[Note: click to the previous page for discussion about the chosen benchmarks if you missed it!]

So, with all that said, here’s the BASE CONFIGURATION that we’re going to start off with. It’s all in the build posts above, but just to reiterate (or for those who immediately click the link in the first post to get down here) :

1ASUS M2N32-SLI Deluxe, final BIOS (5002)                          Cost:   $0 (scrap find)
2Antec Earthwatts EA-750 PSU                                               $0 (scrap find)
3Athlon 64X2 4200+ (K8, 2.2GHz)                                            $0 (scrap find)
4ATI FirePro MV 2260 PCI (NOT PCIe), 256MB (RV620 / HD3450 equiv.)         $5 from FreeGeek
5DisplayPort to DVI adapter                                                $1.89 shipped from Ebay
63GB (2x1 + 2x512) mismatched low-rent PC2-4200U                           $0 (scrap find)
780GB SATA HDD                                                             $0 (scrap find)

Display: LG monitor on DisplayPort->DVI, 1024x768

( ^^ I decided to run all tests at 1024x768 for consistency unless noted otherwise. Not everything supports 1360x768 and I didn’t feel like setting a different monitor up in this spot.)

CPU-Z, because I know someone's gonna want to see it:

(I forgot to run GPU-Z on the FirePro, will add it in later.)

And the results of the first round of benchmarks:

^^ click to enlarge, because holy butt.

(Sorry about the boring spreadsheet - I'll find a nicer way to display these as the build progresses. Just putting this up in the interest of MOVING THINGS ALONG.)

Not bad so far, even though we're using the crummiest configuration I could make and are badly GPU-limited by the weirdass PCI FirePro. Most things ran at any rate. I probably have a different definition of "playable" than a lot of gamers who run stuff from this era, but I can get by at 10FPS in a lot of cases (Prince of Persia, e.g.) It didn't run much better than that on the old P4.

Cinebench on the CPU is pretty slow, as you'd expect. The OGL test wasn't exactly hot shit either.

The PCI FirePro gives all kinds of weird problems in some stuff, other things it runs perfectly well. For example, in NFS2010 there is severe scenery pop-in and pop-out that makes it unplayable and awful to look at regardless of the low framerate. Unreal Tournament (UT99) refuses to run in any accelerated mode other than Direct3D, 16 bit colour. Picking 32b or OpenGL just crashes it instantly. I had this same problem on the P3 laptop setup too. UT2004, on the other hand, runs totally fine. Some benchmarks like GLExcess reliably run better in 32b color than 16b.

3DMark06 crashed on the HDR/SM3.0 tests, just like on the P3, even though this card technically does SM3.0. Magellan (fr-063) crashed immediately but the readme warns that it almost certainly will with less than 1GB VRAM. I don't even have AGP-type shared system memory to fall back on, so yeah, pretty much expected.

These issues will most likely disappear with a more conventional gaming GPU (spoiler: I'm already ahead of the thread, and they totally do), but after my smashing success with the FirePro in the other thread, I really wanted to see it in action in a more conventional machine of its era.

Crysis didn't make this round of benchmarks because I just forgot. 🙁 I'll go back and do it later.

...AAAAAAND, with that done, let's see about making some improvements! Should be fun, right? Stay tuned for the next episode folks!

So clearly from the above results, we're GPU limited. Let's do something about that. Just as a reminder, I'm using one of these:

...wait a minute, I'm using one of them? There are two sitting right there! Surely

xjas wrote:

So now I have two of these bizarre things, and if you think I'm not gonna try to Crossfire them, feel free to believe what you want. 😜

Ahahahahaha.

The RV620 chip in these cards "should" support Crossfire, although since there's no link connector, it just goes over the PCIe bus. You might notice we aren't using the PCIe bus. Well, I popped the second card in, it was picked up fine by the drivers & everything seemed to work. I couldn't find any option to disable or enable Crossfire in the version of Catalyst this driver package comes with, but what the hell, I fired up a game, and:

...I don't know what crossfire is supposed to look like, but according to Afterburner, both GPUs are under load and clocked at "running speed" (they clock down to 125MHz idle.) So, um, win?

