Was the 4Q prediction an existing one or from this CC? If old: I think the outlook for the GPU division and increased *gasp* ASPs for the CPU division as well as potential for chipset sales for 4Q are pretty good. Not enough to erase the operating deficit, but enough to get relatively close (compared to the rest of this year).
The AMD Execution Thread [2007 - 2017]
- Thread starter overclocked_enthusiasm
- Start date
-
- Tags
- amd
- Status
Was the 4Q prediction an existing one or from this CC? If old: I think the outlook for the GPU division and increased *gasp* ASPs for the CPU division as well as potential for chipset sales for 4Q are pretty good. Not enough to erase the operating deficit, but enough to get relatively close (compared to the rest of this year).
What I heard in this call lwas them reiterate the possibility that they could get to break even in 4Q. Tho it would take a nearly perfect quarter to do it. I'll check the transcript when it comes out.
Edit: http://seekingalpha.com/article/50491-amd-q3-2007-earnings-call-transcript
Edit: http://seekingalpha.com/article/50491-amd-q3-2007-earnings-call-transcript
Bob Rivet said:
You have to think that with higher ASP cpu's coming on stream, the potential of rv670 to continue the gpu growth and Christmas obviously being a good quarter that breakeven is a distinct possibility.
The big problem will be Q1 and Q2 when Intel gets up to speed with 45nm and may once again to start the prices going down whilst still having good margin due to the smaller process. If that happens then AMD's higher ASP for cpu looks under threat again.
We still do not know how K10 will scale either, that could have just as much impact on ASP for that product.
The big problem will be Q1 and Q2 when Intel gets up to speed with 45nm and may once again to start the prices going down whilst still having good margin due to the smaller process. If that happens then AMD's higher ASP for cpu looks under threat again.
We still do not know how K10 will scale either, that could have just as much impact on ASP for that product.
Another minor nitpicking:
http://www.amd.com/us-en/Corporate/VirtualPressRoom/0,,51_104_543~121478,00.html
http://www.amd.com/us-en/Corporate/VirtualPressRoom/0,,51_104_543~121478,00.html
http://www.fudzilla.com/index.php?option=com_content&task=view&id=3882&Itemid=35
Phenom X4 vs. Yorkfield tested in Crysis
Expreview has managed to get their hand on an AMD Phenom X4 and an Intel Yorkfield based QX9650 and have run the two against eachother in Crysis to see which chip is the better one.
Also thrown in for comparison was an Intel Core 2 Duo E6850 and a Core 2 Extreme QX6850. As you might have already realised, the Intel chips are clocked a fair bit higher than the Phenom X4, so to make things a bit more even, the Phenom X4 was overclocked to 3GHz.
This is still not quite a fair comparison, as the Phenom X4 had it's memory running at 375MHz due to limitations of AMD's integrated memory controller which prevented it to run at 400MHz as per the Intel based test systems. The Phenom X4 system also had a lower bus speed, due to Intel's move to 1,333MHz bus.
The Phenom X4 was tested using an RD790 board while the Intel CPU's were tested on a P35 based board. The graphics card used wan an 8800 GTX and the test results are looking quite promising for AMD, at least if the Phenom X4's will retail below Intel's asking price. The only problem is that the Core 2 Duo E6850 seemed to outperform it.
The Crysis time demo was run 5 times and Expreview took the average of those runs, with the Phenom X4 scoring an average 46.48fps while the QX9650 managed 49.95fps. The Core 2 Duo E6850 scored an average of 49.19fps while the Core 2 Extreme QX6850 managed 49.92.
You can check out the full test results and some screen shots here in IE only as the site doesn't work in Firefox.
http://news.expreview.com/2007-10-29/1193590532d6599.html
Phenom X4 vs. Yorkfield tested in Crysis
Expreview has managed to get their hand on an AMD Phenom X4 and an Intel Yorkfield based QX9650 and have run the two against eachother in Crysis to see which chip is the better one.
Also thrown in for comparison was an Intel Core 2 Duo E6850 and a Core 2 Extreme QX6850. As you might have already realised, the Intel chips are clocked a fair bit higher than the Phenom X4, so to make things a bit more even, the Phenom X4 was overclocked to 3GHz.
This is still not quite a fair comparison, as the Phenom X4 had it's memory running at 375MHz due to limitations of AMD's integrated memory controller which prevented it to run at 400MHz as per the Intel based test systems. The Phenom X4 system also had a lower bus speed, due to Intel's move to 1,333MHz bus.
