AMD responds to concerns of lackluster POV-Ray scores

[r0ck]

http://www.uberpulse.com/us/2007/05/amds_barcelona_lack_of_performance_still_unknown.php

The objective of our demo was to show performance scaling from our current dual-core processors to our upcoming quad-core processors within the same thermal envelope and drive home the point through a real-world demonstration that customers could expect to see 2x the performance without an increase in power consumption.

We saw an 80% increase, not 100% in performance. And that's nice that their new chips nearly double the performance "in the same thermal envelope", how about measuring the power being burned then?

At some point before our launch you can plan on us showing a demo of our parts vs. Intel's high-performance processors.

Sounds like not anytime soon. And Computex is around the corner.

One thing to note, the system we showed, while it was a 4P, it was running only 6GB of memory.

I don't think POV-Ray needs that much RAM anyways. Boot uses just 4GB and gets 4700.
http://www.bootdaily.com/index.php?option=com_content&task=view&id=460&Itemid=56&limit=1&limitstart=5

 

[BaronMatrix]

http://www.uberpulse.com/us/2007/05/amds_barcelona_lack_of_performance_still_unknown.php

The objective of our demo was to show performance scaling from our current dual-core processors to our upcoming quad-core processors within the same thermal envelope and drive home the point through a real-world demonstration that customers could expect to see 2x the performance without an increase in power consumption.

We saw an 80% increase, not 100% in performance. And that's nice that their new chips nearly double the performance "in the same thermal envelope", how about measuring the power being burned then?

At some point before our launch you can plan on us showing a demo of our parts vs. Intel's high-performance processors.

Sounds like not anytime soon. And Computex is around the corner.

One thing to note, the system we showed, while it was a 4P, it was running only 6GB of memory.

I don't think POV-Ray needs that much RAM anyways. Boot uses just 4GB and gets 4700.
http://www.bootdaily.com/index.php?option=com_content&task=view&id=460&Itemid=56&limit=1&limitstart=5

Ouch. The problem with any specuation at this point is that we don't know what speed they were running at but, a look at QuadFX at 3GHz shows it gets 1620 or so while Qx6700 gets around 1850. That means that Intel is scaling at a lot more than 100% or the test systems are that much different.

But then also, QFX is only 2 CPUs\4 cores while the Opteron system in the tests are 8 cores, so how is POVRay really scaling if 4 extra cores gets no real gain?

I'm not sure what the deal is but something is wrong.

 

[turpit]

Ouch. The problem with any specuation at this point is that we don't know what speed they were running at but, a look at QuadFX at 3GHz shows it gets 1620 or so while Qx6700 gets around 1850. That means that Intel is scaling at a lot more than 100% or the test systems are that much different.

But then also, QFX is only 2 CPUs\4 cores while the Opteron system in the tests are 8 cores, so how is POVRay really scaling if 4 extra cores gets no real gain?

I'm not sure what the deal is but something is wrong.

"That, Baron, is a lucid, intelligent, well-thought out objection."

 

[Wombat2]

At some point before our launch you can plan on us showing a demo of our parts vs. Intel's high-performance processors.

AMDs handling of this incident doesn't inspire much confidence.

The sheer lack of enthusiasm of AMD spokespersons for benchmarking K10 against Intel makes me worry for K10s final performance.

If K10 is sh*t there is no point in trying to hide it. Marketing wont hide that. AMD should have learnt that from the R600 debacle.

 

[BaronMatrix]

Ouch. The problem with any specuation at this point is that we don't know what speed they were running at but, a look at QuadFX at 3GHz shows it gets 1620 or so while Qx6700 gets around 1850. That means that Intel is scaling at a lot more than 100% or the test systems are that much different.

But then also, QFX is only 2 CPUs\4 cores while the Opteron system in the tests are 8 cores, so how is POVRay really scaling if 4 extra cores gets no real gain?

I'm not sure what the deal is but something is wrong.

"That, Baron, is a lucid, intelligent, well-thought out objection."

So its not overruled? If it is I'm wearing my "Ol Dixie" tie and tails. It's just wierd.

 

[r0ck]

Ouch. The problem with any specuation at this point is that we don't know what speed they were running at

I think the scores could be a lot better had they not used High Efficiency parts but they were trying to defend their sad scores with "we only used 6GB" and yeah.

http://www.hothardware.com/articles/Core_2_Extreme_QX6700_QuadCore_Kentsfield_Performance_Preview/?page=3
http://www.lostcircuits.com/cpu/amd_quadfx/11.shtml

Some reviews show POV-Ray scores for Kentsfield near ~2000 and some around ~2600. Different betas maybe, but AMD should know to choose the best to represent themselves.

