CPU price/performance graph

[halbhh]

Hotfoot, one more important question:

As you can see, at resolutions that actually matter, the x2 6000 is close to the e6600 at gaming:

http://www.anandtech.com/cpuchipsets/showdoc.aspx?i=2933&p=8

But your numbers make the difference look significant on the graph, as if the e6600 is a lot better at gaming.

Anand shows the difference at the e6600 better by about 4% roughly (mathemetical average from 4 popular games).

So this broad measure shows the e6600 4% better at games over the x2 6000.

Your metrics show a 9% superiority at games.

I don't think your game metric is broad enough. For instance, Oblivion at 1600x1200 is very cpu demanding, and very representative of a good critical measure of gaming performance, when averaged with other popular games. Witness the much lower frame rates from Oblivion vs their other games benches.

This is more meaningful, because the human eye cannot distinguish any differences above 70 fps. So we need demanding games that cause lower frame rates to have a something meaningful for real world results.

 

[HotFoot]

Yes, the numbers in the table are one week old, but the graphs are current. I thought I noted that in my update post. The updated table will be posted after I get home. This is why the prices don't match, as you've pointed out. For now, if you want to double-check things for me, please open the spreadsheet file in the .zip download. Thanks for your watchful eye!

As for the poor performance of the 6000+, I very much expect that this is due to the lower memory clock. As explained elsewhere in the forumz, the memory clock is determined by dividing the CPU frequency by an integer. For example, consider DDR2 800 MHz memory being paired with the 5600+ and 6000+ processors, witch are at 2800 and 3000 MHz, respectively. The highest 2800/i, where i is an integer is 400 MHz, which gives you DDR2-800. Moving up to 3000/i, i must be increased to keep the result at or below 400 MHz, so i=8, and the memory base speed is then 350 MHz, for DDR2-750. See this link for a table of CPU frequencies and DDR2-6400 actual memory speeds.

The sweet spots for maximising your memory speed are then CPUs with frequencies that are multiples of 400, if you're using DDR2-6400 memory, if you're going to run at stock settings.

 

[HotFoot]

I don't think your game metric is broad enough. For instance, Oblivion at 1600x1200 is very cpu demanding, and very representative of a good critical measure of gaming performance, when averaged with other popular games. Witness the much lower frame rates from Oblivion vs their other games benches.

This is more meaningful, because the human eye cannot distinguish any differences above 70 fps. So we need demanding games that cause lower frame rates to have a something meaningful for real world results.

I totally agree. Do you suggest, then, that Oblivion would be an appropriate addition? I think the gaming results this week are more representative since F.E.A.R. has been added. Again the numbers in the table are old, so if you want to see the advantage of the E6600 over the 6000+ in games this week, the scores used in the graphs are:

6000+: 1.62
E6600: 1.73

So the edge of the E6600 stands at 7%, which is a little closer to your investigation's result of 4%. Next week I will see about getting Oblivion results into the mix and we'll see how that works out.

 

[halbhh]

Yes, the numbers in the table are one week old, but the graphs are current. I thought I noted that in my update post. The updated table will be posted after I get home. This is why the prices don't match, as you've pointed out. For now, if you want to double-check things for me, please open the spreadsheet file in the .zip download. Thanks for your watchful eye!

As for the poor performance of the 6000+, I very much expect that this is due to the lower memory clock. As explained elsewhere in the forumz, the memory clock is determined by dividing the CPU frequency by an integer. For example, consider DDR2 800 MHz memory being paired with the 5600+ and 6000+ processors, witch are at 2800 and 3000 MHz, respectively. The highest 2800/i, where i is an integer is 400 MHz, which gives you DDR2-800. Moving up to 3000/i, i must be increased to keep the result at or below 400 MHz, so i=8, and the memory base speed is then 350 MHz, for DDR2-750. See this link for a table of CPU frequencies and DDR2-6400 actual memory speeds.

The sweet spots for maximising your memory speed are then CPUs with frequencies that are multiples of 400, if you're using DDR2-6400 memory, if you're going to run at stock settings.

Thanks! You and Jack have cleared this up for me. I've only overclocked once, and when I saw it made no difference for my situation I lowered my 4200 back to base speed in order to use QuietN'Cool, which I like for my silent computer. Memory speed has never mattered for me, but it's nice to learn this detail so easily, thanks again.

btw, I see your other post re the gaming results. I'm sure there is no absolute standard, and it would be nice if you could indeed have just 3 representative games for your own efforts to be modest! I do see Oblivion used a lot, since for instance it was used in the article on the 8800 graphics vs cpu. I get the vague impression it's a nice mix of cpu and graphics demand, but it does indeed require a powerful graphics card to make that the case! Nonetheless, for someone building a new build or upgrading right now, it won't be that long until that level of graphics power is reasonably priced. Just 4 months at this point until ATI helps out I believe.

