Cerb
Elite Member
- Aug 26, 2000
- 17,484
- 33
- 86
Both the Anandtech and Xbit tests have been skewed. They simply don't have better parts or information (or are still held to some NDAs). The UT2k3 score from the Xbit preview is promising. But that's all any of the gaming stuff is: promising.
With preproduction processors on preproduction motherboards running preproduction chipsets and using preproduction drivers...it's bound to be better once we really get it. In the Xbit review, the UT2k3 benchmark was the only balanced test done. It performed extremely well. But still just beating the P4 which will likely be eclipsed before the A64 release.
At the same time, they couldn't run many tests.
The servers are ready. But they aren't performance optomized.
The desktops are not, but they are, and need more tweaking (at least the chipsets do).
If anything, the reviews so far have shown how much the Athlon's power has been wasted, and definitely a good showing as to why they went with the integrated memory controller instead of an HT-like solution. Also remember these are the first cores of a new line of chips. With all the time taken and patents filed, I imagine AMD has even better stuff around the corner. And we should all know that the waiting game is a dangerous one. Better to get your foot in the door and then open it wide later than the stand there and wait until you have something miraculous.
Result: Both AMD and Intel look to be doing very well, and we appear to be getting back to the Willamette vs. Tbird days, when the differences in benchmarks were commonly 20% or more between where they suck and where they excel. This is likely a good thing for AMD as far as servers go, because most of the time boxes will have a single task or at most very few tasks.
With preproduction processors on preproduction motherboards running preproduction chipsets and using preproduction drivers...it's bound to be better once we really get it. In the Xbit review, the UT2k3 benchmark was the only balanced test done. It performed extremely well. But still just beating the P4 which will likely be eclipsed before the A64 release.
At the same time, they couldn't run many tests.
The servers are ready. But they aren't performance optomized.
The desktops are not, but they are, and need more tweaking (at least the chipsets do).
If anything, the reviews so far have shown how much the Athlon's power has been wasted, and definitely a good showing as to why they went with the integrated memory controller instead of an HT-like solution. Also remember these are the first cores of a new line of chips. With all the time taken and patents filed, I imagine AMD has even better stuff around the corner. And we should all know that the waiting game is a dangerous one. Better to get your foot in the door and then open it wide later than the stand there and wait until you have something miraculous.
Result: Both AMD and Intel look to be doing very well, and we appear to be getting back to the Willamette vs. Tbird days, when the differences in benchmarks were commonly 20% or more between where they suck and where they excel. This is likely a good thing for AMD as far as servers go, because most of the time boxes will have a single task or at most very few tasks.
True....BUt...What we need to be asking ourselves is will this be enough??? I mean this was using the opteron which has the 1mb of l2 cache and the dual channel intergrated mem controller (I know Anandtech stated the programs used are low bandwidth intensive and low cache intensive, but some of those same benches showed gains when northwood arrived versus the willamette and that was based on l2 cache enhancements). The athlon 64 will be less according to AMD's initial reports. Then throw in the fact the Prescott will arrive in the 3.2 to 3.4ghz with the 1mb l2 cache, more l1 cache, advanced HT and other instruction sets, and an already stable platform. I think it has been stated that hammer (with all the past hype) needs to not only just continue status quo like the bartons and keep close to P4 it needs to LAY DOWN THE HAMMER....I mean the p4 is supposedly at the .09micorn process supposed to scale quite high...Will AMD get same scalabilty??? Have their yield issues truly been solved and allow them to ramp way high???
Cerb
Elite Member
- Aug 26, 2000
- 17,484
- 33
- 86
Scalability they have. I think the Bartons have proven that for the standard K7. Many, including me, thought the 2400+ was the end of the rope for the Socket A. Well...it wasn't. I imagine both K8 series have plenty of room to grow.
