Archived from groups: comp.sys.ibm.pc.hardware.chips (
More info?)
On Tue, 29 Jun 2004 03:57:11 GMT, Robert Myers
<rmyers1400@comcast.net> wrote:
>
>Tony Hill wrote:
>
>> On Fri, 25 Jun 2004 17:48:50 GMT, Robert Myers
>> <rmyers1400@comcast.net> wrote:
>>
>>>The method used to estimate the performance improvement of Celeron D
>>>over Celeron--running at an unrealistically slow 2GHz--systematically
>>>underestimates the performance improvement to be expected from running
>>>at more like 2.6GHz with larger cache. The direct comparison chart
>>>shows a performance improvement of 11.7% for Business Winstone 2004; I
>>>get an estimated improvement of over 24% by pulling numbers off the next
>>>chart, where the processors are run at full speed.
>>
>>
>> Much more important that the clock speed is the bus speed. In the
>> 2.0GHz comparison both Celeron chips were run with a 400MT/s bus
>> speed. The idea behind this was clearly to test the differences that
>> the cache and core make, which is interesting in it's own right, but
>> not entirely applicable to determining what chip to purchase.
>> Comparing the old Celeron 2.6GHz to the new Celeron 330 you are
>> looking at the effects of a new core, doubles L1 cache, doubled L2
>> cache and a higher bus speed all combined together.
>>
>
>Double, double toil and trouble...mutter, grumble...details of memory
>operation and latency...
>
>I guess I just don't really understand the point of the 2.0GHz test.
It's admittedly not the most useful test in the real world. I believe
the main idea was to contrast this to the Northwood vs. Prescott P4
comparisons, where the latter was usually somewhat slower given
identical clock speed and bus speed. With these Celerons that trend
is noticeably reversed, the Prescott-esk Celeron D is almost always
faster (often by a margin of over 10%) than the Northwood derived
Celeron of old.
>When the processor runs at the higher FSB speed, the memory latency
>doesn't change because it's still the same memory running at the same
>conditions (2 x 256MB DDR400 @ 2:3:3:6).
To a rough approximation, yes.
>With fixed memory latency, processor frequency, memory bandwidth
>requirements, and cache size requirements all scale together. A one
>third reduction in processor frequency is like, all other things being
>equal, a one third reduction in required cache size and memory bandwidth
>for equal performance. Things don't scale quite that neatly, but such a
>simple-minded theory goes a long way toward explaining 11.7% gain vs.
>24% gain for doubling the cache size, with the processor frequency being
>reduced by one third for the lower of the two estimates.
Err, I think that might be a bit of an oversimplification of things
here!
>There's little point in having extra cache if you don't have the
>bandwidth to get the data in...but there's very little point in having
>the extra bandwidth if there's nowhere to put the data. The more
>meaningful comparison is the 2.66GHz Celeron D against the 2.6GHz
>Celeron. The discrepancy between 2.66GHz and 2.6GHz can be fixed with a
>little systems engineer's body english.
What's perhaps most applicable would be a dollar for dollar
comparisons, ie probably a Celeron D at 2.53GHz vs. a Celeron at 2.6
or 2.7GHz for the time being. After all, the real reason for looking
at the Celeron in the first place is low cost.
>What this little diversion highlights is that, as Intel pushed the clock
>on Celeron from 1.7GHz, the cache situation, not very good to begin
>with, got more and more dire, and the Celeron D is an inevitable course
>correction.
Combine that with somewhat restricted bandwidth and the whole memory
subsystem of the old Northwood-style Celeron pretty much blew. The
increased bus speed and doubled cache are rather obvious solutions
and, not surprisingly, they result in a fairly noticeable increase in
performance.
What I found funny was that some people were surprised by this fact!
>> However as you mention, most OEMs systems are limited to an Intel-only
>> platform for whatever reasons (95%+ marketing), these new Celerons
>> make some sense. HP and Dell get these chips for DIRT-CHEAP. I would
>> hazard a guess that they pay somewhere on the order of $30-$40 for
>> these Celeron processors, so really there is no real price advantage
>> for going for an AMD chip. You end up with somewhat lower
>> performance, but it's close enough that no one will notice.
>
>I mentioned the actual economic realities because one could come away
>from a discussion like this wondering how it is that Intel ever sells
>processors. Some buyers, it is true, are gullible shoppers at a place
>like CompUSA, but not all of them are by any means. Intel hardware
>commands a modest brand premium, but not as much for buyers of OEM
>hardware as for home-builders.
Definitely not. I've had the chance to get some exposure to the inner
workings of computer building and whatnot at large OEMs recently, and
basically the hardware itself seems to be a pretty much non-issue
cost-wise (the one exception to this rule seems to be motherboards...
perhaps not by coincidence since there is TONS of competition in the
retail market pushing prices down). The cost all seems to come from
support and all the extras like management, marketing and logistics.
As a interesting side note, I now finally understand Dell's thinking
behind their "white box" systems. I'm sure that they are able to
churn out the hardware for next to nothing as long as they have zero
support costs associated with it. Offload the support to some other
company (the retailer in this case) and Dell's per-unit cost are tiny.
I'm still not sure that the idea will actually translate into a
commercially viable product, but at least I understand where they're
coming from now.
-------------
Tony Hill
hilla <underscore> 20 <at> yahoo <dot> ca