Well, maybe. Despite this result, I really couldn't see any difference in framerate in every game I tried. Either the second card wasn't doing anything, or they were both crunching out frames but the whole setup was so bus-limited it just didn't matter. (I didn't go through everything, but Serious Sam, UT2004, Prince of Persia, and Fez all gave the same result.)

Honestly, I wasn't expecting much, and that's exactly what I got with this test, but I had to know. 😜

A pair of these things is a viable option if you really want to run quad monitors in on a non-PCIe board though.

I wonder if they'd play nice with the HD3850 AGP in the same machine? Project for another day...

With that out my my system 😜 let's see about a REAL GPU upgrade. Bring on the next contender!

Still sticking with Team Red, we have a HD4670 PCIe, in a slightly-upgraded variant from Sapphire with 1GB vram rather than the more typical 512MB. This is a really nice budget-midrange card from 2008 and is probably a pretty good match for my motherboard in terms of what people would have been running back then.

It's also fairly close in performance to the HD3850 AGP that was in the original P4, so it'll hopefully show off well how much better the "new" architecture I've upgraded to is.

Incidentally, I have a second, identical, one of these so we COULD crossfire it... but the other one's in my rackmount server thing and is a pain to get out. Actually, this one's going IN there (eventually) once I'm done with it here, to see if I can crossfire it under Linux. Maybe I'll get a result in this machine first.

Up & running...

Also, yes it can! This is where we introduce Crysis as a benchmark. We're running the demo version straight from, er, GeForce.com so I hope they didn't bork it on ATI cards on purpose. 😜 Seems to go fine here though.

It might have run on the FirePro but I don't imagine that being anything other than a miserable experience. I just didn't have it installed & didn't think of it before I swapped the card.

We're running at 1024x768 with all details on high (no "ultra" setting in the demo.) I'm taking the benchmark at this spot, which was basically stuck into the map by the devs to show off their graphics engine. It looks gorgeous, and already we're getting a perfectly playable (by my standards) ~25FPS. Sure, there's some slowdown, but nothing game-breaking.

Crysis came out in 2007, but this would have been a mid-range rig already with a "new" video card in 2008. What was all the fuss about again?

Here's the blob of benchmarks and tests from this round. Again, sorry about the spreadsheet. I'm open to ideas on how to display these a little less terribly. We're in the cyan column this time, and note that I added a units column so you can see what's being measured in each case:

Some interesting stuff to note - obviously the HD4670 blows the FirePro out of the water in nearly everything. Most 3DMark scores have doubled or tripled, and some game results have increased by an order of magnitude. (Check out Fez there, jumping from 16 to 129 FPS, or even NFS:HP2 which went from 33 to nearly 200.)

We're also seeing some small improvements in CPU scores (e.g. Cinebench)... I think these might be real gains due to the CPU no longer having to talk to the GPU all the way down in the bottom slot over the "old" PCI bus. Or maybe they're just within the margin of error. Hard to say.

There's one big exception though, and that is NFS2010. It eked out 9FPS on the old card, and on this one a whopping ... 14. I played a hell of a lot of this game on the P4/HD3850, most of which was around 10-15FPS, and we're still there. So, is it GPU limited on this setup? Or is it just so unoptimized for AMD's architecture that a couple of CPU generations makes no difference? Guess we'll find out!

BTW, if you're wondering if I also tried to crossfire the HD4670 with the glitchy, dying PCIe 3850 from page 1, the answer is of course I did, and it didn't work. 🙁

Bonus content! Here's GPU-Z for the HD4670 if anyone needs it:

BTW check out this Bus Transfer Rate comparison from Final Reality. The blue graph is the HD4670, the brown one is the FirePro. Nothing unexpected, but I just find it hilarious to see a discrepancy like that from two cards nearly the same age & generation. "Well, there's yer problem."

