DELL overestimated INTEL
DELL execs explain why they took so long to go AMD64.
Jeff Clarke, senior vice president (SVP) in Dell's product group said,
"We overestimated Intel's ability to deliver and catch up."
I remember one DELL VP said Intel could do 64 bits in a nanosecond...
At the same time, Dell "underestimated" AMD, admitted Brad Anderson, another Dell SVP in the products group. "We believed that AMD's ability to sustain its performance advantage was not going to continue," he said. "We missed that."
I hope these DELL execs read my blog and understand AMD has all the grand masters in computer architecture: Opteron, Alpha, Power, UltraSparc, MIPS, Itanium and PA-RISC. Intel simply lacks the brain power to compete against these grand masters.
I imaged a meeting among AMD and Intel designers. The Intel team won't even be able to understand what AMD folks are talking about.
All of Intel's stuff seem so primitive -- Pentium Pro technology is all they got.
Woodcrest is ultra low end 2P stuff.
The Register has an interesting article and stock price chart for AMD, HP, SUN, IBM, INTC, DELL and SGI. The more AMD a company had, the better its stock performed. The more Intel a company had, the worse it performed. Arguably, SGI had more Intel than Intel itself. Intel has diversified itself. SGI is pure Itanium, and SGI BKed. Intel had more AMD than DELL, because Intel borrowed AMD64 instruction set and became an AMD64 clone maker.
Michael Dell and his people loved Intel, they still do, they may even hate AMD for whipping Intel. But they love themselves more. When AMD is growing, and Intel shrinking, Dell's fragile financials will not be able to sustain its valuation. The Enron scenario and the thoughts of Feds knocking their doors forced them go AMD. That's the only way DELL can find growth and save themselves from the Enron day, if all possible.
I saw this last year, and I told AMD and Dell folks about this last year. It's called inevitablity.
Labels: dell
posted by Sharikou, Ph. D. @ 9/14/2006 11:10:00 AM
72 Comments:
I wish I could reply but you make no sense.
Sharikou,
Could you in plain language tell us what is core duo and core 2 duo, their availability and the key diffrence. I am confused.
And from the same article you quoted in your link:
But Dell refused to speculate about how many AMD-based systems it will sell going forward. "More than zero, less than 100%," quipped Michael Dell, company chairman.
Furthermore, Dell does not believe that AMD Opteron chips hold an advantage anymore, thanks to Intel's new Woodcrest Xeon chips. "Until very recently, AMD did have the performance and price-per-watt advantage," Hand said. However, "Intel's new designs versus AMD are very competitive."
To be honest.. I doubt anyone reads your blog for facts. Especially not executives. You're not that bright... really. The reason many people read your blog is for a laugh or two. If you actually think you're pretty intelligent then head on over to the THG forums for a debate. My name is ElMoIsEviL, I'll debate you (and beat you). I'm not a fanboi either.. so goodluck.
It would seem there are even more people contradicting your Intel laptop explosion theory.
Link.
Core Duo is Yonah, the native dual core version of Dothan (the second version of Pentium M for Centrino) with shared L2 cache.
Core 2 Duo is Conroe which is about the same level of upgrade over Yonah as K8 is over K7. C2D has increased instruction issue, greately enhanced SSE performance, and 64 bit capability. Compared to Yonah, it is faster at the same clock, has better FP and SSE performance, has 36 bit addressing, and can run AMD64 code.
Suggesting the Woodcrest is Pentium Pro is ridiculous. Pentium Pro led to PII/III which was developed into Pentium M which was developed into Yonah which was developed into C2D.
It is incorrect to call Woodcrest primitive. For single and and dual socket systems it is a very good chip and roughly equal to Opteron.
Its disadvantages versus Opteron include that it only has 36 bit addressing instead of 40. It can't scale beyond dual socket. It uses a much older cache synchronization protocol called MESI whereas Opteron uses MOESI. And, Woodcrest has to communicate over the FSB for both I/O and cache coherency whereas Opteron uses the HyperTransport bus.
"All of Intel's stuff seem so primitive -- Pentium Pro technology is all they got."
And AMD ? what they hv is leftovers of Alpha - which is even more primitive.
BTW doesn't "X86" say it all - Both Intel and AMD are living on architecture which is decades old.
If your business is selling computers, then you must be able to understand that industry very well.
That means the architectures, the lingo, the systems, the trends, the companies, the market, etc, etc.
What DELL executives basically admited to here is that:
"We don't really know what we're doing"
"We guessed and thought we understood, but we were wrong"
As glad as I am that DELL has decided to sell AMD technology, I only have one suggestion for them:
Hire people who have a clue!