The Phenom X4 was tested using an RD790 board while the Intel CPU's were tested on a P35 based board. The graphics card used wan an 8800 GTX and the test results are looking quite promising for AMD, at least if the Phenom X4's will retail below Intel's asking price. The only problem is that the Core 2 Duo E6850 seemed to outperform it.
The Crysis time demo was run 5 times and Expreview took the average of those runs, with the Phenom X4 scoring an average 46.48fps while the QX9650 managed 49.95fps. The Core 2 Duo E6850 scored an average of 49.19fps while the Core 2 Extreme QX6850 managed 49.92.
You can check out the full test results and some screen shots here in IE only as the site doesn't work in Firefox.
http://news.expreview.com/2007-10-29/1193590532d6599.html
Last edited by a moderator:
Beating the dead horse that is AMD is fashion.
But the fact that the Phenom X4 could be o/c'd to 3GHz on a 790FX platform is rather promising, given that the highest iteration of K10 is currently only clocked @ 2GHz.
Bouncing Zabaglione Bros.
Legend
Promising as in "not as bad as we thought it was going to be"? Even so, if AMD doesn't have confidence enough to ship at 3 ghz, then it may not be reliable enough to clock at speed.
It's kind of moot if Penryn is shipping at 3ghz and overclocks to 4 ghz, and the Core 2 Duo with only two cores is still beating the Phenom in games.
Very true.
I'd say this is more a function of games not being coded to take advantage of > 2 cores yet, than an indicator of any inherent performance weaknesses on Phenom's part.
A shortfall on a chip's target workload doesn't sound like an inherent strength.
Software that is written for more than two cores shows no real change to the performance advantage shown by Conroe, nevermind Penryn.
Phenom, barring some kind of undisclosed erratum fix, has some notable inherent weaknesses.
Going from Barcelona's results, Phenom's memory latency is not going to be much better than Core2's.
Blame the L3 cache or some kind of arbitration logic issue.
Phenom will have half or less of the cache its competitors will have, and half of that cache is over twice as slow.
Phenom's coherency latency between cores on the same exact die--the one big advantage a native quad-core should have over an MCM--was shown by Anandtech to be no better than a Core2 MCM.
That last one really baffles me.
The problem is worse because the shared L2 between cores on the same die for Conroe means data sharing is three times as fast half the time.
I'm curious what this means for the dual-core Kuma. Hopefully, the reduced number of cores will speed up whatever arbitration issue that is butchering any advantage Phenom should have as a native quad-core.
Software that is written for more than two cores shows no real change to the performance advantage shown by Conroe, nevermind Penryn.
Phenom, barring some kind of undisclosed erratum fix, has some notable inherent weaknesses.
Going from Barcelona's results, Phenom's memory latency is not going to be much better than Core2's.
Blame the L3 cache or some kind of arbitration logic issue.
Phenom will have half or less of the cache its competitors will have, and half of that cache is over twice as slow.
Phenom's coherency latency between cores on the same exact die--the one big advantage a native quad-core should have over an MCM--was shown by Anandtech to be no better than a Core2 MCM.
That last one really baffles me.
The problem is worse because the shared L2 between cores on the same die for Conroe means data sharing is three times as fast half the time.
I'm curious what this means for the dual-core Kuma. Hopefully, the reduced number of cores will speed up whatever arbitration issue that is butchering any advantage Phenom should have as a native quad-core.
The same holds true for Intel's quad parts. It's a software issue.
The performance advantage of Conroe/Penryn as compared to... K10? Too early for that. Super-linear scaling is being seen with o/c'd K10 parts right now. I think we need to wait for higher-clocked parts & faster RAM w/desktop chipsets to really call this one.
Of course there have been rumors of "major errata fixes" moving to later K10 revisions (specifically B2). Rumors are rumors though.
Let's not attempt to extrapolate latency numbers for a potential Phenom part without accounting for:
1) registered RAM required for Opterons
2) faster memory used for Phenoms (800-1066 DDR2 vs. 533 or 667)
3) IMC & L3 clock increases that increase with core clock
I think that all adds up to significantly lower latency for Phenom compared to Opteron. Some mildly-overclocked Opty 2350 numbers from XS Note the drop in latency from 76ns @ 2GHz core clock to 69ns @ 2.2GHz, and 63.4ns @ 2.4GHz. You don't see Intel latency numbers scaling like that
I think this is a function of the extremely low clockspeeds at which K10 MPUs are currently shipping. See the aforementioned super-linear scaling and significant memory latency decreases.
