You posted about the new 68050-Fpu.
How does it compare to the 68060-Fpu?
Does it include all instructions of the 68060?
What precision does the 68050 FPu support?

Thanks in advance

Both are similar.
First of all the 68060 FPU is done.
The 68050 is nearly done.
It will certainly take a few more weeks for it to get 100% finished.

Angel of Paradise wrote:

Does it include all instructions of the 68060?

Yes.
Actually we plan to include even instructions that the 68060 had only in software.

Angel of Paradise wrote:

What precision does the 68050 FPu support?

Currently the FPU does support full 80Bit precision.
We consider to offer a "light" version which has 64bit precision.
Today its common to have 64bit FPU precision.
PowerPC and many other systems have 'only' 64bit.

This means for "normal" software this precision is fully enough.
The advantage of the 64bit precision would be saving in chip real estate - which is at least interesting.

I wonder what the experienced FPu coder think about this.
Feedback?

        
                   68882          68050   Latency
FMOVE  mem,reg       40              1         
FADD                 75              1    8
FSUB                 75              1    8
FMUL                 95              1    8

Because of different sequential dependancies in different code the
the FPU performance ratio from a 68050 to a 68882 will vary.

With full sequential code the 68050 is only about 10 times faster than a 68030 with 68882.

This means the 100Mhz 68050 reaches then only a performance comparable to an 68882 clocked at 1 GigaHerz.

For code with little sequential dependencies the performance ratio will go much higher.

The peak performance of the 68050 would be  equal to an 68882 running at  13 GigaHerz .

As the 68050 can fuse a FMOCE and FMUL in a single cycle, at can do this combination in 1 cycle while these two instructions would need 135 cycles on the 68882.

Interesting?

Great!! Simply great!! :)

                     68040 Latency   68050   Latency      80486
   FMOVE  mem,reg        3              1                     3
   FADD                  3 7            1    8             8-20   
   FSUB                  3 7            1    8             8-20
   FMUL                  5 9            1    8               14
  

A proper testcase like the Mandelbrot from SP or a Matrixmul testcase will IMHO make most sense.
This will give realistic numbers.
 
 
  Cheers

It means that we can have a VBL Quake I on the NATAMI (software rendering or Gl rendering, GPL code EXTERNAL LINK ).
Quake I was the first game to use massively the x86 FPU unit for software rendering.
Every gamer that had a Cyrix 166 change to the Pentium 133, because the FPU was two times faster.

I'm not found of quake, but it would be a good mixed benchmark (integer/memory/fpu) to see what Natami has in the guts compare to a 96's PC without 3D card (Quake II has a software renderer also).

Amiga port
http://planetquake.gamespy.com/View.php?view=Quake.Detail&id=326#Files
ClickBoom amiga commercial port
www.lemonamiga.com: EXTERNAL LINK 

video on youtube : EXTERNAL LINK  Amigas equiped with full 68060@50, fast gfx card and AHI compatible soundcard would have seen a good 8-10 fps. My own setup AGA +68040/25 would get around 4 fps with postage stamp size window and 2x2 pixel mode ;) Game uses CD Audio for music (Not Recorded).

But its to big a testcase to learn anything from it.
With such a big testcase you will get a "score" but you will have no clue why your system scored this score.

A smaller testcase allows you to analyse both the code and the CPU behaviour reaction on it. And only by analyzing you can learn from it and improve our evolve your CPU.

This means for me as CPU developer such a big testcase is of little value.

My point was that "Quake I" is FPU bounded for the software renderer, it will not help to design the FPU, but will help to have an appreciation of its relative efficiency in an FPU bound application.