Then, and only then will they be able to use Micheal's JIT business savy to make DELL what it once was.
If anyone who knows anything about transistors could comment on the two following, it would be great.
First (read it some place called the internet):
[...] to continue
shrinking of the transistors. Material limits will mean shrinking
transistors (called Moore's law of scaling) will likely slow down -
might take 3-5 years per generation instead of 2-3. This will allow
rest of the industry to catch up with Intel.
Can this really be true? Or is it over generalized?
Second:
I read some place that AMD was going to release a new, faster transistor for their next generation processors (K8L, K9 or something, I forgot).
So if I understand it correctly transistors can have revisions too? I thought these things were so basic that they all smelled, looked and behaved the same.
My question then is what are the fundemental differences (apart from switching speed) between AMD and Intel transistors... if any?
Thanks!
“I wish I could reply but you make no sense.”
You have not been paying attention and are perplexed just like Intel/Dell vip’s. Those whom don’t understand are to far behind and need to keep up.
If you actually think you're pretty intelligent then head on over to the THG forums for a debate.
That's a place infested with low IQ rednecks who can't really afford ther beloved Conroe chip.
I make predictions based on my logical analysis, and most of my predictions come true. That's a capability even Dell folks have to respect. I told them the prospect of Enron time, I told them they must go AMD before the end of 2Q06 when they had no such plans to do so.
Any information on how E521 and C521 are selling in comparison with E520? Only sales figure will tell if AMD is getting any foothold.
"Any information on how E521 and C521 are selling in comparison with E520? Only sales figure will tell if AMD is getting any foothold."
Not until the end of this year, since the estimate shipping date is in mid-October. These AMD-based machine are for the Holiday season sales (we learned that a few months back from Taiwanese sources).
Furthermore, Dell does not believe that AMD Opteron chips hold an advantage anymore, thanks to Intel's new Woodcrest Xeon chips. "Until very recently, AMD did have the performance and price-per-watt advantage," Hand said. However, "Intel's new designs versus AMD are very competitive."
"[...] to continue
shrinking of the transistors. Material limits will mean shrinking
transistors (called Moore's law of scaling) will likely slow down -
might take 3-5 years per generation instead of 2-3. This will allow
rest of the industry to catch up with Intel."
It seems the "normal" scaling will continue until 45nm around 2009. Maybe 32nm will also be "on schedule."
What happens after that is difficult to tell. But if Intel would be slowed down, then so will everyone else, too. Unless, of course, there is a technological breakthrough (which is needed to continue the size shrink and performance improvement) at somewhere else.
I've known people making transistors with feature size ~7nm in labs, using nominal Si lithography. It was 5 years back and it didn't mass produce then. The guy went to work at Intel, and I won't be surprised if Intel gets it to mass production first.
"My question then is what are the fundemental differences (apart from switching speed) between AMD and Intel transistors... if any?"
AMD uses IBM's SOI technology; Intel uses its own bulk Si approach with advanced tweakings. Their makings are very different from each other. Their behaviors are mostly the same (switch on and off in a controlled manner), but with different characteristics (such as threadhold voltage, saturation and leakage current, conductance and capacitance, and heat dissipation, etc.).
Transistors do have "revisions," which is mostly transparent to circuit designers. Some changes in transistor properties however require different circuit design practices in the lower (physical) layer.
Modern transistors are small, but very complex; it can have many design variations, and it can take tens of steps to make one. But fabrication means you get to make billions and billions of them at once, in the same tens of steps. That's why larger wafers (e.g. 300mm)) is much more advantageous than smaller ones (e.g. 200mm) economically.
To be honest.. I doubt anyone reads your blog for facts. Especially not executives. You're not that bright... really. The reason many people read your blog is for a laugh or two.
I was going to respond but this pretty much sums up my sentiments. Sharikou posts pure speculation and then calls AMD masters of the universe. You are a cartoon Sharikou.
I make predictions based on my logical analysis, and most of my predictions come true.
You are soon to have a long list of failed predictions.
Just to name a few that will fail:
1. K8L will reach stock speeds of 4-5ghz (you said this in a comment somewhere)
2. Dell will go near 100% AMD in 3Q08
3. Intel will be bankrupt in 2Q08
4. AMD will surpass Intel's marketshare in 3Q07
5. AMD will have 40% marketshare in 4Q06 (you are vague about this one by saying run-rate. Probably because you know it will fail soon)
Dell does not believe that AMD Opteron chips hold an advantage anymore, thanks to Intel's new Woodcrest Xeon chips.