The key is how fast they can get those speeds out. 2GHz around september probably won't be enough unless Intel stops at 3.4GHz this year. With luck either they will get higher, or the Opterons, which we know will be that high, will be released (the 100 series) for the NVidia platform, and they will have a high-end competitor there. I feel sure they will be able to keep up, but exactly by what degree has yet to be seen. Better yet will be the said 100 series Opterons and/or the high-end A64s being cheaper than competetive P4 parts...but given it's all new stuff, that's a bit iffy.
The key is how fast they can get those speeds out. 2GHz around september probably won't be enough unless Intel stops at 3.4GHz this year. With luck either they will get higher, or the Opterons, which we know will be that high, will be released (the 100 series) for the NVidia platform, and they will have a high-end competitor there. I feel sure they will be able to keep up, but exactly by what degree has yet to be seen. Better yet will be the said 100 series Opterons and/or the high-end A64s being cheaper than competetive P4 parts...but given it's all new stuff, that's a bit iffy.
Well they shifted to the .13micron process and it bought the K7 more life...no surprise there...Intels .13 saved the willamette which appeared to start hitting the wall at 2ghz and apparently can go to mid 3ghz range, but intel will switch to .09 process to gain more life to near 5ghz. AMD is still on .13 right???
Remember the scalability issue of recent has been more about the problems associated with implementation of the SOI process which was supposed to buy AMD plenty of room to ramp speed up quite rapidly. Thisis one of the concerns many have shown in other threads about timeliness and already K8 delays. Many thought they would have been to 2ghz range nearly 9months ago which they were not. Who knows they solve the SOI process and maybe they ramp to the moon, but to say theu have plenty of room could be too early to tell. They obviously run at lower speeds then Bartons....
I agree I think AMD will always find a way to keep up, but what I am saying and have so repeatedly that with all the hammer hype starting even as far back as Q4 2001 when I still ran athlon the talk was about annihilation and technology that would overwhelm the P4. Now we are talking about even keel stuff. I have to say that the delays have really hurt the overwhelming success this processor could have had. Where is the shock and awe??? I am not as impressed.
Also remember in the past when the p4 was doing poorly ppl stated wait for optimization...well it ended up happening with someprograms implementing SSE2, but for the most part ramp up speed and enhancements in cache was the real savior for the P4. The SSE2 should be their for AMD consideing P4 has already laid that ground work. I am wondering aside from platform enhancements where is the cpu enhancements going to come from and what % is likely to result from whatever is there to optimize (at the cpu level)??? I just want ppl to be realistic about how much will the athlon64 look intially at launch from what we see now at same mhz level, even if they may succeed in ramping it to 2-2.2ghz levels by then...
Lets review the score of the opteron in desktop performance as seen in anandtech's review...(including gaming scores eventhough there may be an issue with use of pci card in a pci-x slot versus not on the other platforms for the opteron...remembering unlikely the athlon64 desktop platforms will have this, and that agp interface may not show same leads)
P4 3.0c -vs- Opteron 244 (1.8ghz) ; Opteron 244 (1.8ghz) -vs- Barton 3000+ (2.167ghz)
<<Content Creation>> : P4 3.0c (+21.9%) ; Opteron 244 (+8.4%)
<<Business Winstone>> : opteron 244 (+19.5%) ; Opteron 244 (+2.0%)
<<UT2003 Flyby>> : opteron 244 (+7.5%) ; Opteron 244 (+11.6%)
<<UT2003 Botmatch>> : opteron 244 (+9.9%) ; Opteron 244 (+19.6%)
<<Quake 3 Arena>> : opteron 244 (+7.5%) ; Opteron 244 (+44.1%)
<<Jedi Knight>> : oPteron 244 (+12.7%); Opteron 244 (+86.9%)
