2.4B/2.4C/865PE/875P/nForce2/2500+ Comparison...
Round 2...grab a brew, cuz there's a ton of stuff to check out here.
A lot's changed in the last six months, so we're going to take a close look at the current mainstream PC platforms, and try to dig a little deeper than we're used to seeing from typical reviews.
We're going to look at the nForce2 version 2.0 with a Barton 2500+, and then the 2.4B/533 and 2.4C/800 procs with both the 865PE and 875P.
For the nForce2, we'll be using the Abit NF7-S ver. 2.0.
For the 865PE, the MSI 865PE-LS, and for the 875P, the Abit IC7.
Common hardware is exactly:
-Radeon 9700 Pro, Cat 3.2, default Peformance settings
-Soundblaster Audigy
-2x256MB Kingston HyperX PC3500
-Western Digital Raptor 10k SATA Drive
-Samsung 48x24x48x16 Combo Drive
-Floppy
-Explorer 3.0 Mouse USB
-MS Internet Keyboard PS/2
-Windows XP Professional, all updates, DX 9.0a
When I say exactly, I mean I used the exact same pieces for each scenario...I wanted to eliminate any possible variance.
For a brief moment, I was going to do a stock cooling comparison, but as I learned last time, this makes AMD rather boring. So, for the Barton, I ended up using an SLK 800-U cooler, and a freakin' loud Delta 6800rpm fan. I had a regular 80mm on there, but it just couldn't handle 2.0v, so I went to the hair dryer. For Intel, I stuck to my guns, and stayed with the stock Intel heatsink.
Everything else is bone stock...no other mods. I'm doing it this way, because I wanted to represent what can happen for the average, everyday air overclocker, just looking to get their bang-for-the buck. Hopefully anyone, from the hardcore to the newbie, can take something good away from the comparisons and tendencies of these platform results.
The tests run are Sandra 2003 Memory Bandwidth, Comanche default 640x480, Quake III 1.17 demo001 default Fastest, Jedi Knight II jk2ffa default Fastest, Unreal Tournament Flyby and Botmatch with Anand's HQ scripts at 640x480, and 3DMark 2001 SE at both 640x480 and 1024x768 default. What I'm trying to accomplish here is a good platform versus platform comparison, so I used the lowest default settings on each bench, to help take away the video card bottleneck. I also ran 3DMark at default, because it's so recognizable. Should prove interesting.
Let's roll...
Stock comparison...
http://www.monsterboxpc.com/roundup2/stock.gif
There's the whole gang at stock speeds. Note that both the P4 boards default to 4:5 with a 533 proc. Also note that the 865PE is at CAS Latency 2.5 at 1:1...this is because the board would default to 5:4 speeds any time CAS was set to 2. This actually doesn't matter that much anymore with the current P4's...I'll go into that more when I look at the boards specifically, but just note the fact for now.
Right off the bat, notice that P4's are really strong in Comanche, QIII, and JKII, while the Barton excels at UT2K3 and 3DMark. This trend will continue.
Also notice that, with the P4 533 proc, the 865PE and 875P are practically identical (in fact, the MSI board beats the Abit). This is because the 875P's PAT feature was only designed to work with 800 FSB procs. With a 533 processor, the 875P is, really, an 865PE...they're the same silicon. I'll again get into this later.
Last note to take away from the stock comparisons...those are damn nice gains by the 800 FSB P4 part, and when PAT kicks in, it's a beautiful thing ;)
In fact, it pretty much throws the AMD Performance Rating versus P4's clockspeed out of whack...where the 2500+ pretty much handles a P4 2.4/533, the 2.4/800 turns around and smacks the 2500+, especially on the 875P.
Stock is boring...let's dive a little deeper...
A closer look at the Barton/nForce2...
Version 2 of the nForce2 chipset just improved on what was already a nice chipset. It allowed me to hit 200FSB in dual-channel with nice timings, where my old board would crap out at 190 in dual-channel.
Barton, however, seems to be a mixed bag. It's faster than the non-Bartons clock-to-clock, but it's, well, kinda quirky when overclocking.
First, it took a full 2.0 core voltage for me to be able to hit the same 2400MHz clockspeed that I was able to nail with my 2400+ proc at 1.85v. It would seem the extra cache makes it a little tougher to push the clockspeed.
2nd, FSB and memory timings adjustments are simply 'different' than with the non-Barton procs. For example, at 200 FSB, if I set Precharge delay/Cycle time to 6, it'd lock up 3DMark. Turns out, this didn't matter, because 7 seems to be a really well-performing setting for the nForce2. I literally got better results from 7 than from 5 or 6 (5 was actually really bad). The one that blew me away was CAS Latency. Again at 200, any time I set it to 2.5, it'd crash 3DMark (sometimes even rebooting the machine). CAS 2, however, was gold. Why the heck would a more aggressive setting be more stable??? It'll take a smarter man than me for that one, but it's true. And neither one of the above occurances happened with a non-Barton. Oh well...I'm happy with 2/7/2/2 (I really hate memory timings after all this).