Dude, Woodcrest is only for ultra low end 2P. There is no woodcrest for 4P. Nowadays, all companies are touting their 8P 16 core Opteron workstation/servers. DELL is simply not in that lucrative market.
Hes got you there. Maybe you should take Sharikou more seriously. You may not beleave him but he has been right as he says so. Things do tend to come true he says. So I wouldn't be surprised.
I been here for awile and seen it happen. Some passer-byers or new people automaticly just give their 2 cents in when they prob never read his intire bio. Well if you did then you would know hes very smart and knows what hes talking about.
If you don't well then good for you. Your too close minded to see the truth Intel has a cloak over themselfs with to mask its nearing end. Because Intel can't scale or put more then P2 systems out will make it fail. Why would I want a server jamed with only P2 systems will just take up 4 times more space.
Now if I go opteron I can go P8 ways in one blade. Very reduced space by 4 times and more then enough multitaksing power. Can do what intels server space can do thats 4 times larger only Opterons in 1/4th of that space could domonate it. Guess why Google and yahoo went AMD for servers just to name a few.
You maybe able to put your 2 cents in but you can't prove it. P2 systems can not equal the power of P8 dual or P8 quad cores. Sorry but its impossible unless you overclock the system by 800% or even 1600% oviously impossible for 4 cores to do the work of 32 or even 16. Silly little kids and their fantisy's Isn't it SOOO CUTE! Awww...
So where does the magical 14 or 30 extra cores come in to beable to match a opteron system with hmm? Got ya there. (BTW Each cpu is assined to a task allowing different programs to be exicuted on seperate OSes and physical hardware. This allows multi taksing to be amazing on each blade holding dual or quad OS'es running at the same time in one blade at a time or more even. This allows for all the cores to be used in the intended way. Software limit solved, sorry no software can't use more then 2 cores speach today. In that configuration it does.)
"I been here for awile and seen it happen. Some passer-byers or new people automaticly just give their 2 cents in when they prob never read his intire bio. Well if you did then you would know hes very smart and knows what hes talking about."
Maybe you could enlighten us as to his education. All we know is he's a 'Freelance Journalist' who's fraid to own up to his blog.
http://www.theinquirer.net/default.aspx?article=34392
AMD doesn't have the better deals with DELL you say? Well let me pop your callur for ya! POP PPPPP POW POPED YOUR CALLUR!!!!!!!!!!!!!!!!!!!
Anonymous said...
"Well let me pop your callur for ya! POP PPPPP POW POPED YOUR CALLUR!!!!!!!!!!!!!!!!!!!"
Did you even look at the links they put in the article, probably not.
If you had you would have noticed the rather peculiar price differences relating to the E6400.
I wrote an email to the person who wrote the article...
I just have to ask you did you notice the price increase going from the "Intel® Celeron D® Proecssor 346" to the E6400 was $430.
The increase in price would probably have to include a different motherboard and then the cost of the processor, we all know Dell is not using the top of the line Intel, ATI or Nvidia chipset for the E6400 processor.
Dell gets better prices than Newegg, yet they are ripping people off, thats $430 for a motherboard and processor + the $75 Celeron.
So $500+/- from Dell...
Newegg...
Intel Celeron D 346 = $75
Intel Core 2 Duo E6400 = $227
Intel BOXD946GZISSL (Intel® Graphics Media Accelerator X3000) = $83
So if you go to Newegg and buy all the parts its $385, then subtract $75 for the Celeron you don't need, and you total $310.
I would have to say that Dell knows how good the new Core 2 Duo chips are and are charging a premium for them, or people have no clue, and Dell is ripping them off.
That should be your next story.
"Dell exploits computer shoppers"
Thanks,
enumae
PS: The motherboard I used could also be used with the Celeron so theoretically you could subtract that as well.
Glad to see you believe everything you see, and do not look at the details.
" Anonymous said...
To be honest.. I doubt anyone reads your blog for facts. Especially not executives. You're not that bright... really. The reason many people read your blog is for a laugh or two. If you actually think you're pretty intelligent then head on over to the THG forums for a debate. My name is ElMoIsEviL, I'll debate you (and beat you). I'm not a fanboi either.. so goodluck.
12:53 PM, September 14, 2006 "}
been there, and admins are more than happy to ban anyone who tries to say positive stuff from AMDers.. its like "we mods go with the crowds of the local community, and if they're all intel kisser,s we will be intel kissers and kick anyone supporting AMD who makes a good debate, even if they're not trolling"
Intel's only got Pentium Pro, eh? Well, that speaks volumes for AMD then, since that means AMD's X2 can't outdo Pentium Pro technology in the vast # of benchmarks out there.