<<Commanche>> : opteron 244 (+7.6%) ; Opteron 244 (+12.8%)
<<Divx/Xmpeg4.5>> : P4 3.0c (+48.8%) ; Barton 3000+ (+1.4%)
<<WMP 9 encoding>> : P4 3.0c (+41.6%) ; Opteron 244 (+4.5%)
<<3DMax5 (Test 1) >> : P4 3.0c (+65.4%) ; Barton 3000+ (+13.2%)
<<3DMax5 (Test 2) >> : P4 3.0c (+39.5%) ; Barton 3000+ (+12.2%)
<<3DMax5 (Test 3) >> : P4 3.0c (+27.8%) ; Barton 3000+ (+4.5%)
<<3DMax5 (Test 4) >> : P4 3.0c (+56.2%) ; Barton 3000+ (+13.6%)
<<3DMax5 (Test 5) >> : P4 3.0c (+73.3%) ; Barton 3000+ (+23.8%)
........................................: P4 (avg. +52.4%) ; Barton (avg. +13.5%)
<<Lightwave (Test 1)>> : P4 3.0c (+9.3%) ; Opteron 244 (+24.0%) ***SSE2***
<<Lightwave (Test 2)>> : P4 3.0c (+3.5%) ; Barton 3000+ (+22.7%) ***SSE2***
<<Lightwave (Test 3)>> : P4 3.0c (+27.3%) ; Opteron 244 (+47.2%) ***SSE2***
........................................: P4 (avg. +13.4%) ; Opteron (avg. +16.2%)
Gaming: Opteron 244 (avg. +9.0%) ; Opteron 244 (avg +35.0%)
Multimedia/ Encoding: P4 3.0c (avg. +45.2%); Opteron 244 (avg. +1.6%)
3d rendering(using avg's): P4 3.0c (avg +32.9%); Opteron 244 (avg. +1/4%)
Now how much will the smaller cache (inintialy) and single memory controller Athlon64 do less then the OPteron at same speed??? That is the equestion, and think of it as 2 ways; 1 currently with same optimizations of opteron, 2 and with a reasonable amount of ptimization and platform tweaking....
trust me ppl we can gather observation and analysis from this data...
Why would I need a link when YOU YOURSELF said as well that AMD brough in IBM to help them out with SOI-problems?
What YOU don't get is that the tested Athlon 64 was far from being the final product! If it were, they would be releasing it right now. But they are not. You are whining about Athlon 64 and you base your opinions on benchmarks that were done on beta-hardware and beta-drivers and on benchmarks that were done on server-platforms. You just don't have any real evidence to suggest hos well Athlon 64 is going to perform.
The products and platforms that are being tested are SERVER-products! God, they are testing 1U-rackmountable servers! They are NOT designed for desktop-applications! In server-tests, Opteron did really well, and those are the things the tested hardware is meant for!
Like I have said repeatedly, there will be two version of Athlon 64: one with 256kb of L2 and other with 1meg of L2. Of course, Athlon 64 will also have higher clock-speed (around 2GHz and up, whereas Opteron is 1.4-1.8GHz), propably a new revision of the mem-controller (lower latencies) and desktop-tweaked platforms/chipsets. Of course, the single-channel mem-controller is a drawback, but OTOH, Athlon 64 will be running of faster RAM, which partly compensates for that.
Why don't you just wait 'till they release the goddam chip before bitching and moaning about it's performance, OK?
No one is doing any of that...Seems like discussing this in a hardware thread must have been wrong, huh????roll;
I guess we wont be able to discuss when they officially launch either cause it will be new and not many manufacturers will have mobos out...Right???: roll;
I guess maybe sometime next year like Q2 2004 we can discuss it.......damn!!!
I don't need a link to the IBM thing as I have read that...What I haven't read is whether or not they have worked out all of their issues...
Was I OK (with you) in reviewing the desktop/workstation numbers on the OPteron??? I mean it is out for purchase so I can review it, right???
I guess we wont be seeing YOU much in this thread since YOU have nothing to offer to this discussion...We will see YOU later this year!!!! Bye Now
I was posting a reply yesterday when the power here went out. Came back two hours later and I had a blown up power supply. Then I went out partying and forgot about it ... literally.
Well, today I read over at Ace's the follow up to the thread and understood were the lower latencies came from. Just a correction:
The P4 scored 65.8fps. It was the Athlon XP that scored 39.7.
