Re: F21_vs_Alpha

[Eugen Leitl <ui22204@xxxxxxxxxxxxxxxxxxxxxxx>]

On Thu, 10 Aug 1995, Jeff Fox wrote:

> Dear MISC readers,
> 
> Mike wrote:
> 
> >Yes, maybe I was a little opptimistic here.  I _was_ focusing on integer
> >benchmarks because (1) F21 would be trounced in a floating point 
> >comparison, and (2) it seems that many applications could be written
> >using integer math, based on many comments here and elsewhere.  Of course,
> >most scientists and engineers only know or want to use floats.  I have 
> 
> I agree with you about floating point.  I hope that if F21 can demonstrate
> that it can act as a digital lab instrument it may be useful to some
> scientists and engineers will recognize that it may have some
> usfulness for this.  Something in between a conventional microcontoller
> and a conventional workstation.  Some of the features of a workstation,
> some of the features of a microcontroller.  I have seen some very nice
> digital instrument software that runs on PCs with an A/D board and
> acts like a storage scope or logical analyzer or spectrum analyzer and
> I think that that sort of thing might be quite useful on a single
> F21 or network of F21.

The problem of scaled integer: if their resolution/dynamic range is
too small one has to do rescaling (with shifts) by hand. This is lossy,
messy (you have to be aware how numbers are represented, keep track of
the values, write some primitive numerics opcodes, etc.). Most
scientists have never heard of scaled integers, do not believe you
if you tell them such thing does exist. And they won't take additional
trouble in programming which scaled integers bring.

Moreover, with 21 bit integer math the number of problems you can handle 
is still relatively small (1:10^6 resolution is not very much). 

On the whole this does not bid well for the average scientist 
using the F21.

The problem of storage scope: 6 bit A/D is a bit crude to do
any scoping without an external A/D. 6 bit is good enough to do
so some cruddish voice I/O, but is just too few bits for
most sensoric apps.

> My image of the F21 is something cheap enough to be used in 
> embedded apps, and delivering high end PC power as a single chip.
> As a multiprocessor fairly high numbers of nodes can be connected

I don't think you can get very much beyond 8-16 nodes for most
applications using a single serial ring, however fast. But 16 nodes
is still very good, and are still in reasonable price range.

> before the network becomes the bottleneck in many applications 
> while still staying at low prices.

The need for external SRAMs is unfortunate. The additional board
space, design complexity, sockets, etc. pushes price way up.
Small on-chip SRAM can make the board faster and cheaper without 
external SRAMs, though the CPU price may rise by about one order of 
magnitude.

> >not figured out how the F21 compares to Alpha on integer tasks.  Maybe Alpha
> >would kill it.  I am hoping that a ring of F21s has a chance.
> 
> F21 is not intended to deliver Alpha performance per node. 486DX/4, Pentium,
> P6 maybe, but Alpha is the current top of the line.  The 21164 is

I know you are not trying to do wild claims on F21 for the sake of
seriosity, Jeff, but I think your estimate from the 3d package you
wrote (i486/40, you said?) is too low. Using 3- and even 4-dim matrix
maths with lots of multiplications is not the only way to do fast 3d.
Multiplication is the weak side of F21, and 3d matrixmult makes it
show more than the average code mix would. In fact I am amazed
you still have the i486/40 performance in spite of above handicap.

Apart from clipping with logical masks and sin/cos lookups for
rotations, there exists even a "clean" (i.e. the mathematician will
accept) way of reducing the computation amount for 3d purposes
(some recent German c't, I can look it up, if needed).

> listed as potentially getting two integer and two FP instructions
> per clock, and at up 300 mhz.  Yes, 1200 mips is the peak advertizing
> mips.  Now in discussions in comp.arch where HP and DEC people were
> comparing benchmarks the HP guys liked to mention several published
> Alpha benchmarks that are getting 3 or 4 clocks per instruction, not
> 4 instructions per clock.  So although the Alpha might actually
> get 600 integer mips on some benchmark in the real world it could
> easily drop to 75 or 100 mips.  (yes the thing is 64 bits wide too!)
> Likewise F21's 200+ mips is peak advertizing mips.  In the real world
> the performance will be 50 mips.  It doesn't take much data memory
> access for F21 or out of cache operation on Alpha to see the
> real world numbers drop.

I don't know if you noticed, but _all_ recent chips, P6, the new
Power and certainly Alphas deliver much less performance than
the original claims. And this with benchmarks tailored for this
kind of number-crunching. I take is as a general symptom for deep 
internal (architecture) trouble. In fact, it is highly amusing 
to plot the number of transistors (please add the P6 cache resources, too) 
vs. the real-world (e.g. those benchmarks) peformance. The only thing 
that saves them, is the $/transistor drop trend. Once substrate size and 
structure size starts reaching limits, this trend will saturate/break. 
Each new generation of wafer steppers price grows exponentially. UV won't
suffice for ever, then they'll need xRay synchrotrone rad for litho,
defect yield will go up, and all this won't be cheap.