Dude, Woodcrest is only for ultra low end 2P. There is no woodcrest for 4P. Nowadays, all companies are touting their 8P 16 core Opteron workstation/servers. DELL is simply not in that lucrative market.
Who cares? They have Tulsa. Tulsa proves that all that is needed to crush AMD Opteron is an old Netburst dual core Xeon with a bit of extra L3 cache.
Guess why Google and yahoo went AMD for servers just to name a few.
Google went with Woodcrest based servers - they decided AMD is not reliable enough with their old junkyard processors.
http://www.cmpnetasia.com/ViewArt.cfm?Artid=28813&Catid=1&subcat=9
"DELL is simply not in that lucrative market."
It's lucrative, but not volume, which is what Woodcrest is.
Scientia:
Its disadvantages versus Opteron include that it only has 36 bit addressing instead of 40.
How many dual-processor servers need more than 64 GB of memory? How many DIMMs does Opteron DP need to populate on the channels to exceed 64GB and can it be done?
It can't scale beyond dual socket.
"Can't" is a strong word. Doesn't the upcoming quad-core show that it could (think two sockets on each of the two buses)?
It uses a much older cache synchronization protocol called MESI whereas Opteron uses MOESI.
Both MESI and MOESI are equally venerable. If Woodcrest sits alone on the bus, why would MOESI make any difference? Anyway, can you point out a study which quantifies the performance difference between the two protocols?
And, Woodcrest has to communicate over the FSB for both I/O and cache coherency whereas Opteron uses the HyperTransport bus.
Actually, snoop filters block pretty much all the I/O traffic from reaching the bus. I/O does cut into the bandwidth of Hypertransport, so this is advantage Intel.
>>>Furthermore, Dell does not believe that AMD Opteron chips hold an advantage anymore, thanks to Intel's new Woodcrest Xeon chips. "Until very recently, AMD did have the performance and price-per-watt advantage," Hand said. However, "Intel's new designs versus AMD are very competitive." <<<
And you believe this after dell's track record?
They did say competitive, not better...
https://beta.blogger.com/comment.g?blogID=9020108538297955304&postID=955126355635466373
Scientia from AMDZone said...
Red? The only one I know of who is talking about Intel's being in financial trouble is Sharikou.
Intel is not going into the red; Intel is shrinking. These are two different things. Intel is in no danger of losing money however they are making less money than they did last year. In other words, Intel has negative growth while AMD has positive growth.
A better indicator than the dividend is how much Intel spends on stock buyback and how much cash and inventory they have at the end of the year.
September 13, 2006 11:19 AM
Sad that an AMDer can see past your crap.
For one your link is not reliable. 2nd of all thats why the US military denyed all conroes into their systems. They are unreliable. Again only for 2P systems hello. Ofcorse woodmistake is better then a opteron 2P but again you fail to see thats all they can do.
Know what scaleing means? 2 or more cpus working together to gain performance. On a legesy FSB keep dreaming. Thats why there are no 4P or 6P or 8P Servers for intel. And thats why Opteron is the best. You also never ansered to my last post.
You can't touch a processor(S) with more cores if you lack the cores to perform with. Woodchest doesn't have the multitasking power but does the performance. Doesn't mean its better. Nothing intel can make can beat a 4P system. AMD's are going quad core to make it worse. 32 cores on a 8P Server and 16 cores on a 4P server.
So how can even 2 woodmistakes with quad cores even out perform the other 8 cores if going agenst a 4P system. Highest end intel blade server vs a half range opteron blade witch fits into a rack of useouly 16 or 32 same as a Intel. Blows old servers out of the water. Yet intels only use 2P systems witch limit it big time.
You can simply not compare a 2P system to a 4P system because they are in 2 totally different ball games. Your trying to say a system can scale like a intels when it simply cant... (A Legecy FSB does not scale limits the cpus full power and speed. Useless 2 processors have to share the same bus makes it impossible to scale.) On the other hand in hyper transport. Each cpu and bridge and socket has its own seprate hyper transport.
Allows it to scale and gane performance with NO BOTTLE NECKS such as in Intel 2P systems.
Final conclusion Unless you get 2 woodmistakes or any cpu from intel to perform 100% faster thats the only way to match a 4P system because you lack the cores and the FSB to perform the tasks. Keep in mind I'm talking about quadros on both sides as well. Wow and a quadro 2P system would be very limited on one 1333 BUS. Only what 166mhz per core when AMD's have 2Ghz per socket both ways on 3 Hyper transport links per cpu.