Kuk
The current Opteron doesn't support DDR400. You can use the memory but its running it at DDR333 speed. You can be pretty sure that by the time Athlon64 comes out, DDR400 is gonna be supported. I could see that accounting for maybe a 10% difference in heavily bandwidth dependent stuff like video encoding, but not the size gap that we saw in that review. And going back to 72-bit rather than 144-bit bus on Athlon64 will much more than counteract the DDR400 advantage.
The idea of a GeForce 440MX card supporting 66MHz PCI is absurd. Such features would never be considered in a budget card (anything with the MX label).
I would wager that by the time the Athlon 64 arrives, there will be a new core revision with an additional FPU pipeline. In webservers & DB servers like the Opteron is intended for, the FPU never gets used, so it's lackluster FPU performance isn't that big a deal. For scientific and graphics workstation apps it matters, and of course for games where the Athlon64 will be used it matters bigtime.
glug,
Some of the Geforce PCI boards for the Mac supported 66Mhz PCI, and these were budget boards.
But once again, the point is that the PCI implementation on the Opteron differs dramatically from that in Canterwood and nForce2.
Some of the Geforce PCI boards for the Mac supported 66Mhz PCI, and these were budget boards.
But once again, the point is that the PCI implementation on the Opteron differs dramatically from that in Canterwood and nForce2.
There is no such thing as a PCI connected thru southbridge on ANY PC system.
Southbridge is connected to Northbridge thru a psuedo-PCI channel (usually). PCI connects to the CPU through Northbridge, which in the case of the Hammer happens to connect to the CPU through HT rather than through EV6 (Athlon) or Netburst (P4) protocols.
There is no Hammer advantage there unless you are just saying that HT is better than Netburst in general.
Southbridge is connected to Northbridge thru a psuedo-PCI channel (usually). PCI connects to the CPU through Northbridge, which in the case of the Hammer happens to connect to the CPU through HT rather than through EV6 (Athlon) or Netburst (P4) protocols.
There is no Hammer advantage there unless you are just saying that HT is better than Netburst in general.
glug,
I think you are confused. You don't get PCI on an Intel Canterwood without ICH5. Look at Intel's web site; the PCI controller is provided by ICH5 (southbridge). Look at the layout diagrams on the Intel developer web site, which clearly show the PCI bus connected to ICH5, not the northbridge. Similarly, PCI support on the nForce is provided by the MCP, which is their term for the southbridge. In contrast, AGP support is on Intel's MCH (Canterwood northbridge) and the nForce2 IGP (northbridge).
On Opteron, you get a dedicated Hypertransport link for PCI-X direct to the processor. The nForce3 for Opteron will consist of only a single chip that uses a 3.6GB/sec link (Hypertransport) to AGP and PCI. Because of this, nVidia claims "the single-chip 64-bit NVIDIA nForce3 Pro architecture provides an inherent performance advantage compared to dual-chip implementations of the same functionality. In addition to overall latency reductions, the NVIDIA nForce3 Pro significantly boosts device throughput. "
But the real issue here is that PCI doesn't provide enough bandwidth for geometry, so the cpu must go back constantly for this information. As a result, the latency differences in the PCI implementations becomes a major impact on performance, whereas it would not be with AGP. I have a message from Epic to this effect, which I plan to post after a follow up.
I think you are confused. You don't get PCI on an Intel Canterwood without ICH5. Look at Intel's web site; the PCI controller is provided by ICH5 (southbridge). Look at the layout diagrams on the Intel developer web site, which clearly show the PCI bus connected to ICH5, not the northbridge. Similarly, PCI support on the nForce is provided by the MCP, which is their term for the southbridge. In contrast, AGP support is on Intel's MCH (Canterwood northbridge) and the nForce2 IGP (northbridge).