I guess this is not enough, is it?
AMD and Toshiba climb semiconductor rankings, IC Insights reports
Despite continued heavy losses, AMD has entered the IC Insights’ top 10 rankings for the first time in its history in 10th position. According to the market research firm, AMD has seen improved sales over the last two quarters with a sequential growth of 18 percent in 3Q07 when compared to the 12 percent growth experienced in 2Q07.
http://www.fabtech.org/index.php?option=com_content&task=view&id=5653&Itemid=69
Check the article for more text and data.
AMD and Toshiba climb semiconductor rankings, IC Insights reports
Despite continued heavy losses, AMD has entered the IC Insights’ top 10 rankings for the first time in its history in 10th position. According to the market research firm, AMD has seen improved sales over the last two quarters with a sequential growth of 18 percent in 3Q07 when compared to the 12 percent growth experienced in 2Q07.
http://www.fabtech.org/index.php?option=com_content&task=view&id=5653&Itemid=69
Check the article for more text and data.
I think it says that Phenom x4 can't compete per clock with a Conroe dual core in apps that are perhaps dual-core optimized. I'm not sure if/when dual cores will be at a disadvantage to a lower-clocked quad, but it seems to me that dualies aren't even being utilized fully yet.
Super-linear scaling ? You mean as the frequency gets higher the scaling goes above a linear increase? Hmm, sorry but that has never happened on any x86 cpu and I do not expect it to happen on a K8 derrived K10 either. You can have good scaling where you keep pretty close to linear as the frequencies rise, but you cannot go above it.
Can you show any overclocks where it is showing this supposed super-linear scaling? Here's a post by one of the best overclockers in the world
http://www.xtremesystems.org/forums/showpost.php?p=2519472&postcount=170
"There is no strange scaling behaviour within same CPU type."
Scaling above the linear maximum amount seems to be more straws being grabbed by people wishing this processor to be performing better than it is.
My point is that software that is written to take advantage of more than two cores shows a performance advantage for Core2.
There is no advantage for Barcelona until it is in a multisocket system, where the bandwidth and latency limits finally start to hinder Intel's scaling.
Phenom is likely to be single-socket except for a handful of enthusiasts.
I haven't seen any superlinear scaling, at least nothing significant beyond some cyclical variations due to the coarse memory frequency scaling thanks to the CPU-dependent frequency divisor (an inherent headache to the IMC if there ever was one).
1) The comparisons for Opterons were against Core2 chips running on FB-DIMM boards, which is the best-case comparison AMD could hope for.
2) Core2 chips on the desktop will be running the same memory type.
3) The gap recently was something like 50ns versus 70ns.
Phenom at release and for a quarter or two will likely peg it at 60 vs 70 ns.
The gap really isn't as big as it used to be.
The IMC's location can only do so much when the latency of the memory device itself is now dominant.
2 GHz to 2.2 GHz is a 10% clock increase, yielding a 10% drop in latency.I think that all adds up to significantly lower latency for Phenom compared to Opteron. Some mildly-overclocked Opty 2350 numbers from XS Note the drop in latency from 76ns @ 2GHz core clock to 69ns @ 2.2GHz, and 63.4ns @ 2.4GHz. You don't see Intel latency numbers scaling like that
2.2 to 2.4 is 9% and a 9% drop.
Sounds linear.
At 3.0 GHz, that would hint that an ideal scaling would lead to a return to roughly where the A64 is now. I feel that it will wind up marginally worse than A64 at 3.0GHz, as the L3 penalty in an exclusive cache setup cannot drop to 0.
edit: This would be at the same memory speed grade, since DIMM latencies are a more dominant factor.
Any gains Phenom expects over A64, I'd expect to come from it having support for faster grades in the future.
Phenom's competition is likely to have the same benefit and will keep the difference in the same range.
The coherency latency for Barcelona is 145ns. (edit 152)
That's a lot of superlinear scaling to arrive at something that is more than incrementally improved over an MCM (Penryn's higher bus speed should also lead to an incremental improvement).
Addendum:
Just to bring this closer to the financials side:
The baffling latency for cross-core communication shows just how marginal the native quad-core design is over an MCM.
AMD should have been able to figure out the latencies of MCM snooping on a shared bus.
More importantly, how can they engineer a quad-core with inter-core communications taking about as long as a round trip to and from memory?
Last edited by a moderator:
Check the XS link I posted previously. Super-linear scaling is observed in the SPI tests.
- Status
