Register for free to join our community of investors and share your ideas. You will also get access to streaming quotes, interactive charts, trades, portfolio, live options flow and more tools.
Register for free to join our community of investors and share your ideas. You will also get access to streaming quotes, interactive charts, trades, portfolio, live options flow and more tools.
economaniac,
Itanium also has 48 GB of memory vs. 32 GB of Opteron (if I am reading these correctly).
And it looks like Opteron is running in 32 bit mode using these 32 GB of data. (PAE works on Opteron?) That's interesting. There should be a substantial boost once Opteron runs in 64 bit mode. But even this result is very impressive, IMO.
Joe
chipguy,
Do you design full custom integrated circuits for a living?
Well, don't quit that day job.
Have you done product engineering and yield analysis? What about devising standard cost models for ICs? If not you don't have the slightest clue about most of the factors whose consideration went into my comment.
Don't pound that chest so hard. You may injure yourself. I don't want to be responsible for that.
BTW, you started with 3 datapoints: pricing of 3 chips, of a boutique server chip and you went on a tangent from there, presumably applying logic of pricing of volume product.
I used the same "logic" and concluded that it is a heroic struggle to yield a single good Itanium die.
Joe
RE: Tom's review
Are these guys pathetic or what?
Joe
Elmer,
Let me say that there's no denying that Hammer looks pretty damn good in 2 & 4 way configurations. Now, can they produce it?
Over last 3 years I have been following AMD closely, I go between "can they produce it" and "can they sell it". I spend more time worrying about "Can they sell it".
Joe
Elmer,
Yes you're right but this small set are AMD handpicked benchmarks. Don't you think Intel could handpick a set of benchmarks that told a completely different story?
Yes, more info will come out over the next 24 hours.
Joe
Elmer,
Intel beat Opteron in SPECint with the 3.06GHz P4.
That's not exactly true. Here are some numbers:
Peak / Base / Avg
Opteron 1170 / 1095 / 1133
Dell 1130 / 1085 / 1108
Intel 1107 / 1099 / 1103
Itanium 719 / 711 / 715
Elmer,
As I predicted, Hammer fails to match Intel SPEC performance despite their promises.
Good prediction.
I don't mean to rain on your parade, but there are 2 issues that could swing things from -5% to 5% or so such as:
- adopting DDR-400
- running in 64 bit mode (submission is in 32 bit mode)
I don't think AMD is going to go with DDR-400 in the servers (and neither is Intel). But they may in the workstation market. As far as 64 bit OS and compilers, they will come and be improved to the point that they will outperform the best 32 bit code.
BTW, what are the base and peak scores of the mysterious 3 GHz P4 with the patch applied?
Joe
Elmer,
Despite claims that Hammer would beat anything Intel has to offer, Hammer gets it's but kicked by Itanium in TPCC and edges out Xeon (for now)
I don't think you are reading these correctly. You have to pick a processor, and than look at it across all benchmarks, and see which one is an overall winner. If you pick Opteron vs. Itanium, Opteron is the overall winner. If you pick Xeon whatever, Opteron is the overall winner.
The fact that all the Intel processors combined are able to steal a win here and there vs. 1 concrete Opteron processor doesn't mean that Opteron is not the overall winner. With your approach, you could even add Alpha to the mix and point to some benchmark where Alpha is still leading and say that Intel is ahead.
Joe
I wonder about this slide:
http://www.amd.com/us-en/Processors/ProductInformation/0,,30_118_8796_8800~69673,00.html
SPECweb®99_ssl (Zeus) 4P Servers
They have Itanium 2 1.5 GHz listed. I am a little confused by this. Is this Madison?
Joe
chipguy,
If true the current process + current stepping yield vs frequency curve peak occurs around 1.5 GHz or so. I guess we won't be seeing 2.4 GHz parts any time soon.
Using your "logic", I have to conclude that for Intel to make a single Itanium is a mighty struggle.
Joe
wbmw,
Check out the tables in the following link. They show which modes support adaptive voltage and frequency adjustments, and which processors have the following settings.
I eventually concluded that from the previous review, and posted here:
http://www.investorshub.com/boards/read_msg.asp?message_id=921624
I haven't read the the current review yet.
Joe
belgiangenius,
I'm a very unimpressed AMD shareholder.
"Paper trading" doesn't mean that you are actually a shareholder.
Joe
Elmer,
One of the glaring issues is that SOI hasn't bought AMD anything.
It will be only a few hours before we know the answer, so there is no point in jumping to conclusions. But if these Opterons really dissipate power in 40s Watts, I think SOI bought a enough.