On Opteron, you get a dedicated Hypertransport link for PCI-X direct to the processor. The nForce3 for Opteron will consist of only a single chip that uses a 3.6GB/sec link (Hypertransport) to AGP and PCI. Because of this, nVidia claims "the single-chip 64-bit NVIDIA nForce3 Pro architecture provides an inherent performance advantage compared to dual-chip implementations of the same functionality. In addition to overall latency reductions, the NVIDIA nForce3 Pro significantly boosts device throughput. "
But the real issue here is that PCI doesn't provide enough bandwidth for geometry, so the cpu must go back constantly for this information. As a result, the latency differences in the PCI implementations becomes a major impact on performance, whereas it would not be with AGP. I have a message from Epic to this effect, which I plan to post after a follow up.
mechBgon
Super Moderator<br>Elite Member
- Oct 31, 1999
- 30,699
- 1
- 0
KenAF, in the first place, has it been confirmed that the video card was in a PCI-X slot rather than in a PCI slot? Yes or no?
Secondly, if the card was in a PCI-X slot on the Opteron board, and if it benefitted from a low-latency connection as a result, you say "no fair." But an AGP card would be connected very similarly, by way of the AGP tunnel, wouldn't it? Would it still be "no fair" if it got higher framerates than Canterwood in an AGP-to-AGP comparison as a result of another low-latency connection? That seems like a backwards viewpoint... "no fair, it works better because of its superior design." (and remember, PCI-X video cards are on the horizon, for what that may be worth)
Maybe you should be satisfied that your objection has been duly noted by everyone who's interested, and wait for nForce3 Pro so we can see what the facts are. Persisting on your holy quest to discredit one gaming benchmark is starting to look like tunnel vision on your part.
Secondly, if the card was in a PCI-X slot on the Opteron board, and if it benefitted from a low-latency connection as a result, you say "no fair." But an AGP card would be connected very similarly, by way of the AGP tunnel, wouldn't it? Would it still be "no fair" if it got higher framerates than Canterwood in an AGP-to-AGP comparison as a result of another low-latency connection? That seems like a backwards viewpoint... "no fair, it works better because of its superior design."
(and remember, PCI-X video cards are on the horizon, for what that may be worth)Maybe you should be satisfied that your objection has been duly noted by everyone who's interested, and wait for nForce3 Pro so we can see what the facts are.
Persisting on your holy quest to discredit one gaming benchmark is starting to look like tunnel vision on your part.mech,
The problem is that PCI doesn't have enough bandwidth for the geometry, so the processor ends up going back constantly for this information, which would not happen with AGP. Thus, PCI latency becomes a key factor to performance.
Good question. My inquiries with Anandtech have been ignored, so I don't have an answer. But I am not entirely clear on how standard PCI support is provided on the Opteron, and whether or not they have the same (or some degree of) latency advantage as the PCI-X interface.
No. The latency difference isn't an issue with AGP; see the remarks from Johan @ Aceshardware above.
The problem is that PCI doesn't have enough bandwidth for the geometry, so the processor ends up going back constantly for this information, which would not happen with AGP. Thus, PCI latency becomes a key factor to performance.
All of the gaming benchmarks, not just one. The actual developers of these games don't have trouble criticizing Anand's conclusions, so I suspect Anand may have received emails from them as well in regard to the review.
Anand's hardware has so far ignored my inquiries, and even deleted my posts (possibly by accident). I want to see them respond or update the review. If appropriate, I want to see them note the latency issue in the review. So I won't be "satisfied" until they do...or post the Nforce3 tests.
mechBgon
Super Moderator<br>Elite Member
- Oct 31, 1999
- 30,699
- 1
- 0
Summary for those not inclined to read long argumentative stuff: Games run well on powerful CPUs/platforms. Opteron is a powerful CPU/platform. Strangely enough, an Opteron system does pretty well in gaming benchmarks. Weird, huh? :Q[/i]
Ok, KenAF, so you just gotta argue. Well, I'll be your punching bag for a couple minutes, then I had another look at the two pages of gaming benchmarks.
