3DMark03 anomalies? No surprises there at all

Mike

Article is really interesting if you compare the results of each of the three P4 processors against each other on a test by test basis.

In a perfect world the P4 800 MHz bus version should be on average 3 per cent faster than the standard P4 (because of the faster memory) and the P4 large L3 would be 3 percent faster than the P4 800 MHz (if you check the memory intensive graph results at the end of the test you see this).

What is interesting in the individual tests (where they run entirely in a 512 kb cache) is when the P4-8 time equals the P4-L3 but not the P4-5 and where the P4-8 is faster than either the P4-L3 or the P4-5. A confirming test is that the P4-8 numbers should be 2 per cent slower than the P4-L3 or the P4-5 in cache resident tests because of the difference in clock speeds (3066 vs. 3000).

What this comparison tells, in that there is 15 per cent discrepancies in some of these tests, is that the P4-L3 has new optimizations/errata fixes incorporated in its cache unit but does not have the same improvements/errata fixes in its bus interface unit (the P4-8 is actually a later mask and has all of the accumulated improvements). It also shows that the fixes to the P4-8 are far more significant than the increase in the bus speed. Being as the Prescott will further "improve" the P4 (correct the current bottlenecks, "fix" the original mistakes in design), we can look on a fairly spectacular increase in performance from this chip, especially in a multi-tasking situation.

Name supplied

Just wonder what all the who-ha is about.

Lately, the Opteron is getting a lot of press, which I know is good seeing that the server market generates a higher price point, however, unless I'm wrong, all indications that the AMD64 will be under 2.0 ghz lead me to believe it will still be a CPU that (in 32 bit mode) is on par, sometime ahead, sometimes behind, with the P4, let alone what happens once the Prescott comes out.

My question therefore is, what happened to HAMMER HAMMER HAMMER and with AMD leading the pack, pushing by INTEL and leaving them in the dark? Is there some unknown that perhaps AMD will surprise us with, i.e. high clock speed? Isn't that what SOI was about, isn't the Opteron running cool to the touch?

Just wanting to know what the deal is, and hope to see some story about it. HAMMER HAMMER HAMMER? Well then, what is it then, and if not, what's the problem here?

Cheers
frankii

Extra Athlon64 pin mystery solved

Mike,

The whole world+dog (shouldn't you copyright that and produce a logo?) knows that the Athlon 64 3400+ is going to be an Opteron 2.0GHz that's simply been re-badged. Reading between the lines, this has been obvious for months now. The 754-pin AMD64 will of course be the 64-bit replacement for Duron, and Athlon FX will probably never be launched, as this is just a way for AMD to re-use Bartons that didn't quite make the grade, probably due to the difficulty with manufacturing on-board cache.

I suggest this whole confusion of 940/939/754 pins is caused yet again by the spinmeisters or 'marketing' division, who couldn't decide whether to launch 940 or 754 pin Athlon 64's in time for AMD to ramp production. Now, in order to get Athlon 64 out of the door in time for it's September launch, AMD are going to have to use re-badged Opterons (the 146 or 246 at 2.0GHz). As the spinmeisters told yourselves over a few shandys a while back, the Opteron makes a rocking games machine.

The rebadging from Opteron to Athlon 64 is so as not to upset the "Corporate" sector, where AMD is desperate to be taken seriously.

I strongly suspect that the 939 pin Athlon64's which come later will be almost identical to their 940-pin brethren, but maybe with one of the extra hyper-transport links disabled (to stop people dualling them up), hence removing one of the pins. And we all know these chips are going to work in 940-pin boards.

I for one am looking forward to these new processors. Pentium has been around for just over 10 years now, and the 32-bit architecture has grown fantastically since it's launch. Just think what wonderful possibilities are heading towards us with the dawn of "64-bit processors for the common man," not least of all for the bean-counters.

Eva owes me a drink!

John Parlato

ATI promises never to paper launch products

Greetings

Fascinating.

I defy them to produce this product - "The TV WONDER™ family lineup now includes the new TV WONDER™ PRO - Remote Control Edition for $99..." - announced in April of this year. I challenge them because I'd rather like to buy it, my desire stymied by a total lack of existence.

Here.

Andrew Peppard

Microsoft installs Open Source systems in SWAT venture

I couldn't help but think that you're going to get a lot of feedback about the phrase "MS is better at fixing bugs than the Open Source community". The fact that Ballmer has the balls to say such a thing is typical MS arrogance and will surprise no one. However I might just venture to say that, following this comment, the Open Source community is indeed going to have to make an effort - to pick themselves off the floor and stop laughing their heads off.