> 
> Still I like to use Alpha as a comparison of the real difference
> between the extemes.  Alpha is BIG, HOT, and EXPENSIVE.  I have

Have you seen the substrate size? Have you seen the power dissipation
(30 W?), have you seen the package type they are using? Alpha is
a monster.

> collected chip data spreadsheets, and the differences between
> Alpha and F21 just jump out at you.  Watts vs milliwatts, 10M vs

milliwats (and 50% at 3 v) is great for portable machines.
There are very small HDs needing very little power, now (1-2 W).
Mobile computing all day long with cellular modem and a head 
up display (may be with voice input!). Now do this trick
with an i486: 2-3 h at best. People don't like carrying
big battery packs about which run dry too soon.

> 15k transistors, hundreds of times larger and more expensive die
> even when Alpha is made with a smaller process.  Basically it
> looks like 1000/1.

But the Alpha is not $1000, or is it?

> This does not reflect the fact that to get maximum performance
> out of Alpha you need a lot of other very expensive parts like
> very fast memory and GAS cache controllers and three levels of
> cache etc.  As if starting off with 1000/1 ratio were not
> already enough.

Agreed, the assotiated circuitry overhead is much bigger
in Alphas. F21 wins.

> I think the Alpha is a great chip, but I can't afford one.  If
> you want to make a comparison I think you should say how does
> one Alpha compare to 1000 F21?

One of the things of an Alpha (it purported to run better
on a Windows NT): you just don't see the additional power
as compared to a Pentium. Ok, video is noticeably faster.
But all the MIPS are being gobbled up by one heck of a big
OS. If one needs 16 MByte of RAM just to run the OS (SGI 
Irix needs 48 to do anything sensible with it), the cache
is full with OS code, and all you can do is using a C compiler,
the amount of power at the user front end is ... moderate.
Now look at the price, friends and neighbours.
 
One thing where F21 wins house-high and most people keep ignoring:
it is a simple machine which you can learn in a week. With this
knowledge, your degree of control is much higher. Control is power.
Have you seen the audiovisual demos they did/do on the Amiga
hardware? MIPS 0.2-0.6, (7.1 MHz), 16 bit and video access.
Have you seen what they did with this kind of simple hardware?
Remember what they did with the C64? 

A higher degree of control is equivalent to a lot of real MIPS.
Extra MIPS you get from tailored code, what you couldn't do
on a bigger machine (too complex, not enough documentation,
nonlinear tweaking behaviour).

I have never noticed people going for this kind of argument. Why?

> If you are paying $250,000 per node you definitely want the
> biggest fastest killer node you can get.   Alpha is one of those
> killer type CPU.  If you are spending big bucks then $3K or $4K
> for a CPU is a good fit.  You just wouldn't to put a single
> $1 CPU on thousands and thousands of dollars worth of memory.

The CPU price is negligeable in relation to the board, DRAM, SRAM
stuff. As I said, the thing is too cheap.

Apropos, I recently found out that our SP2 is really a 70 node
dedicated workstation cluster (fat nodes and skinny nodes,
RAID, Power CPUs and a fast crossbar for connectivity, AIX).
Seen this way, the SP2 is not even a standard maspar 
supercomputer, but a modified workstation cluster.

> F21 is not designed to run the same software as Alpha, nor is
> it going to produce Alpha level performance per node.  It will be
> closer on performance than it is on any of the other numbers
> in the spreadsheet.  On many problems I expect a few F21 will
> get better performance than a single Alpha.

Xactly. Even two F21 (apart from bus width) have a higher
memory bandwidth than the Alpha.

> It would be better anyway to compare a P64 type processor to
> Alpha.  In .5 or .35 micron a MISC processor could get a clock
> with very high speed, and on stack operations run the CPU at
> about 13 times the memory speed.  I expect that Chuck will offer
> a combination of on-chip ram and rom and multiple busses and

I hope you are talking serial buses here?

> more memory management and multitasking support in the hardware

MMU sounds good. Hardware multitasking (2nd register bank for
OS?) switch, too.

> by the time he is working with 64 bit chips.  But of course it
> is hard to predict what will be a priority in the future.
> The P64 I describe above would be several times more expensive
> than F21, but not several hundred or a thousand times more expensive.

10? 20? $20-$40 would be grand.

-- Eugene

P.S. I hope I haven't hurt anybody's feelings in my previous,
     somewhat wild post(s)? Or why are you so death quiet
     out there? Is MISC alive?

> Jeff Fox
> 
> peak in
> Jeff Fox
> 
>