UT2003: Of the two UT2003 benchies, which one's known for being the video-card-limited one, is it Flyby or Botmatch? It's Flyby, isn't it? And what do we see in Flyby? That the results are all pretty close to the 80fps mark, because... gee, the video card probably is maxed most of the time during Flyby, since all the CPU has to do is move the point of view through the map as fast as the video card can draw frames. Right? So if the PCI-X bus was used, and was such a big advantage, where is the massive difference here?
Now look at Botmatch, where the CPU is handling a lot of physics and AI. The P4 3.0C is winning by 10% over the XP3000+, just like it did in Anand's review of the P4 3.0C before. Are you surprised that the Opteron, which is clocked a bit lower than the 3000+ but armed with SSE2, twice the L2 cache and an advanced low-latency memory controller, is performing better than a 3000+ (and a 3.0C)? Why would you be surprised at this result? I'm certainly not. One look at the 1.7GHz Celeron's poor Botmatch results is enough to confirm that CPU/platform power has a LOT to do with your botmatch fps... look at this graph. Don't tell me that the huge, huge differences in that graph are all due to differing AGP-port latency. CPU power can make for very large performance gains in Botmatch.
Comanche 4: The above goes for Comanche 4 as well. Flight sim framerates are well-known for reflecting the CPU/platform's FPU power, not the video card's capabilities. Ask at any flight-sim forum if you need to hear it from them. And there are Comanche 4 results at the same page I linked just now that clearly illustrate that a powerful CPU can make for very large differences in Comanche 4 benchmarks.
Quake 3 Arena: The main anomoly in the whole thing has nothing to do with Opteron, per se... it's the low performance of the 3000+ in Quake 3 Arena that's odd, but you can hardly write that up to the Opteron's PCI-X bus on a game that came out when a 32Mb video card was cutting-edge. Simple geometry, simple textures, simple everything. If anything, I'd be inclined to write it up to it being OpenGL instead of Direct3D... my nForce 220D board and nForce2 board (R.I.P.) like Direct3D better than OpenGL in trueSpace 4.3, for instance. An alpha Hammer 800MHz performed on par with a 1.6GHz Williamette in Quake 3 Arena about a year ago, and it's hardly a surprise if a 1.8GHz production Opteron slightly outperforms a P4 3.0C today.
If you have counter-arguments to make, please don't start them all with "Johan at Aces Hardware says..." If you are still shocked and discontent that an advanced 1.8GHz CPU can run with a P4 3.0C, and want to blame the possible PCI advantage, well heck, I'm not saying it's impossible for you (and Johan) to be correct. But I don't think Anand misrepresented anything, with wording like
Ok, KenAF, so you just gotta argue. Well, I'll be your punching bag for a couple minutes, then
I had another look at the two pages of gaming benchmarks.UT2003: Of the two UT2003 benchies, which one's known for being the video-card-limited one, is it Flyby or Botmatch? It's Flyby, isn't it? And what do we see in Flyby? That the results are all pretty close to the 80fps mark, because... gee, the video card probably is maxed most of the time during Flyby, since all the CPU has to do is move the point of view through the map as fast as the video card can draw frames. Right? So if the PCI-X bus was used, and was such a big advantage, where is the massive difference here?
Now look at Botmatch, where the CPU is handling a lot of physics and AI. The P4 3.0C is winning by 10% over the XP3000+, just like it did in Anand's review of the P4 3.0C before. Are you surprised that the Opteron, which is clocked a bit lower than the 3000+ but armed with SSE2, twice the L2 cache and an advanced low-latency memory controller, is performing better than a 3000+ (and a 3.0C)? Why would you be surprised at this result? I'm certainly not. One look at the 1.7GHz Celeron's poor Botmatch results is enough to confirm that CPU/platform power has a LOT to do with your botmatch fps... look at this graph. Don't tell me that the huge, huge differences in that graph are all due to differing AGP-port latency. CPU power can make for very large performance gains in Botmatch.