I still love AMD's non-locked multiplier, and this time around, I only managed to corrupt my OS once when trying to push this sucker (imaging rules). Let's play around a little...
Different levels of 2000MHz...
http://www.monsterboxpc.com/roundup2/2000.gif
Fun to look at. I bumped up the clockspeed to 2000, at took a look at that 2000 at 133, 166 and 200 FSB's. This board cooperated with me marvelously as well...CPU-Z reported 133.6, 167, and 200.5, which gave me 2004, 2004, and 2005, repectively. Can't get much tighter than that.
This chart gives you a good idea of how nicely the nForce2/Barton benefit from FSB gains...AMD's upcoming push to 200 FSB parts will definitely help.
Let's go...
Barton at 2000 and 2400MHz...
http://www.monsterboxpc.com/roundup2/20002400.gif
The left two columns are at 2000, to see what the 200 FSB brought to the party, and the right two columns to see what the extra 400MHz of clockspeed is giving up. The total gain from 2000/133 to 2400/200 is simply delicious.
Just like last time, notice that column 2 versus column 3. The push to 200 FSB gave me much greater results in everything but Comanche (a VERY CPU dependant benchmark) than the raw 400MHz of clockspeed did (column 1 versus 3, for those in the back of the class).
2400MHz was everything 2.0v and the hair dryer could give me. And 200 FSB was a good as I could pull off with these timings in dual-channel.
I happen to like dual-channel...it seems a tad bit more responsive to me in actual use. Still and yet, let's take a look at single-channel. Last time, we figured out dual-channel was worth about 9-10 points of FSB...
Single-channel versus Dual-channel...
http://www.monsterboxpc.com/roundup2/singledual.gif
...there is is with 8 FSB. CPU Z says that's 200.5 versus 208.8, which results in 2406 versus 2401. The 200 dual-channel is just edging by the 208 single-channel scores in everything but the 3DMarks, so this is basically tie-ballgame. I'll just go ahead and say dual-channel is worth 9 FSB.
Good fun. Let's move on to Intel, where there's a lot more shakin' and bakin' going on...
A closer look at Intel...
This was a load. Just a ton of new stuff, what with two new chipsets and the whole shift to the 800 front-side bus. I'm just going to dive right in to the numbers, and explain as I go:
The 2.4B/533 Processor...
http://www.monsterboxpc.com/roundup2/533.gif
I took a ton of benchmarks with the 533. When I sat back and took it all in, then took a gander at the 800FSB results, I just decided to group all the relavent 533 numbers in one chart. Plain and simple, the 800 FSB parts are where it's at now.
Still, there's some good info to take from all this.
-First, there is a very clear benefit in running an asynchronously-higher memory bus with these boards. A lot more than I would have imagined...one, because 1:1 matches the FSB bandwidth already, and two, because Athlons/nForce2 suck when run asynchronously. So if you can, use that 4:5 divider, folks.
-2nd, as mentioned previously, the PAT enhancements of the 875P chipset do NOT function with 533 parts. The 875P is basically an 865PE when used with a 533 proc. This is pretty evident throughout the whole chart, particularly when looking at the Sandra Floating-Point memory benchmark (the second number), which is typically quite consistant. The differences that exist can simply be chalked up to two different brands of boards, where there's commonly slight differences. The biggest differences in CPU Z were at 166, where the 865PE reported 166 versus the 875P's 166.7 (not so coincidentally where the Abit board finally took a few meager victories).
3GHz was as high as I could push this thing...a tad disappointing, as there's some gem 2.4/533's out there (I've had a few), but again, it's not gonna matter...
The 2.4C/800 Processor...
http://www.monsterboxpc.com/roundup2/800.gif
Yum. First we have the 5:4 divider, both at good timings. We can immediately see PAT doing it's thing here, and the good news is, it works asynchronouly as well.
Next, at 1:1. This really is some nice stock-performance here, folks...what's funny is, I 'felt' like I was already overclocking, just because of the 200 front-side bus (I mean, this used to be holy grail material with 533 parts). This feeling went away quickly, though, I'm happy to say.
I said I'd come back to that CAS Latency thing earlier...that's where the column on the end comes in. As I said (and others have done complete reviews on), it really doesn't matter when PAT's latency enhancements are in effect. Sandra has a mini-picnic, but everything 'real' just yawns...heck, some benches are even slower. I will note, though, that with the 865PE (and inherently, the 875PE with a 533 proc...i.e, no PAT), CAS Latency does actually matter. It's a small difference, but it's there.
