... how many FPU's does a CPU have these days?

So, if you have read this and have any thoughts, please post them If you have read this and think I don't know what I am talking about, I won't deny it

Thanks!

IIRC, the first Mac based chip to get an FPU was the 68040 back in 1990.
All PPC architecture has an FPU, AFAIK.

Mr frost Mac was not a PPC, and so had no FPU. You could buy extensions that purported to add that in software.

80486 for Intel. The 8086, 80286 and 80386 processors had a companion 80x87 processor for FPU duties, but they were very rare. Starting with the 80486 Intel incorporated the FPU in the main package, although they produced the 80486SX without for a while. From the Pentium on, there has been no version available without an FPU. The '486 came out in 1989.

The Motorola 680x0 range did it about the same time, the 68881 and 68882 were the companion FPUs for the 68K range up until the 68030. The 68040 had the FPU integrated for the first time, those chips where the FPU failed testing were branded "LC" and didn't have an FPU.

PowerPC always came with an FPU AFAIK.

AMD followed Intels lead, but were a couple of years behind. I think they released a 486 clone in 1992, so 3 years behind Intel. They only introduced their 386 clone in 1990.

The 486SX did technically have a FPU albeit a broken one. 486's with a working FPU were sold as DX and those with a faulty FPU had it disabled and were sold as SX.

Yes I remember the separate FPU but it was years ago, like Big D said. Only really at the beginning of personal computing. I think that they dropped the separate chips as it was costly to maintain two processing lines, for the small numbers needing FPU.


Sent from my iPad using Tapatalk.

Thanks for the replies all, the info is really helpful

Interesting that the initial motorola 680x0's and Intel 80486SX's were able to have their FPU break and still be able to operate/be sold. I wonder if the same would still be possible today or if the entire computer would crash and burn.

Also did some digging around on the internet about apple stuff, it seems the Powerbook 500 still had an expansion slot for FPU's, which was in the mid 90's.

Funny how over time the FPU and associated operations has become a measure of performance (FLOPS), especially for supercomputers (according to an Intel doc) whereas it used to be an optional addon.

Not exactly the same, but multi-core chips can have entire cores fail and still be sold. Also true for on-chip cache; if part fails then disable and sell it as a cheaper model.

Some chips are designed with two or more cores but only only one FPU, and many production ARM chips have no FPU. However, since they're rarely pin-compatible I doubt you could sell one with a failed FPU.

Even the original iPhone had an FPU, meaning that integer arithmetic can be slower than floating point for similar results. Such a far cry from the early days. My Z80 would take about 2 seconds to calculate a single tangent!

Thanks for this - I hadn't thought that the architecture of the FPU woulddramatically affect run times, or that the floating point arithmatic could be faster than that for integers. I did some searching about the ARM processors and found out that the latest A9 CPU's have an FPU (a full VFPV3) that is around 10x faster than the A8's (VFPlite). Just need to find the actual exeuction cycles on the A9 one now for comparison, easier said than done

If I remember right, it was mainly spreadsheets that benefited from having a FPU. I never opted for SX chips preferring the future proofing benefits of DX chips, and the price difference was not too bad. Though friends did.


Sent from my iPad using Tapatalk.

It wasn't just obvious mathematical applications. Graphics also benefited enormously. I had otherwise identical 486 SX and DX machines, and comparing just the screen savers was dramatic! The DX managed about 20fps while the SX was about 0.1fps.

Applications like Auto-cad required the co-pro and simply would not install if it wasn't present.

A construction engineering friend had that problem.



Sent from my iPad using Tapatalk.

The technology improved so much, that they could include an FPU on the same slice of silicon as the CPU itself, just like they are now building GPUs and memory controllers onto the silicon as well, because the size of the gates has shrunk enough to gets billions of gates on a slice instead of thousands...

It was more CAD / CAM applications and scientific software at the beginning. Spreadsheets came to it later in the game, ISTR, because it would have meant two versions of the software, one emulating FP and one using hardware. Given that only very few PCs had a FPU co-processor, it wasn't until the late 80s that they started to write it into the software.

CAD/CAM and scientific software, on the other hand, was another matter, that was so expensive and such a niche market, that they could require the FPU to be installed.

Take a look at the ARMv6 chip inside the Raspberry Pi. This has an optional FPU called VFP2, which can be ignored by using armel machine code. This is used by the official Debian distro for the Pi, so it's pretty easy to check out. The floating point alternative is armfp which is used by Raspbian, currently the recommended OS for the Pi.