 

[halbhh]

THG finally put out another price/performance analysis today. It is interesting, so I recommend reading it. The writer said they had a "two week hiatus" which means they had one out I suppose when we were waiting and waiting. It was practially invisible it seems to me. In fact, to be honest, I searched for one more than once over the last couple of weeks!

 

[HotFoot]

I've updated the table finally. Sorry for the lag behind the graphs and the confusion this may have caused. Next up: Oblivion.

Back to THG's price/performance update, I was also interested to see the phrase "two week hiatus", since I cannot find any updates since January. It is nice to see the new higher-end AMD CPUs added, and once again I'd say I hope to see the lower-end C2D products on the graphs soon. It will certainly be interesting to see what kind of performance the sub-$100 C2D derivatives will have when they arrive over the next months.

As for me, I'm certainly coming to realise that CPU performance is having less and less of an impact on what system I choose to build next. While I do want a minimum standard of performance, even the X2 3800+ would make me happy. Since I want to improve my media centre PC, and reclaim my desktop hardware for my gaming/office machine, I think my next purchase will be goverened by the available chipsets. The AMD 690G is looking very good to me, especially with the integrated DVI/HDMI support. With the current pricing on "low-end" X2's, I should be able to put something together that handles encoding, file sharing, and playback with FAH on the side for super cheap. It was only a year and a half ago that I paid over CAN$300 for a PD820. I still get overwhelmed by the progress since then (oh yeah, and the 3800+ that I wanted at the time was way out of reach at $450 for me).

 

[fishboi]

One cant open up the THG charts and englarge the pics. Boo.

 

[HotFoot]

I think you get a closer look at the graphs if you go into the beta site. The link is in the original article.

As for Oblivion results, I'm having trouble finding benchmarks for my reference chip: the P4 630. Does someone have results for a 630 running Oblivion at 1600x1200?

 

[halbhh]

I think you get a closer look at the graphs if you go into the beta site. The link is in the original article.

As for Oblivion results, I'm having trouble finding benchmarks for my reference chip: the P4 630. Does someone have results for a 630 running Oblivion at 1600x1200?

hmmmm....Well, two ways to do without would be either A) extrapolate from the other chips and then normalize, or B) just normalize all the other chips (without bothering with the 630), since for equal weighting with the other games, you have to anyway (ask me if you need technique).

 

[HotFoot]

I could switch the standard to the X2 3800+ chip, this wouldn't be any trouble. This may be a more representative baseline for CPU performance in 2007 in any case. It will also make it more convenient to keep the origin on the graphs.

 

[halbhh]

no doubt a fine solution, if you prefer it. Kinda interesting actually, since then I can immediately gauge all the numbers vs my chip since I'm exactly 10% more frequency.

 

[HotFoot]

I've not yet found a review that has Oblivion 1600x1200 results for the 4800, 4400, 4000, or 3600 chips. Basically any of the former 1MBx2 cache parts that have been replaced with x.5 multipliers for Brisbane. I've gone through some articles from AnandTech for the results I've got so far, but I haven't found another site that does Oblivion benchmarks. There must be some others out there.

 

[halbhh]

Frankly, I would be inclined to just interpolate those, using frequency for similar cache sizes. If an odd memory speed (below 800) created a penalty of even a big 5% in a chip in a game for instance, but you have 4 games, with 3 already known accurately, then interpolating the 4th game with a 5% error would make about a 1% error in the final average.

But a 1% error, or even a 2-3% error, is just a lot smaller than the effect of choosing certain games instead of others!

Also, we need to remember that expert users need about a 10% difference in speed to be able to detect it! (so 10% would be too much, but 1, 2 or even 5% is unimportant)

Re using a base chip (like the 630): While it's convenient, you can normalize a set of data without a base chip. Any chip can be chosen for a base for a set of data because the ratios are independent:

if x was a base chip (your 630) but it's data isn't known, it does not affect the ratios of chips y and z:

z/y = (z/x) / (y/x)

so you don't need a base chip in order to normalize a set of data to weight into your existing average.

Another way of saying it, is that any low chip (x2 3800) can act as a base for set of data (for one game). Just normalize the new data so that the x2 3800 number is equal to the existing average (1.23), then the ratios for the other chips will be meaningful and ready to weight in at 1/4.

 

[HotFoot]

I suppose for now interpolation would be okay. While it is less meaningful to talk about expected performance rather than observed performance, it is not too bad in terms of performance/price discussions. It will fail to highlight points of interest such as the odd case where higher-numbered/priced chips actually perform worse than cheaper chips of the same line, but these things do, as you say, get washed out in the averaginging process. I expect to see more reviews of the Brisbane lineup as these chips become more popular.