This is a good time to go into memory timings with the new boards, actually. The most sensitive seems to be precharge delay/cycle time again, just like with the Athlon. This should be the first to go...and amazingly, I actually got a corrupt OS when pushing higher with this setting low (this actually made me happy, in a sic way...of course, I'm disturbed).
It doesn't actually do much for performance, though...I was able to overcome the drop from 6 to 7 with 1 point of FSB. So I just left it at 7 permanently. The next most sensitive was CAS Latency...I was able to hit 212 before I had to drop this to 2.5. And again, there's practically ZERO difference in anything but Sandra with this setting...so unless you're that freak that 'plays' Sandra all day, just set it to 2.5 and forget it (I really don't like Sandra...I just include it for reference). The next most sensitive setting is RAS-to-CAS. Dropping this one actually made a difference. However, it also let me take off quite nicely with raw memory clockspeed. 218 1:1 was the best I could do with it at 2, but I could hit 230 with it at 3. Very nice. Basically, if you can, run this one at 2, but if it hinders you by 2 points of FSB overclocking or more, go ahead and drop it to 3. RAS precharge makes a difference as well (less than RAS-to-CAS), but I didn't find it sensitive at all. In fact, when I dropped it, I couldn't go any higher...so, basically, leave it at 2. Memory timings overall really don't make a huge difference with the 875P in general...I'll save you the pain of 27 charts worth of benches, and just say the average difference between 2.5/7/3/3 and 2/5/2/2 was a measly 0.5-2%.
Hey, this is a good time to look at those fancy memory dividers...
Memory dividers...
http://www.monsterboxpc.com/roundup2/dividers.gif
Wow. I found this one really interesting...and breaking it out into percentages really made it hit home. The first percentage row shows the individual differences on each benchmark when dropping from 1:1 to 5:4. The Composite figure is simply those seven percentages divided by seven.
That's right, folks...the drop from 1:1 to 5:4 is an average of 2%. Overclockers everywhere should be smiling about right now.
In the third score column, I dropped to 3:2, and the following percentage columns are the percentage drop of 3:2 from 1:1 and 5:4, respectively.
3:2 is a 4.91% drop from 1:1. And 3:2 is a 2.85% drop from 5:4.
This whole chart actually shocked me...but in a good (GREAT) way. I was expecting MUCH greater hits. I'm SO glad to be wrong on that one ;)
Good, good stuff...let's roll on...
Different levels of 3000MHz...
http://www.monsterboxpc.com/roundup2/3000.gif
Another neat one. 3 G's with the 533 and 800 parts on both boards.
800's where it's at, folks. And PAT makes it that much sweeter.
Hyperthreading or not...
http://www.monsterboxpc.com/roundup2/ht.gif
Hyperthreading, as cool as it is, has been known to hinder benchmarks a tad. I left it off in all these tests as well, because I wanted to compare the 2.4's directly (i.e. the 533 doesn't have it).
What I DID find cool here is, the 'HT hit' was practically zero on the 875P board...nice! Maybe a little PAT side-benefit here? Whatever it is...coolness.
Ok, let's bring on the finals...
'You always saved the best for last'...
http://www.monsterboxpc.com/roundup2/max.gif
There's the whole gang again, in their Sunday best.
Truth be told, these are all really strong, but the clear winner is the 875P/800 FSB part. And there's some headroom in there...I got to Windows in the 280's, and could even pass benchmarks. 3250/271 is the best I could say that it was 100% stable, with Hyperthreading on as well. Pretty damn swanky for a stock-Intel heatsink accomplishment.
Now, I'm gonna dive in real deep. I hadn't really seen this out there yet in this kinda of detail, and was always curious, so I did it...enjoy:
Getting freaky...
I wanna see exactly where all of these gains are coming from.
Furthermore, I want to see it AMD versus P4 style.
Sooo...
Clockspeed + FSB scaling...
http://www.monsterboxpc.com/roundup2/clockfsb.gif
PHEAR my mighty hyperthreaded dual-channel P4 1.6GHz!!! Man, that benchrun felt like an oldschool timewarp. Glad it didn't last long.
This represents a 50% overclock for both procs, both equal clockspeeds and FSB's. Again, in each percentage column is each benchmark's benefit, with the Composite being the average. The Efficiency percentage is the composite percentage gains divided by the actual overclock, 50%. Cool, eh?
Both are quite efficient...the P4 gained a little more. I think what we're seeing here is the P4's longer pipeline starting to pay off. While it's not so good at low clockspeeds, it gets better as you ramp it up.
Note those 3DMarks...this is telling us it's time for bigger, badder video cards. The next 'leap' is going to have to come from our friends at nvidia and ATI. This becomes painfully obvious as the CPU clockspeed rises.