 

[turpit]

So its not overruled? If it is I'm wearing my "Ol Dixie" tie and tails. It's just wierd.

Sorry, I should have added "sustained"

 

[r0ck]

Hello, do you have something to add?

 

[yourmothersanastronaut]

"That, Baron, is a lucid, intelligent, well-thought out objection."

:trophy: :trophy:

And Bytch, these are workstation setups, for professional graphics design, visualization and other high-power computational duties like that. In light of your recent mental breakthrough that these 16-core setups aren't for consumers:

 

[Wisecracker]

Fud . . .

Any POV-Ray bench should include the version# and whether they are running the actual render benchmark tool test (or simply timng a run on a scene or an object). They should also capture a screenshot of the message area because POV-Ray does not report their benchmark tool results on a separate screen. Like this:

This is v3.6.1b (one core!) of an Opty 270 showing 75.47 pixels per second with a total time of 1953.86 seconds.

Average PPS CPU time shown @ Boot Daily is a FUD. It's Pixels Per Second (higher the better) or Total Time (lower the better)

One additional note: The POV-Ray SMP release povwin-3.7.beta.20b has *Timed-out* as of May ist and the *so-called* tests cannot be reproduced.

What a crock . . . .

 

[r0ck]

What points are you trying to make?

Are you accusing AMD, Intel, or both of FUD because they didn't give full blown numbers and specs and screens?

http://www.xbitlabs.com/articles/cpu/display/intel-v8_11.html
X-bit got similar scores to Intel and Boot.

My initial post was responding to AMD's "we got bad score booey only 6GB" when V8s with 4GB get better. And deducing their statement of when they'll launch hard numbers, and asking them to prove 80%, not 100%, performance gain with 0 increase in watts.

 

[BaronMatrix]

Ouch. The problem with any specuation at this point is that we don't know what speed they were running at

I think the scores could be a lot better had they not used High Efficiency parts but they were trying to defend their sad scores with "we only used 6GB" and yeah.

http://www.hothardware.com/articles/Core_2_Extreme_QX6700_QuadCore_Kentsfield_Performance_Preview/?page=3
http://www.lostcircuits.com/cpu/amd_quadfx/11.shtml

Some reviews show POV-Ray scores for Kentsfield near ~2000 and some around ~2600. Different betas maybe, but AMD should know to choose the best to represent themselves.

Even worse, the AMD demo used 82xx and the LostCircuits used 280 and the 280 system is just slightly behind the 82xx system. How could that kind of scaling happen?
Barcelona has L3 which lowers latency and widens paths everywhere so there is no "theoretical" bottleneck.
A 4P AMD system is the fastest out there for servers. It always scales around 80-100% above dual so why is the 4P 82xx system only getting 2000?

 

[r0ck]

http://en.wikipedia.org/wiki/List_of_AMD_Opteron_microprocessors#Opteron_8000-series_.22Santa_Rosa.22_.28high-efficiency.2C_90_nm.29
Slowest Socket F 8 socket chip is 2GHz. So...
1) AMD underclocked them.
2) Bad scaling for 2 to 4 sockets. There are no Intel 2 to 4 socket scores to compare to though so "bad" is subjective.

http://siliconvalleysleuth.co.uk/2006/10/intel_shines_on.html
It's not a core scaling problem since here it gets a 17x speedup.

 

[BaronMatrix]

http://en.wikipedia.org/wiki/List_of_AMD_Opteron_microprocessors#Opteron_8000-series_.22Santa_Rosa.22_.28high-efficiency.2C_90_nm.29
Slowest Socket F 8 socket chip is 2GHz. So...
1) AMD underclocked them.
2) Bad scaling for 2 to 4 sockets. There are no Intel 2 to 4 socket scores to compare to though so "bad" is subjective.

http://siliconvalleysleuth.co.uk/2006/10/intel_shines_on.html
It's not a core scaling problem since here it gets a 17x speedup.

Hey, I beleive that it should scale better. I just can''t understand why there is seemingly NO scaling from 2 to 4 cores. It definitely deserves investigation.

 

[bfellow]

perhaps they were accidentally used a bad batch of Barcelona chips without the latest revisions?

 

[Viperabyss]

Hello, do you have something to add?

yeh. you poor boys need to get lives.

 

[m25]

Whatever they say, it's just tremendously CRAPPY that a Barcelona core has more or less the same IPC of a current K8 and power envelope and other BS have nothing to do with this, blah!

 

[turpit]

I think thats kind of the conclusion everyone is coming to. Though "4000 pixel...' sounds great, it doesnt mean anything without a whole lot of other info thats missing. An unqualified number.

 

[WR]

This is v3.6.1b (one core!) of an Opty 270 showing 75.47 pixels per second with a total time of 1953.86 seconds.