The bottleneck I figured it made sense to take care of next was RAM speed. I originally used two mismatched pairs of low-end PC2-4200 (533MHz), which was already way underspecced for this board. Well, for once in my life I went out & bought good RAM for this project. Here it is:

Yes, it did come in that gloriously pink & shiny makeup box, and hell yes I'm keeping it.

I have more, but this will do for now; I want to see the gains due to memory speed, not quantity. We're still bumping up from 3GB to 4, but nothing I can do about that. Incidentally this is the most expensive component in this system so far at twenty bucks out of pocket.

So first we pull the old stuff out, and

Oh for fuck's sake. I knew the crowded components in that area were going to be an issue. With a bit of "not-quite bending" I was able to get this top stick out without pulling the whole damn thing apart, mercifully. Needless to say the new RAM went in the BOTTOM two slots.

Here's the old stuff vs the new for comparison:

^^ It's starting to look a bit "blingy" in here. It would have looked cooler if I'd gone for OCZ Gold, but the Platinum has better timings.

I didn't realize it at first, but the timings for this stuff are listed rather conservatively in the SPD. I think they do that to make it compatible at 1.8V. It doesn't say it on the sticker, but its rated timings require 2.1V which is "out of spec" and needs to be set manually.

Fortunately this board has a ton of options for RAM tweaking in the BIOS (and you can up the voltage in 0.25V increments on another screen) :

I had to learn basically everything about RAM tweaking in the span of a day at this point. I still don't have any idea what I'm doing, but I got something stable that worked, so I'm claiming success. The "stock" timings it booted with were 6-6-6 18 24 2T @ 1.8V; I went with 5-4-4 15 20 1T @ 2.1V for my two modules.

I ended up running a bunch of benchmarks at 666-18-24 2T before I fiddled with it & left them in just to see how things improved. That was pretty interesting!

Incidentally I'm introducing a few more "workstation"-oriented benchmarks here, as I really wanted to see results from the RAM. First is SuperPi at 1M and 4M digits to measure pure single-threaded computation, PassMark v8 for both an overall system & memory benchmark, and ScienceMark, which is a whole suite of multi-threaded number crunching benchmarks designed to simulate mathematical workloads using common libraries (BLAS) & languages (Fortran, etc.) found in the sciences. (I'm a physicist and I do work like this all the time, so this is interesting to me.) I chose ScienceMark's single-precision matrix multiplication & memory benchmarks, rather than running the whole thing.

The BLAS algebraic library, which this uses, has specific optimization for 3DNow! which is cool. Put those creaky instructions to good use.

So finally, here are the results:

Note we have two new columns, one for the RAM upgrade at 6-6-6-24-18-2T and the next one at 5-4-4-15-20-1T. Honestly, I was blown away with how much of a difference the faster RAM made. I was expecting a little bit here & there, mostly in the computational stuff, but virtually every game and 3DMark benefited, sometimes by a lot. UT2004 gained 32FPS from RAM speed alone, for example.

Even the timing made improvements across the board; not always by a huge amount but noticeably & consistently. You can see the result in the 3DMark scores a little, but the computational benchmarks really showed it - check out my 4M SuperPi time dropping by 4 seconds from the timing alone. Not bad.

In conclusion, if you don't want bottlenecks, buy decent RAM and set the timing right. Even on creaky old platforms like AM2 that "the internet says" aren't very sensitive to RAM speed. Lesson learned.

Guys, I'm thinking of discontinuing this thread. I've been having a lot of fun with this build, but the lack of response here has been pretty underwhelming & I have other projects to move on to. I'm not whining or begging for "likes"; if nobody's interested than fair enough, but it's not worth the hours of work I've put into writing everything up if it's not going to entertain or benefit anyone else.

As usual, I'm a few steps ahead of the thread in real life - we still have at least two more GPU upgrades to get through (and you better believe I'll be putting that "SLI" board to good use), another RAM upgrade, a CPU, cooling, and I was going to try my hand at overclocking & tweaking which is another thing I want to learn about. If you've been following along and want to see the thread continued, just say so 😀 otherwise I'm gonna put the energy into something else.