Let's break it down even further...
Raw FSB scaling...
http://www.monsterboxpc.com/roundup2/fsb.gif
Here we have simply a 50% FSB gain...clockspeeds for both are the same.
Quite interesting, no? The Athlon gains a higher benefit from raw FSB gains. A fair guess would be because AMD's FSB is dual data rate, versus the P4's being double that (quad data rate). In any event, they both like FSB, and it's more important to the Athlon.
Once more into the fire...
Raw clockspeed scaling...
http://www.monsterboxpc.com/roundup2/clock.gif
The script is flipped. What we have here are equal 50% clockspeed gains, with both FSB's being equal. The P4's gaining roughly 10% more out of it's clockspeed gains than the Athlon, at an efficiency clip of nearly 20% better.
It appears that the P4 still has a pretty good ways to go, while the Athlon line is nearly played out. Perhaps if AMD could figure out a way to go quad-FSB...but we already know they've got their eggs in the Hammer-basket.
In the last roundup, we had a virtual draw. This time around, I've got to hand it to the 875P and the 800 FSB P4 parts. The benchmarks show it. Furthermore, and to some of you, more importantly, the 875P/P4 has too many 'intangibles'....kick-butt stuff that doesn't show up in benchmarks:
-CSA Lan: The Athlon/nForce2 doesn't have an answer for this. CSA Gigabit LAN hauls butt, but even better, it's got it's own dedicated bus. You do not have to worry about flooding your PCI bus with CSA Lan anymore. You can get Gigabit with the AMD setup, but it will be on the PCI bus.
This feature isn't all that important to me, nor the majority of home users, but for those with pretty intense networking needs, this is a deal-breaker.
-ICH5/ICH5R: ICH5 supports two SATA ports natively....nForce2 does not. Furthermore, both those SATA ports, as well as the ATA100 ports, are no longer on the PCI bus either. Intel went out of it's way to clear up some bottlenecks, and it shows. Lastly, the ICH5R version supports RAID 0 (and in the future, RAID 1). Here's the thing...each of those SATA channels is it's own bus. That's right, combined together, it's 300MB/sec of potential (well, 266MB/sec is the limit of the bus, but it's not like we'll be there any time soon):
http://www.monsterboxpc.com/roundup2/drives.gif
That's a pair of 10k SATA Raptors, folks...on the onboard SI controller with the nForce2, on the ICH5R controller with the 875P. Note that PCI limit for the nForce2's buffered read. Then note that better-than-single-channel-SATA150 161MB/sec the ICH5R puts out...all the while not being on the PCI bus. Good stuff.
-Hyperthreading. Virtual dual-processing. It's now for the masses, available on all 800 FSB parts down to 2.4. This, to me, is still the big one, and still the hardest to describe. Again, the best example I can give is, this is an All-in-Wonder Radeon 9700 Pro. With Hyperthreading on, I can start capturing video, fire up Comanche, and not drop a frame of video.
If I turn off Hyperthreading, or do this on the Athlon/nForce2, I drop 30-40% frames...i.e. pure garbage video. And it doesn't have to be as strenuous a task as Comanche is...I can fire up Quickbooks or Excel without Hyperthreading, and get frameloss. Heaven forbid I compress a file or something like that. In a nutshell, Hyperthreading makes mulitasking smoother overall, and actually makes some actions possible that weren't before. You'd have to go dually with Athlon to accomplish that.
There you have it. For the time being, Intel's simply got AMD on the ropes. 865PE is cool, but 875P paired with an 800 FSB proc is phenomenal. This new chipset just feels like something bigger and better, and it is.
AMD's game is it's bargain...in fact, if one were to go AMD with the thought of overclocking, I'd tell 'em to just grab one of their $50 Thoroughbreds, and go to town. That's AMD's clear advantage...but unfortunately, it's also really hurting their bottom line as a company. So outside of the cost advantage of low-end Athlon overclocking, it's all 875P and the 800 FSB Pentium 4's.
Good times. This was a lot of work, and a lot of fun.
I'll be back when the Hammer drops...and see how the Prescott counter-punch hits.
Cheers ;)
Addendum:
Since I don't like going back and changing articles after the fact, I'll just add this on the end.
The PAT results above are correct, but the 'cause' is actually different.
Here is how PAT actually works:
There are two paths...a PAT path, and a standard path.
Both 865PE and 875P have both paths.
The 875P is ALWAYS on the PAT path. It is on this path with a 533 proc or an 800 proc, with any ratio. PAT is always on.
The 865PE is also on the PAT path when a 533 proc is installed. However, it is NOT on the PAT path when an 800 FSB proc is installed (unless 'tricked' to do so).
That's it in a nutshell...thanks!