To bench POVRay, you'd select from the menu, "Render" -> "Run Benchmark (all CPUs)". A message comes up reminding you that benchmark figures are not comparable between versions 3.6 and 3.7.

For example, version 3.6 (single-core) yields ~170 PPS (867 seconds) on a 2.93GHz Conroe. Version 3.7b20 (dual-core) yields ~911 FPS (287 seconds) on the same system. I ran these two just now without optimizing, but the figures are nowhere near each other.

One additional note: The POV-Ray SMP release povwin-3.7.beta.20b has *Timed-out* as of May ist and the *so-called* tests cannot be reproduced.

Temporarily set your system clock to last year while benchmarking.

Hey, I beleive that it should scale better. I just can''t understand why there is seemingly NO scaling from 2 to 4 cores. It definitely deserves investigation.

That's because the HE line of dual-core K8 goes up to 2.6GHz, and there sure isn't an HE Barcelona expected at 2.6GHz.

Assuming good scaling and the standard benchmark, the K8's were running at 1.6 to 1.8GHz, and the video states the K10's were running at the same clock. That's not the best way to argue that people would get 2x performance per watt from a simple CPU swap... because it's simply not true for POVray if you start with a 68W K8 at 2.6GHz.

Barcelona has L3 which lowers latency and widens paths everywhere so there is no "theoretical" bottleneck.

Every level of cache increases minimum latency. It's the hope of the designers that average latency goes down due to cache hits, which save tremendous time over fetching from RAM.

POVRay is not bandwidth constrained in modern systems. Why else does a 16-core Clovertown speed up so perfectly? If POVRay were bandwidth-intensive, such results would make a joke of those complaining about the antiquated FSB.

 

[BaldEagle]

Well if it is a POS then what do you expect them to do;
stall giving information (Done),
have a demo nobody is allowed to talk about (Done),
hype up whatever redeeming characteristic it might have like average power consumption (Done).

Oh my, I guess it must really be a POS.

 

[Wisecracker]

Temporarily set your system clock to last year while benchmarking.

You read my mind I ran it back to April.

Went to 1138.37 PPS in v3.7b20 4-way Opty's at 2GHz 4Gb PC3200 ECC DDR.

@ r0ck:

My point was that Boot Daily showed the **Average PPS CPU time** (their exact words) at 4719.32 when it would actually be around 600. My **PPS CPU time** in the above bench was 288.13.

AND THEN Boot Daily provided their own *extrapolation* . . . simply fud . . .

So I'll do my own extrapolation:

8-way Xeons @ 2.66GHz = 600 PPS CPU time
4-way Optys @ 2GHz = 288 PPS CPU time

288 x 2 = 576 x 1.33 = 766.08 PPS CPU time
(clock-to-clock // core-to-core)

So Boot Daily had their $4,000 workstation's arse kicked by my $1,000 rig :lol:

It's all fud . . .

 

[r0ck]

I have no idea what you are saying
http://www.lostcircuits.com/cpu/amd_quadfx/11.shtml
http://www.xbitlabs.com/articles/cpu/display/intel-v8_11.html

 

[Periander]

Well if it is a POS then what do you expect them to do;
stall giving information (Done),
have a demo nobody is allowed to talk about (Done),
hype up whatever redeeming characteristic it might have like average power consumption (Done).

Oh my, I guess it must really be a POS.

Occam's razor in action!

 

[turpit]

We can only refer to R600 as a debacle if they dont match nvidias driver maturation in 6 months.

I give you my vote to call R600 a debacle.

Realistically, Ive seen conflicting benches, and didnt go into depth on the uarch evaluation, but it looks like it has buttloads of potential.

Seems very much like QFX. Like QFX, R600 seems to be lacking something to scream, and it doesnt seem to be the drivers. Where QFX lacked the CPUs (K10) for it to do what AMD intended, it seems to be R600 needs 65nm to shine.

I feel bad for all the folks waiting on drivers. While ATI has built some mind boggling hardware, they suck relative to Nvidia in the driver support arena, and I doubt the need to speed up R600 is going to change that.

I honestly dont get what ATI was thinking. If R600 was intended for 65nm, why not hold off till 65nm was up and running? Why not play the game, do a node shrink of the 580s, get the node process perfected, then start pumping out the 600s.

 

[JAYDEEJOHN]

This is off topic, but to follow up on what you were saying, everyone knows about the nVidia debacle concerning their 88 series. Law suits etc. Maybe the new DX10 arch itself is such a change that its hard to do. nVidia didnt even have SLI drivers until very recently for Vista. Anyways back on topic, maybe I missed this but what version of povray is being used in the "bench" ?