Comanche 4: The above goes for Comanche 4 as well. Flight sim framerates are well-known for reflecting the CPU/platform's FPU power, not the video card's capabilities. Ask at any flight-sim forum if you need to hear it from them. And there are Comanche 4 results at the same page I linked just now that clearly illustrate that a powerful CPU can make for very large differences in Comanche 4 benchmarks.
Quake 3 Arena: The main anomoly in the whole thing has nothing to do with Opteron, per se... it's the low performance of the 3000+ in Quake 3 Arena that's odd, but you can hardly write that up to the Opteron's PCI-X bus on a game that came out when a 32Mb video card was cutting-edge. Simple geometry, simple textures, simple everything. If anything, I'd be inclined to write it up to it being OpenGL instead of Direct3D... my nForce 220D board and nForce2 board (R.I.P.) like Direct3D better than OpenGL in trueSpace 4.3, for instance. An alpha Hammer 800MHz performed on par with a 1.6GHz Williamette in Quake 3 Arena about a year ago, and it's hardly a surprise if a 1.8GHz production Opteron slightly outperforms a P4 3.0C today.
If you have counter-arguments to make, please don't start them all with "Johan at Aces Hardware says..."
If you are still shocked and discontent that an advanced 1.8GHz CPU can run with a P4 3.0C, and want to blame the possible PCI advantage, well heck, I'm not saying it's impossible for you (and Johan) to be correct. But I don't think Anand misrepresented anything, with wording like
He takes the trouble to try to give us a preview of gaming performance, and gets a kick in the teeth for his efforts. Give it a rest, maybe?
That's the thing, it's not about "fair" or not, but rather whether it's a good indication of gaming performance. Most people will be using AGP video cards. The point of the preview was to use similarly configured graphics subsystems to give you an indication of what the comparative performance when using an AGP card would be. Since in this case, the PCI implementations do not comparably match the AGP implementations, this is a questionable indication of performance when using an AGP video card. It's certainly good information for those with PCI video cards, but how many people are there who uses that?
Ok, Aces hardware has posted actual benchmarks to prove what we have been saying.
Unreal Tournament 2003 allows the use of a "Null renderer" (software renderer) which has been recommended by Tim Sweeny of Epic Megagames for Opteron testing. This means that no data will be sent to the videocard--the cpu does 100% of the work without involving PCI transfers or differences in PCI implementations.
As you can see, the results are quite different than reported by Anand when testing over the PCI bus. The Opteron 1.8 and Athlon XP 3000+ perform about the same in Unreal Tournament 2003, both significantly outperforming the XP 2200+. The P4 3.0 outperforms the Opteron 1.8 by about 10% in the game.
Nothing beats hard, cold proof, ehh?
Unreal Tournament 2003 allows the use of a "Null renderer" (software renderer) which has been recommended by Tim Sweeny of Epic Megagames for Opteron testing. This means that no data will be sent to the videocard--the cpu does 100% of the work without involving PCI transfers or differences in PCI implementations.
As you can see, the results are quite different than reported by Anand when testing over the PCI bus. The Opteron 1.8 and Athlon XP 3000+ perform about the same in Unreal Tournament 2003, both significantly outperforming the XP 2200+. The P4 3.0 outperforms the Opteron 1.8 by about 10% in the game.
Nothing beats hard, cold proof, ehh?
By that reasoning... If you plan on using a PCI GFmx, then the Opteron "is just the ticket".Originally posted by: mechBgon
If you want to play UT2003 with no video output, then it looks like a P4 3.0C is just the ticket, yep!
Ken's point is a very valid one. With Ace's followups, it's apparent that the PCI bus is the bottleneck of Anand's (and others') tests, and the gaming tests aren't the direct results of the cpu. The fact that the different speed XP's had the same results was the red flag.
mechBgon
Super Moderator<br>Elite Member
- Oct 31, 1999
- 30,699
- 1
- 0
Are you sure?Originally posted by: Wingznut
By that reasoning... If you plan on using a PCI GFmx, then the Opteron "is just the ticket".Originally posted by: mechBgon
If you want to play UT2003 with no video output, then it looks like a P4 3.0C is just the ticket, yep!