This would be 1/2 of the SOI benefit, which is low power at relatively low clock speeds. There is the second half, if AMd persues it, which is to achieve high clock speed at high power dissipation.
Joe
wbmw,
Doesn't make much sense. If the TDP is 80W, then vendors will be paying for 80W cooling solutions, regardless of the actual CPU power. Guiding conservatively here costs money.
This seems like a week FUD.
Was it you who commented about how difficult it is to launch a new speedgrade of the server chip, because all of the testing involved (I presume the time and the cost of it).
Well, if it costs so much to validate the next server speed grade, AMD approach is exactly correct. You spend dollars in single digits (maybe $10) now on a slightly oversized cooling solution, but save $100,000s later. And, the computers purchased today can be upgraded in the future - by the vendor (releasing faster speed grades) and by the customer.
Joe
chipguy,
My observation was based on the (planned) steep ramp of SOI wafers (since there is no K7 on SOI, I think it would imply K8). Well, unless AMD plans on manufacturing SOI based K7 using 90nm process. I doubt it.
Joe
combjelly,
I am sure the info is correct up to Q4 2002. Even Q1 2003 may be right, even though the SOI wafer starts seem to be a little on the high side at some 1,000 wpw throughout the quarter. This would imply production of some 50,000 good die per week or some 650,000 per quarter. That seems a bit on the high side (unless there is a desktop / workstation Opteron chipset available prior to September 2003 A64 launch date to soak up all these Opterons.
At some point this quarter, probably very soon, A64 wafers need to be started. Or actually, some of the Q1 wafer starts may already be A64 starts, so that 650,000 number may be only half Opteron and half A64.
I think the ramp will probably turn out to be less steep, to reduce risk of potential infrastructure problems, or possibly problems with the new process.
Either way, it looks like a year from now AMD plans some 90% of wafer outs to be Hammer.
Joe
Dan,
From the presentation, they said: "initial Sledgehammer vehicle now in tape-out" (page 13). That was in November 2002. So AMD may first 90nm hammer silicon by now.
Joe
sgolds,
with 180 to 130nm shrink, there were possibly 2 sets of issues, one for bulk one for SOI. I don't know what problems were encountered where.
On the 90nm process, if it performs adequately, it should be still producing chips well into 2005. Even if Intel goes to 65nm in late 2004 / early 2005, there will still be need for lower end chips, even if the high end start to migrate to another process technology.
Joe
econo,
Your comments make sense. Another issue is that AMD has a fab that is capable of producing competitive chips well into 2005 on 90nm, before any costly upgrades will be needed. AMD doesn't have an option of not using Dresden to the fullest extend. The depreciation charges are hitting the income statement every quarter, and will continue to do so for some time.
Before we talk about production (even though there can be some production at IBM), AMD needs to figure out if they are going to adapt IBM's 65nm process to 200mm equipment in Dresden or if they are going to switch to 300mm. In late 2004 / early 2005, there should not be any premium for 300mm capable equipment.
Joe
sgolds,
I think there will be 90nm process from AMD. I think there will be some clarification of the roadmap during the shareholder meeting. I don't think we should take on, not very clear statement about 65nm as a definite answer of what AMD plans.
One thing is that the 90nm will is developped for Dresden's 200 mm production lines, and IBM's 65nm is I believe targetted for 300mm IBM fab. It may be customized for Dresden, or Dresden may switch to 300mm, but it is all unclear now. I think it would be too optimistic to expect 65nm anytime before 2005, and AMD will need to have a competitive, volume product throught 2004.
To get 65nm from Dresden in 2004 is just way too optimistic.
Joe
Keith,
I don't think 90nm was mentionedduring the CC. But 90nm processors are on AMD's roadmap:
http://www.amd.com/us-en/Processors/ProductInformation/0,,30_118_608,00.html
Joe
wbmw,
I am talking capabilities. When you have Megabytes of L2 (actually L3), 64 bit addressability, good I/O bandwidth, it will have good database performance. Whether people buy is a different question.
For the last 2 years, the answer was obvious. Itanium just wasn't ready, the software wasn't there despite thousands and thousands of press releases (or commitments as you called them) stating otherwise. Customers have sayed away in droves.
In H2 2003, it may become actually possible to get some productive use out of Itanium, if one is determined enough. How customers will vote is yet to be seen.
Joe
yb,
I think there are different requirements for main processor in database server, for example, and a processor in the blade. I think a processor like Madison is well suited for cetralized database server, while blades are more likely to be web servers.
Realistically, we have to expect Opteron to first make it into 2-way servers. They are actually a much bigger market processor-revenue-wise. If there are 4 or more way Opteron servers, that will be a nice bonus.