As for the baseline chip, I have adopted the X2 3800+, so this should no longer be an issue. The P4 630 is old news now and I don't expect review sites to continue showing results for this chip. The 3800+ should be relevant until at least 2008. If I'm still doing these graphs at that time, I'll look at changing chips again yearly.

I don't think I can justify interpolation for the FX-60, since this is a different platform. I would like to keep this chip on the charts for another 6 months or so, so I would like to find Oblivion benchmarks for it.

 

[halbhh]

Perhaps you could make a conservative fudge factor for the performance penalty those odd memory speeds in the Brisbanes cause, like an average penalty amount over a few games vs the interpolated theoretical fps, from what benches do exist, if any. I imagine any particular fudge factor would be specific to one game only, but if you found any consistency across games, then you'd have an easy fudge.

 

[HotFoot]

The memory speed issue is actually addressed, which you'll see in the next update. I have a stepwise linear regression of game performance with CPU speed (as you suggested) as well as memory speed. I had tried adding L2 cache to the mix, but there wasn't enough data. As far as significance, for the Oblivion 1600x1200 benchmark, I found a 25 fps difference over the CPU frequency range, and a 5 fps difference between the best memory speed (400) and worst memory speed (350). All the data will be available in the spreadsheet in the .zip folder on the next update.

 

[halbhh]

8)

 

[HotFoot]

The graphs have been updated.

Major changes this week are that the X2 3800+ processor is now the baseline used for normalisation (previous versions used the P4 630) and the inclusion of Oblivion 1600x1200 benchmarks in the gaming performance index. Of special note for the latter point is that benchmarks were not found for several of the AM2 processors. Expected scores were estimated using CPU and memory speed relations within the AM2 processor family. Please take a look at the spreadsheet under "mapping" (.zip file) to see the technique used.

As for the pricing updates, the Intel processors have not moved an inch in the last week. The same goes for the majority of the AMD processors, except the 3600+ became even cheaper ($95) and the 5200+ fell $10. Availability of the FX-60 continues to dwindle, and it is no longer available even at TigerDirect. I found the lowest price using PriceGrabber at $499.99, which is $140 more than last week. I expect this processor will have to be retired from the graphs very soon.

I appreciate all the feedback I've gotten so far. Ciao!

 

[HotFoot]

The graphs and table have been updated with pricing from 2007-03-24.

 

[halbhh]

The AMD pricing has gotten so low that it's concievable to build a gaming machine as a seperate computer, and have two machines, one for on-line with security software, and the other for games only, with low system resource overhead, which allows the cheap 4200 or 4600 cpus to be great gaming choices even for the cpu intensive games.

 

[corvetteguy]

The AMD pricing has gotten so low that it's concievable to build a gaming machine as a seperate computer, and have two machines, one for on-line with security software, and the other for games only, with low system resource overhead, which allows the cheap 4200 or 4600 cpus to be great gaming choices even for the cpu intensive games.

Or not.

 

[HotFoot]

The concept of having an "unplugged" computer these days is novel, but I really wonder what the gaming penalty is for having your firewall and antivirus running in the background. On multi-core machines, this load can't be significant.

On the other hand, it's true that good CPUs are getting very cheap. I'm being enticed by the X2 3600+/AMD690G combination for my HTPC. Throw in $60 for a good PSU and another $100 for a big hard drive and I'm set. Conveniently, the cost of this project is going to be about the size of my tax refund.

 

[halbhh]

There's a good example.

I'm pretty well computerized actually. There are the two desktops for internet stuff for my wife and I, and then the laptop for around the house, and finally the dedicated writing computer which is intentionally off line.

If I had some program I needed to run constantly that took enough processing I'd throw in another, but I don't. It's amazing what you can build for a few hundred now.

 

[HotFoot]

I noticed that TG Daily has put up a 30-Mar update to the price/performance charts, but I can't get to the second page:

You are not authorized to view this resource.
You need to login.

What's up with that? How many accounts do I need to have here? I'm hoping it's just a glitch with the new website.

 

[HotFoot]

The graphs have been updated: I felt compelled since the pricing on most AMD chips has tumbled since the cuts yesterday. I have included the QX6800 from Intel as well, but since this is not yet available at Newegg, I'm just using the tray price of $1200 for now. I'm missing only the FarCry benchmark for the QX6800, so hopefully that will show up soon and the gaming performance metric will be complete.

I have retired the FX-60 part. It has become generally unavailable, and if you can find one, it's probably very expensive. When you can buy a new faster CPU/motherboard/RAM combination all for less than the price of a single FX-60, you know there's no point even considering the old chip.

It's very interesting to look at the general performance zoomed view. The AMD chips, since they've gotten so cheap, are well ahead of the Intel ones in the same price point. I know this situation will change once Intel's price cuts come into effect, but right now it's very hard to suggest Intel below $250. I look forward to seeing how the prices match up by the end of the month.