Ken's point is a very valid one. With Ace's followups, it's apparent that the PCI bus is the bottleneck of Anand's (and others') tests, and the gaming tests aren't the direct results of the cpu. The fact that the different speed XP's had the same results was the red flag.
It seems to me that, if you're trying to gauge real-world gaming performance, taking the Direct3D and OpenGL APIs out of the picture is not the right answer, because it eliminates the PCI-bus unknowns but introduces others. What optomizations did AMD make for D3D and OpenGL that are going to waste? How about 3DNow, something that the Opteron loses while the P4 never had it? Ah so... I'm not completely discounting KenAF's and Johan's concern, but the cure is worse than the disease, when you start proclaiming that the true gaming performance is illustrated by... software rendering?
We'll see when nForce3 or K8T400M boards arrive. Edit: I invite the skeptics to look again at XBit Labs' preview of the Athlon 64 at 1.6GHz, using an AGP Radeon 9700 Pro. That should help eliminate the PCI question without tossing hardware-accelerated 3D rendering out the window in the process. For the lazy, look here for the test setup and here for the UT2003 results. And remember... that's just 32-bit
Note I made a few corrections to this post...
From what I understand, all the normal Direct3D and OpenGL operations are still used. With a null renderer, you just don't get anything on the screen. Thus, the null driver delivers the performance you would expect at 1280x1024 if your video card was infinitely fast, and only limited by your processor's performance. That's not exactly the case, because modern video cards are able off-load certain tasks (geometry?) to the video card, that are probably performed by the processor in case of the null renderer. Admittedly, I am still not fully clear on what the null renderer does.
Johan notes in his test that he expects the Athlon64 to do better in Unreal Tournament 2003, even without the extra memory channel, because of the improved latency offered by DDR400 memory, rather than the dual channel DDR333 on the Opteron that also uses ECC and less aggressive memory timings. Unreal Tournament 2003 is a special case in this regard, because it benefits more from improved latency than extra bandwidth, as per Tim Sweeney.
This appears to be born out in the Xbit Labs review, which shows a 1.6GHz Athlon64 with DDR400 turning in Unreal 2003 results comparable to the P4 3.00/800 with the Radeon 9700 Pro (note the A64 does enjoy a slightly better configuration than the setup in the separate P4 3.00 review). It will be interesting to see whether the nForce3 boards coming for Opteron permit overclocking of memory to DDR400 speed, and whether they also support more aggressive memory timings. If so, we could see a decent boost for Opteron on nForce, even for null renderer tests (which would be pointless once we get AGP, of course).
mechBgon
Super Moderator<br>Elite Member
- Oct 31, 1999
- 30,699
- 1
- 0
I would like to point out that the K8T400M chipset used at XBit does have memory tweaks in the BIOS (see page 4). And that chipset supports Opteron, not just Athlon 64. I can't envision an nF3 board without tweaks either. Ergo, Opteron boards with tweakable memory settings are going to happen. If Athlon 64 is enjoying some of its performance lead at XBit due to the tweakability of a workstation-class board, well, Opteron probably will too. Granted, Opteron's on-die memory controller mandates Registered ECC memory, and that's going to incur a little performance hit.
Also remember that Tim Sweeny is predicting about a 15% boost from 64-bit UT2003 versus 32-bit, so looking forward, we've got to keep that in mind for future games. Anyone want a 15% boost in Doom III, raise your paw
TRENDING THREADS
-
Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)- Started by DisEnchantment
- Replies: 25K
-
Discussion Intel Meteor, Arrow, Lunar & Panther Lakes + WCL Discussion Threads
- Started by Tigerick
- Replies: 23K
-
Discussion Intel current and future Lakes & Rapids thread- Started by TheF34RChannel
- Replies: 23K
-
-
Facebook
Twitter