Intel to use new way of numbering [SFGATE... gigahertz isn't everything]

[truthisfreedom]

http://www.sfgate.com/cgi-bin/article.cgi?f=/c/a/2004/03/20/BUG4E5O7VK1.DTL

Intel to use new way of numbering
Model designations to focus on features other than speed

Matthew Yi, Chronicle Staff Writer

Saturday, March 20, 2004

Intel Corp., which has touted the need for microprocessor speed, on Friday said that gigahertz isn't everything in PCs and that it plans to roll out a new model numbering system for its Pentium chips this summer that focuses on other features.

The new model numbers -- 300, 500 and 700 series -- will focus on chip features, such as improved security, less power consumption and how fast data can move in and out of the microprocessor, said Don MacDonald, vice president of Intel's corporate marketing.

<snip>

For example, Intel makes two versions of its 3.4 GHz Pentium 4 chips. The extreme edition, which is designed for hard-core gamers, has larger memory as well as a bigger pipeline to get data in and out of the microprocessor.

For desktop PC chips, the 300 series will represent Intel's Celerons while the 500 series represent Pentium 4s. The Pentium 4 Extreme Edition will fall under the 700 category. For notebook computers, 300, 500 and 700 will represent Celeron Ms, Pentium 4s and Centrino's Pentium Ms, respectively.

Despite the new plan, Intel is not giving up on clock speed. It plans to list the frequency as one of its features.

<more>

[punpirate]

1. Hmmmm...

... this probably means that Intel has hit the wall regarding speed increases, at least for now.

[Warren Stuart]

2. A chain is only as strong as it's weakest link

What's the point of developing super fast chips if the computer can't handle the speed? The bus speed on the motherboard and chipset, as well as the memory speed are limiting factors.

It would make sense to work from the bottom up, develop the weaker (or slower) parts and this will bring the overall performance up.

But you may be right, eventually they will reach a limit with current technology.

[Yemp4734]

6. Actually many times the bus is way overboard

For example, take AGP cards. Go from AGP 2x to AGP 8x. You might see a frame or two increase/decrease when you benchmark your video card. The card never even exceeds the bandwidth of AGP 2x... let alone AGP 8x. Great for marketing though.

[democracy eh]

3. its all about marketing

Intel's next generation chips will be faster, but will run at lower clock speeds. Most casual computer users believe that Mhz = speed. After years of using Mhz as the be all and end all, Intel is in a pickle as to how to sell consumers a "slower" chip.

AMD and Apple have been trying to shatter the magahertz myth for years.

[LifeDuringWartime]

4. its all about the pipeline cycles

i think that's what they're called. also, one of the reasons the G5 is so damn fast is because of its 1GHz bus. if it weren't for the advanced cooling design, you could make a grilled cheese on the case.

[democracy eh]

5. mmmm grilled cheese on a G5

maybe that is how I could justify buying a G5! I wouldn't need a toaster oven.

[priller]

8. The amazing thing about my G5 is how QUIET it is

It's pretty amazing that having 9 fans would make it quieter.

[Atlant]

7. Yes, pipeline cycles and stages

Modern computers are, essentially, assembly lines for your data.
Just as in a car assembly line, each piece of your data passes
through many stages in what computer scientists and engineers
call "the pipeline", and at each stage in the assembly line/pipeline,
some small primitive operation is done to your data. Because of
this, a single computer instruction may take several (or even
many!) clock cycles from start to finish, but the computer still
runs fast as the blazes because it can have many instructions
"in flight" (in the pipeline) simultaneously. This is just like that
car assembly line where a given car may take several hours to
build (from start to finish) but at the end of the assembly line,
you'll see a new car roll of the line every minute or so.

In addition, modern processor chips have multiple pipelines
(multiple assembly lines) so they can actually retire (complete)
multiple instructions with each clock cycle. (For you computer
types, it's not unusual to be able to retire two integer operations,
two floating-point operations, a load/store, and a branch
instruction in each clock cycle.)

But all this "parallelism" (simultaneous, coordinated operation)
doesn't come cheap. Because of that, these vastly complicated
processors may run each individual clock cycle more slowly
than a simple processor would. But because they have far
more functional units simultaneously processing, they get
more work done per unit of time.

It's because of this that MHz have *NEVER* been a fair
comparison. If you compare two processors, one that
runs at 3 GHz but only retires (completes) one instruction
per clock cycle with a second processor that only runs at
2GHz but retires up to five instructions per clock cycle,
the 3GHz processor "sounds" faster, but in fact may only
be accomplishing 30% of the work that the 2GHz processor
is accomplishing in any given unit of time.

Intel has lived by the MHz myth, pumping out processors
that clock ever faster, but this is starting to catch up to
them. Many apparently "slower" processors are actually
distinctly faster performers when you measure how much
work they can actually do.

Atlant
(Old-time processor designer)

[priller]

9. And if you have to stop that beautiful parallelism

If you hit a mispredicted branch or something like that, and have to restart the pipeline, then your overall effective rate really goes downhill.

[Atlant]

11. Which, of course, is why...

> If you hit a mispredicted branch or something like that, and have to restart the
> pipeline, then your overall effective rate really goes downhill.

Which, of course, is why modern architectures spend a lot of effort
to *NOT* branch at all and, if they have to branch, to try *VERY*
hard not to mis-predict the branch.

Branch-minimization techniques usually include some sort of
conditional execution where the processor can test some
condition, then use the results of that test as a "predicate"
to allow or suppress the execution of subsequent instructions
Alpha allowed one predicate; IIRC, IA-64 allows five predicates.
This allows algorithm such as the max () function to not branch;
it simply tests A>B, sets max=A, and suppresses the setting of
max=B if A>B. No branch needed, just one "predicated" instruction
whose execution may then be suppressed; the pipeline roars on
at full speed.

Loop unrolling is another powerful technique. If we perform, say,
five passes of the loop before finally looping back, we've then
reduced by a factor of five the number of branches-back that
must be performed by the processor.

Meanwhile, branch-prediction takes place on several levels. Some
architectures simply assume that rearward-facing branches (as in
loops) are usually taken and forward-facing branches are usually
*NOT* taken. Other architectures allow an explicit "branch prediction"
bit to be set by the compiler; this bit indicates whether or not the
compiler thinks the branch will usually be taken. And other archi-
tectures keep track of the last action at the branch and assume
that the next time the branch is hit, it will go the same way.

(By the way, it's not unheard of for a processor to begin processing
down *BOTH* directions of the branch if all of the necessary instructions
are in I-cache, at least to the point where a "side effect" will modify
state that would be externally visible.)

But your point is still entirely true: if a branch *IS* mispredicted,
then the costs in modern processors to discard the in-process
work and restart the pipeline can be *VERY* high. Which is why
so much effort is spent on the approaches that I've described
above.

Atlant

[Atlant]

12. IBM Journal of Research and Development...

It's not available tonight, but the IBM JR&D had a wonderful series
of articles about their Power4 architecture (the basis for the current
PowerPC-970 chips that are used in the Mac G5s); these articles
discuss many of the trade-offs that we're discussing.

When the web site is back up, this link should get you started:

http://www.research.ibm.com/journal/rd46-1.html

Here's a link that's working tonight:

http://www-1.ibm.com/servers/eserver/pseries/hardware/whitepapers/power4.html


Meanwhile, Intel has plenty of articles about the IA-64 architecture
as well. It's not clear yet whether the IA-64 will make a significant
impact in the marketplace, but it's certainly an interesting architec-
ture to read about.

And then there's the poor lamented DEC Alpha...

Atlant

[Florida_Geek]

10. Funny AMD has been doing this for some time now

:)