I think it is possible for AMD to add more cache to Opteron, if it becomes more popular, but I think the approach is bottom up, unlike Itanium which is top down.
AMD doesn't have the deep pockets of Intel to conjure up an all-powerful CPU and hope that buyers show up. AMD has to go incrementally, like with x86-32 to x86-64. Likewise, if there is a strong adoption of Opteron, higher end models with more L2 can be added later on, which would be a less risky proposition than releasing a $4000 6MB L2 processor from the start.
Opteron competition is Xeon, not Itanium. There is a market share to be taken away from Xeon. Itanium has no market share. It is just trying to take some from Sun and IBM (and from discontinued technologies such as Alpha).
Joe
yb,
I don't think AMD is skipping 90nm. As far as timing, there are number of thresholds that the process needs to pass before we see a chip made on that process technology at Newegg. I don't think anybody said that there will be 65nm chips sold at Newegg in 2004.
Joe
sgolds,
AMD is going to have a pretty bad YoY quarter as well, but hopefully at least slightly up over last quarter.
I think Opteron revenue at this point (without workstation / desktop motherboards) are extremely hard to project, so I am afraid that AMD will not project a breakeven Q2. Flash would have to stellar to achieve that. I wonder if it will be possible to achieve at least cash flow breakeven, or positive in Q2.
Joe
What was the expected range of Intel revenue and ESP? It seems good at first sight.
Joe
Tenchusatsu,
some board rework (not-so-good case),
By board, you mean motherboard?
Joe
wbmw,
PowerNow mode will be somewhere in the middle of 502 seconds and 174 seconds, but it will not be 174 seconds.
No, that's not correct, in all adaptive modes, the result is in range of 174 to 179:
http://www6.tomshardware.com/mobile/20030410/sony_vaio-10.html
The adaptive modes on the graph are:
- Office desktop AC + DC
- Portable laptop AC + DC
- Max Batter AC only
- Presentation DC only
Likewise, the Pentium M SpeedStep mode will be between 154 seconds and 420 seconds. THG obviously aimed at testing those two extremes
Actually, looking at this again, while you are wrong it the above statement, I think you may be right overall, because Centrino behaves exactly like Mobile Athlon with PowerNow. In modes that correspont to PowerNow adaptive modes, Centrino behaves like Athlon, meaning it runs top speed when necessary, and it runs in slow speed when forced to that mode, just like Athlon. This would mean that Centrino does have adaptive mode (unless power these modes are hard coded to full speed, which wouldn't make sense).
See the battery modes of Centrino here:
http://www17.tomshardware.com/mobile/20030205/centrino-13.html
and Mobile Athlon:
http://www.tomshardware.com/mobile/20030410/sony_vaio-10.html
Joe
wbmw,
Are you going somewhere with this 'PR stunt' theory?
Not really. Just some innocent stone throwing...
Joe
DewDiligence,
This write-up is from Feb., 2003 and hence may not be the latest word on the Pentium-M, but I have seen other articles similar to this. Are all of these publications mistaken about the power-management feature of the Pentium-M? Very strange…
I am puzzled by it as well, for the same reasons: If the functionality was there, why doesn't anyone mention it in their reviews, why isn't Intel advertising it? Intel is very good at letting everyone know about every single improvement in their product.
Now, I went back to Tom's review, and I came to conclusion that these guys are morons, (or they are purpously slating their test).
If you look at this page:
http://www6.tomshardware.com/mobile/20030410/sony_vaio-04.html
you will see that there are 3 modes that basically disable PowerNow, the greatest feature of AMD processor. These are: Presentation, Always On and Max Battery. Yet, they keep using these over and over, in some cases to the exclusion of other, reasonable (= adaptive) modes, suce as here:
http://www6.tomshardware.com/mobile/20030410/sony_vaio-08.html
The Sysmark test shows that the adaptive mode (Portable Laptop) is within less than 5% of Always On mode, yet, it achieves substantial power savings. Power savings that you will not see, because Tom's reviewers are too stupid to figure out how to show it. If I recall from previous reviews, the adaptive mode power usage is much closer to the slowest speed than to highest speed, which brings me to my original point that only a complete moron would use a mode any other than the adaptive modes.
Now on to wbmw's post:
http://www.investorshub.com/boards/read_msg.asp?message_id=913108
His performance comparisons don't make much sense, because he doesn't understand how the power modes operate (no wonder, since the author at Tom's probably doesn't know either):
In POV-Ray, the Athlon XP's best score with the HP Evo laptop is 174s under portable/laptop mode and 502s under max battery mode. The Centrino laptop from the previous article only took 154s in portable/laptop mode and 420s in max battery mode, making it 13% faster with full power and almost 20% faster on batteries.