As for Monkey Ballmer, I suggest he review the process and compare apples to apples, meaning support for the tiny, single-user individual. Corporations pay juicy contracts for MS support, and MS is therefor a bit less overbearing when a big company comes in and says there is a problem. For those accounts, MS will not brush them off like an annoying fly, MS will send a well-groomed consultant to find the problem and fix it. But for Joe User, it will start by Phone Support.

Phone support cost is through the roof, constantly delayed for minutes at a time, and when you get someone you waste more (paying) minutes getting around the basic stupid version game before finally being able to state your problem. Most of the time, when you've finished all the patching and updating that support has required, your PC is so broken that you end up formatting the drive in disgust. Conclusion One : your problem is gone (and your PC also). Conclusion Two : Phone Support has fulfilled its purpose, which is to make everything YOUR fault and avoid filing a Change Request. If, in your utmost stubbornness, you do manage to make MS acknowledge a mistake in their software, then they will probably tell you that it will be fixed in the next version. Conclusion One : you're still screwed until then, but you've made the phone company very happy. Conclusion Two : when you finally get the new version, you're screwed again because you will have to pay for it, on top of the enormous phone bill you took three months to get over.

If Joe User addresses himself to the Open Source community, he will have only one problem : finding the right forum. The cost of that depends only his internet connection type (dial-up, cable, ADSL, . . .) and the speed of his research. Once the forum is found, Joe User needs only to post his report and wait for an email answer. He will get one, most often in less than a day. If additional info is required, it will be handled by email. The patch will be most likely ready within a week (I'm being generous here).

After reviewing the two processes, I'm pretty sure that the Open Source way is far ahead in the support game. MS does not stand a chance in hell to get such a level of support quality - not until they've hired every single Indian developer there ever will be, that is.

Pascal Monett

Hi,

Please note, that the Quad Xeon-MP chipset uses quad-channel memory, thus giving a peak bandwidth of 6.4Gbytes/sec. Obviously, you'll need to install memory in sets of four.

Despite bandwidth issues, the on-board memory-controller of the Opteron typically offers much lower latency, but in a NUMA-style MP configuration, part of that lower latency is "eaten" by remote-memory accesses via hypertransport ( e.g.: CPU 0 accesses memory via CPU 1 ). Any OS needs to be aware, that the total memory map is contructed in a NUMA-manner, with all inherent drawbacks:

- process migration can be extremely costly and/or result in a performance penalty
- scheduler needs to be NUMA-aware
- the VM subsystem needs to pay special attention Currently, AFAIK only the latest linux-kernels ( 2.5.7x and greater ) are aware of these issues.

Oh well, it's me getting off-topic again .... ;-)

Regards,

Thomas Weyergraf

Does our government have plans for a PC Licence?

Mageeky,

You may be on to something here...

What if we tax people $100 for each stupid PC tech question posted online? Then we could tax people $250 for each ignorant rank like how it's OK to steal music. That would rake in millions the first week... Then we could tax people $10,000 for each illegal music or software download, etc. the potential is endless !!! And the really fun part would be that with all PCs registered we could give dumb people TWO official warnings to pay their taxes NOW or we turn their hard drive into burnt toast! I realize this concept isn't totally my own but it would be so much fun watching billowing smoke pour out of peoples houses and offices as their HD turned to toast, it may be worth the effort to license all PCs.

Whaddaya think???

Randy

Scaling for Opterons and for Xeons

Mike

I just read the bit about Opteron and Intel scaling. I have done some controlled tests that indicate all chipsets are not created equal. Take a look at:

this

and

this.

I have plans to look at other chipsets as well.

Doug Eadline

Logitech Elite Keyboard requires brain the size of a planet

Mike

Must say, it's not often that I get to let out a cynical guffaw, but you [Charlie D.] did it with the Logitech keyboard review.

Danm, I knew that I did need to turn on that pesky LCD screen on my desk for some reason!!

Best

Dave

Big anomalies found in 3DMark03

When 3DMark03 was first released Futuremark/Mad Onion was very clear that they were doing as much as they could to isolate the video card in this benchmark. They were moving away from the 3DMark2001SE model which was becoming a whole system benchmark towards a video card only benchmark.

This is not an anomaly, it's just someone who wasn't paying attention coming up with the astounding discovery that 3DMark does what it was supposed to do, namely test the video card rather than the entire system, pretty well.

Thanks.

Dave Robinson

Inquirer,

Hi, I read the Inquirer all the time. I saw an article about a 350Mhz Pentium doing just as good as a 1.4Ghz in 3DMark2003, and I knew the inquirer would pick it up.

