Question regarding Intel Pentium 4s

dfirestein · Post by **dfirestein** » Mon Jun 27, 2005 8:06 am

Going to probably be buying a new computer tomorrow or the next day, so hopefully this will get answered in time...What is the difference between an "Intel Pentium 4 Processor 650 w/ HT Technology 3.4 Ghz 800 Mhz FSB 2 MB Cache" and "Intel Pentium 4 Processor 550 w/ HT Technology 3.4 Ghz 800 Mhz FSB 1 MB Cache"? Which should I be buying and which will last me the longest before I have to upgrade my desktop? Price difference is about $150.

Deacon · Post by **Deacon** » Mon Jun 27, 2005 8:51 am

If you've got the cash and the motherboard to support it, get the extra megabyte of cache.

RandyWang · Post by **RandyWang** » Mon Jun 27, 2005 1:11 pm

The actual speed difference between the two won't be all that big, but Deacon is right - if you have the extra money, get the one with the 2Mb cache.

Zagaroth · Post by **Zagaroth** » Tue Jun 28, 2005 5:35 am

the 650/550 is just model number. The important issue is the cache. it may not increase pure speed of the processor, but it speeds things up similiar to how an increase in RAM speeds the entire processs up. Think of it as a three step memory:

Hard Drives hold the most, but take the longest to access.

RAM holds less and that memory goes away when it looses power, but it is a LOT faster to access.

Cache is like micro-RAM right on your processor, so you save even more time by not having to go through the mother board to access, but it is very small.

of course, there is a small chance my memory of how cache works is flawed, if so please correct me. It's a detail I haven't paid as much attention to as I should have.

pc486 · Post by **pc486** » Tue Jun 28, 2005 9:10 am

Almost right on the money there Zagaroth. There are things slower than hard drives (tape, offline storage) and faster than cache (registers) but the idea is the same. Zagaroth is refering to memory hierarchy. The only reason it exists is because fast memory is expensive. Wouldn't it be nice if you can load your system up with 100 gigabytes of system memory at the price of hard drive memory (~0.50 USD/Gig and dropping fast)? The price of L1 cache compared to system memory is the same way

.

Since we are talking about cache... be careful. While increasing L1/L2 cache does have a net positive effect on system performance do not think of it as a magic bullet. With most applications these days there is almost no difference between 1MB and 2MB of cache either due to the type of application or the quite common not-cache-aware programmers out there. Even cache size doesn't tell you much because if you compare different chips (say AMD x86 versus Intel x86) as you may end up comparing against a different way of managing cache. Without getting too technical, improving cache management in one area (the number of dimensions it can handle, aka "2-way-cache" vs '3-way-cache") can lead to the equivalent of doubling cache memory or better. It is entirely possible for a chip with 512kB of cache to outperform another that has 2MB. In other words, to judge a chip based on cache size is totally bogus.

Now, since we are comparing two very similar Intel chips and if your mind is set on either of these Intel chips and nothing else (no AMD, Via, whatever) then yes, if you have the money to pay for the extra megabyte of cache go for it. Extra cache will only hurt your wallet and not your performance.

Tagrineth · Post by **Tagrineth** » Tue Jun 28, 2005 6:11 pm

Speaking of cache differences (aside)

With AMD you also get a larger L1 cache - a lot larger than Intel's - but the numbers are deceptive. L1 cache works differently from L2+, and the main reason AMD's is so large is to make up for AMD's weak branch prediction in their processors.

Random GK boost~!

pc486 · Post by **pc486** » Tue Jun 28, 2005 11:49 pm

That's interesting. I wouldn't expect a larger L1 instruction cache size to help out performance much based on failed branch predictions because the pain in a failed prediction is a flushed pipeline rather than a flushed cache (although a flushed L1 instruction cache may happen if it is small enough) . Is there a link to the source of this or a research paper? 'Twould be an interesting read.

Also, Intel better have one hell of a branch prediction mechinism. With a 20 stage pipeline in the P4 series a failed prediction hurts big time.

T8y8 · Post by **T8y8** » Thu Jun 30, 2005 3:24 am

Also, the 6xx lines are 64-bit EMT64 technology.

Tagrineth · Post by **Tagrineth** » Thu Jun 30, 2005 8:28 pm

[quote="pc486";p="513382"]That's interesting. I wouldn't expect a larger L1 instruction cache size to help out performance much based on failed branch predictions because the pain in a failed prediction is a flushed pipeline rather than a flushed cache (although a flushed L1 instruction cache may happen if it is small enough) . Is there a link to the source of this or a research paper? 'Twould be an interesting read.

Also, Intel better have one hell of a branch prediction mechinism. With a 20 stage pipeline in the P4 series a failed prediction hurts big time.[/quote]

Larger L1 -> store more pre-prediction data, makes the "retries" happen faster.

pc486 · Post by **pc486** » Thu Jun 30, 2005 9:44 pm

[quote="Tagrineth";p="514117"]Larger L1 -> store more pre-prediction data, makes the "retries" happen faster.[/quote]
Older AMD Athlon chips hevily depended on a branch predition table which I believe was not kept in cache and thus a larger L1 instruction cache wouldn't effect the branch predition. I don't have my AMD64 architecture books with me so I cannot comment on current AMD technology. Intel hasn't given me any free architecture books except for obscure processors so I know little about Intel's x86 stuff

.

Nit-pick: Instruction retires do not happen faster with better branch predition. Rather the ratio of retired instructions to issued instructions improves. Improving branch prediction algorithms will not make a processor magically execute instructions faster

.

Tagrineth · Post by **Tagrineth** » Sat Jul 02, 2005 9:20 pm

I haven't dealt with this material in years, so I'd rather drop out of this discussion... I can't remember further than what I've already stated.

I can make one comment though...

Nit-pick: Instruction retires do not happen faster with better branch predition. Rather the ratio of retired instructions to issued instructions improves. Improving branch prediction algorithms will not make a processor magically execute instructions faster Razz.

No, retries will not go faster with better prediction, but they will happen less often, in some cases a good deal less often. Intel has a fantastic branch prediction unit in P3 and P4, estimated at well over 90% accurate. If I remember correctly, the AMD K7 "Thunderbird" was estimated at only a little over 60%... or was that K7 vanilla? Damn, I can't remember clearly now. Anyway... research research research and I really don't have time to look all that up at the moment. ^^;

pc486 · Post by **pc486** » Tue Jul 05, 2005 4:59 am

[quote="Tagrineth";p="515166"]I haven't dealt with this material in years, so I'd rather drop out of this discussion... I can't remember further than what I've already stated.

[/quote]
Yeah, I'm hitting my limit too

.

Intel has a fantastic branch prediction unit in P3 and P4, estimated at well over 90% accurate. If I remember correctly, the AMD K7 "Thunderbird" was estimated at only a little over 60%... or was that K7 vanilla? Damn, I can't remember clearly now. Anyway... research research research and I really don't have time to look all that up at the moment. ^^;

Neat. I'll take a look to see if AMD's stuff is up to the P4 90% rate (that's huge!).

Post by **Martin Blank** » Tue Jul 05, 2005 6:03 am

There are some articles that discuss the architectures in some detail over at AnandTech. Don't recall who wrote them, but they went into some pretty significant detail, and are not short overviews.

Real Life Forums

Question regarding Intel Pentium 4s

Question regarding Intel Pentium 4s

Who is online