Ok, in full power, it makes sense, but you can buy 2200+ to deliver Centrino-like performance for the same amount of money or less, making the comparison somewhat arbitrary.
But on battery power, it is still 174, while Centrion drops to it's slow speed and score of 420, which means that Athlon-M performance is 141% faster. So in mobile use, it's no comparison, Centrino is blown away.
As far as power usage and battery time, others have pointed out that the comparison is that of apples and oranges, where screens are different.
Let me add to it that the test does not specify anything about the CPU mode in their Athlon tests. All they show is the battery times using the different settings of the video card, because they keep talking about PowerMizer is a power control app for the video card, not CPU.
So as far as I can tell, there is no way to draw any comparison from such an incomplete info.
Joe
wbmw,
I don't have any better clue to what the problem is. The only thing that I can observe is that Intel waited until all the reviews were published before informing of the halt in shipments. Kind of curious, isn't it?
Joe
yb,
I seriously hope that the Inquirer is right about workstation class motherboards available. But I don't know what they mean by chipsets. AMD already has an AGP chipset component on HT (in unknown state, stability and performance-wise). So it is just a question of someone including it on an Opteron motherboard. If HT is what AMD is selling it to be, it should not be such a major project to do this for mobo partners.
One possible explanation is that AMD did not make the reference designs available, and mobo makers are afraid to venture on their own, so maybe what everybody is really waiting for is for AMD to complete Opteron workstation reference design.
Joe
wbmw,
If that is the case (Banias can adjust the frequency on the fly) what is the point of the whole test (at Tom's Hardware). If the automatic mode is available, I don't see any reason to run in any other mode. Because what do other modes do, just cap the frequency?
For example, let's take the "presentation". If you are doing the presentation, the computer is in the slowest speed (or sleep) all the time, unless you click "next" slide, when there is a flurry of activity (that you want to finish ASAP) and back to sleep. Doesn't it make sense to re-draw at the highest speed?
I still not convinced, and the way Tom did the review makes me less convinced, not more.
Joe
PS: Do you have a link to MPF2002 slides?
wbmw,
If you come across something that explicitly says that the processor can dynamically adjust multipliers, depending on load, please let me know, I would appreciate to know the definite answer. The info in the tech document you linked is not clear.
Based on the info I have come across so far, my conclusion is that this functionality is not there. If it was, it would be clearly stated, and it would be marketed.
Actually, I found one site that suggests this functionality:
http://hardware.earthweb.com/chips/article.php/11055_2108711_3
but I would prefer something more definite.
Joe
wbmw,
There is no indication in the spec that there is an automatic mode analogous to PowerNow automatic mode. All it says that the clock speed can be adjusted by software (such as changeing the power mode usint the mini control panel, I assume), but there is no indication that it can happen dynamically, 10s or 100s times per second, based on the demand on the CPU.
I looked at couple of reviews, and neither mentioned it, which makes me suspect that there is no automatic / dynamic mode for clock speed changes, so I go back to my original transmission analogy.
Joe
wbmw,
Centrino has a power management system that operates exactly the same as the Athlon XP. In other words, performance is varied at multiple levels, depending on the performance requirements of the CPU.
Really? Do you mean that multipliers and voltages can be changed on the fly, for example, while playing a DVD? I was not aware of it. Maybe I need to read (re-read some of the reviews).
BTW, I never argued that Banias is a bad CPU.
Joe
wbmw,
PowerNow has an automatic mode, with variable clock speed based on the load on the CPU. It really makes no sense to run the CPU in any mode other than automatic (unless you are plugged to AC power).
Your comparison is kind of like comparison a car that has transmission vs. car without a transmission under the condition that the car that has transmission is not allowed to used it.
Joe
Elmer,
So first silicon "real" is expected the next month but samples are already in the hands of customers! I guess that aren't real silicon.
Paper?
Joe
Tenchusatsu,
Coherency checks happen on the hardware level (i.e. the microcode level), which is completely transparent to software. There are systems out there which do have interactions between hardware and software, but like I said before, that's big-iron stuff, and not something you'd typically see in an x86 system, especially 4-way.
That's why I said that it would be on the next level. There would need to be some hooks built into the hardware to allow this, which I am not sure Opteron has.
No, it shouldn't. For a 4-node, 16-way Xeon system, the ratio of remote-to-local access times is large. For an Opteron system, the ratio of remote-to-local access times isn't that big.
Makes sense. I did not consider time of remote access, I was just thinking percentage of accesses.
Joe