There is really nothing there. 3DMark2003 doesn't have any AI, or anything like that. That isn't too surprising, since it's just a synthetic graphic demonstration with a FPS counter stuck on it. I can't think of anything that would be required to run on the processor, most modern graphic cards have hundreds of Mhz to utilize anyway.

Like I said, I'm not surprised about it. What it means is that it's even more useless then most people thought it was. Maybe Futermark should just through in abunch of code for the CPU to process, to simulate AI? I don't know, it doesn't really matter. Let's see what Futermark says.

Adam Geldersman
Michigan, USA

3DMark03 focusing on 3D gaming performance? No shock there. There's The 9700 Pro and the 9600, while being close in marchitecture numbers are very differently-abled beasts, as their prices show.

However, 3DMark03 also includes CPU tests which render within the CPU - and the article you refer to shows that the P4 veritably tore the backside of the P2 in those tests by a factor of 50 or so - but that woudln't make for such a wild headline, so wasn't mentioned in your article... or the fact that even with the 9700 Pro, Unreal Tournament 2003, Jedi Knight II, Comanche 4 or Neverwinter Nights are all unplayably slow on the P2.

The conclusion? Shoving a top-end graphics card in a slow old PC doesn't get you great gaming performance, but a midrange card in a decent PC will get you decent gaming performance. Which is commonsense.

Yes, 3DMark03 is not perfect, and can be skewed by silly combinations of hardware (£250 graphics card in £250 machine is not realistic in anyone's book), as can all other benchmarks known to man. At the end of the day, anyone buying hardware on the basis of one benchmark deserves the dissapointment they will almost-inevitably get.

Regards
Dan Lawrance
Technical Consultant

Hello Mike,

I think this whole NV38 thing should be seen in the context of a larger picture. So what adds to it ? This is how the story goes from my point of view, but let me say, that I'm still a NVidia fanboy nonetheless.

First there is the fact, that NVidia has been known for executing their products almost perfectly in the past. Okay, there have been smaller delays but in price and performance terms they were always able to match or overshoot their competition and customer expectations. Nvidia has a well known and proven development cycle; releases of new architectures every year with a rework half a year after to speed up GPU clocks and to add features (TNT to TNT2, GeForce 256 to GeForce 2 GTS, GeForce 3 to GeForce 4 Ti).

But then came the NV30/GFFX5800 debacle. The original plan was to execute like before, since the designs were great. But that thing was horribly late. But even with this late arrival to the market its performance was annoying. Not the mention its cooling "solution". Its seems that TSMCs inability to move the platform to the new .13 process on time was the reason for all this. The result was bad yielding. As a consequence the clock specification had to be taken down to deliver a least a couple chips working, even if they were getting very (and I mean VERY) hot. Thats why the performance was unsatisfactory and that why that dreadful cooling fans were developed in a hurry.

What a nightmare this must have been to the whole NVidia management team. They had to face ATI with an immature product. Nvidias Contracting of IBM as another foundry must have been quite a pill for TSMC to swallow, as it has been NVidias exclusive from the beginning. NV35/GFFX5900 should have been the respin half a year later. The first action was to speed up its development. NV35 was to replace NV30. ("The king is dead, long live the king.").

In the meantime PR and driver development had to deal with the community and kept it in check until NV35 taped out. The whole 3Dmark2003 benchmark issue has its roots here. And has someone seen the news on Omega driver devekpment being suppressed? Could it possibly be that the current reference drivers image quality was a little bit lower to squeeze out a couple of extra fps and NVidia is afraid of anyone noticing it by comparing to drivers patched to optimal visual quality ?

Now where are we now? NV40 is still far away. And where is the respin of the high-end GeForce FX (NV30 as the basis doesn't count anymore - NV35 took its place)? Why not put in a little effort to iron out the glitches that made it in to the hurried NV35 silicon? Has some one read about the PS 2.0 problems with it that NVidia is constantly denying? By the time DX9 applications/games appear on the market, driver development should be able to do a couple of workarounds.

One of the symptoms of the whole story you see now are Nvidia's earnings going down due to higher manufacturing costs. Keeping in mind that TSMC charges for every processed wafer regardless of what its actual yield is. To deliver the planned amounts of NV30 / NV35, more wafers are being need, which drives the production costs up.

Let's hope Nvidia learns from its bad luck and is able to execute NV40 correctly. Never depend on a single foundry again. Never try a new production process on a flagship product. Never cheat on the community, your potential customers.