You aren't being very smart if you still think anything intel has can even compete with that. You know some of you intel fans are TRUE Intel worshipers no matter what anybody says if you truely beleave a woodchest can beat a platform with more cores that can scale. Your hopeless.
Google and yahoo went with 4P and 8P systems from AMD for a reason. They may not use them for 2P systems or may do mixes of intel and AMD. But they did !!!NOT replace WHAT THEY JUST BOUGHT FROM AMD!!! With crappy 2P only systems from Intel. Xeons are limited any 2P system no matter what core it has or how many is limited from Intel until they use CSI then this desussion is over.
AMD wins until Intel can make CSI but by then Hyper transport 3 and 4 will be out. So until then we'll see. But right now AMD rules the server market thanks to 8 way and the upcomming 16 way processors. Thanks to DCA processors scale and thanks to HTL's the processors can scale. IF Intel had Hyper transport their cpus would have a unlimited potensial. But its limited to FSB is what it all comes down to. End of story. Not what I think only fact!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
Again a 2P system would also limit rack space and take up more space. A 4P system can take up less space and do the work of 2 2P systems. A 8P system can do the work of 4 2P systems in one blade. Also save energy and valuble space on the server floor. There are 35w opterons btw. Woodmistake only can go as low as the 80 to 65w mark. AMD's are still the ones that will save the money for a company not Intel.
http://www.channelregister.co.uk/2006/09/14/amd_market_cap/
-raapi
"Google went with Woodcrest based servers - they decided AMD is not reliable enough with their old junkyard processors.
http://www.cmpnetasia.com/ViewArt.cfm?Artid=28813&Catid=1&subcat=9"
Please... Google did not use those machines, Google was going to sell them. And that wasn't because Woodcrest are stable or what, but because at that time Dell only has Woodcrest-based low-end server boxes that are cheap yet with reasonable performance per watt (calling Netburst "unreasonable" is just kind).
I bet internally Google is served better by what you called "junkyard" K8 than by Woodcrest, because Google need scalability, which is what K8's biggest advantage over Woodcrest.
"Actually, snoop filters block pretty much all the I/O traffic from reaching the bus. I/O does cut into the bandwidth of Hypertransport, so this is advantage Intel."
Gosh... you don't know what you're talking about. Do you know what "snoop" is? It's a inter-processor thing, related to the MESI/MOESI protocols. It's nothing particular to I/O. If Woodcrest needs data from a disk, or a network card, or an encryption device sitting on the PCI bus, those data MUST go through FSB and occupy some precious memory bandwidth. Snoop filter can't help that (use the strongest "can't" there).
Besides, HyperTransport switch has snoop filter, too. Oh yes, it has. Just because Intel market it like blue crystals to these websites doesn't make it any special to Bensely (or whatever Intel's platform it is).
Oh man, how rediculous! If you have no idea what you're talking about, why did you even say it in the first place???
The bottom line...
Dell once a marvel of business envied and feared alike for their efficiencies and growth that weren't matched by anyone.
After 10 years HP, Toshiba and others have figured out lean and JIT manufacturing. They have a lot of things dell doesn't and will never; A culture of innovation, a culture of R&D, a culture of support/customer service.
Is it any wonder Dell has fallen on hard times. Dell missed the boat two years ago. They waited and waited to go when AMD was ahead but had no capacity.
Now AMD got no capacity and got no performance chip. AMD is like VIA a few years ago, cheap, avaliable. Just like that ugly girl you knew in high school
Sharikou, which would be superior in overall performance?
fx-62 or the conroe x6800 ee?
Actually, snoop filters block pretty much all the I/O traffic from reaching the bus. I/O does cut into the bandwidth of Hypertransport, so this is advantage Intel.
First of all, all non-DMA traffic, and most DMA traffic (indirectly) touches Intel's FSB. You should hit up google and read up on cache snooping and snoop filters. They do not do what you think they do.
Second, calling Intel's old and slow FSB as a "win" performance-wise goes against the wisdom of almost everyone in the industry. Even Intel is planning on dumping the bus in favor or CSI in the future. The FSB does have some minor advantages in some areas, but performance is not one of them.
Steel Smack said...
"Enumae, that link you provided also has Sony saying it wasn't their fault and their VAIOs aren't having the same problems..."
I know, and it was not a very strong link, but I really wanted to point out that there was no mention of Intel from either Dell or Sony.
"I don't think it's a matter of them ripping off their customers nearly as much as it is trying to screw Intel."
Well isn't it Intel that could screw Dell more right now?
Since AMD can't supply all of Dells needs right now, Dell may wan't to be careful.
But I am still sticking with the thought of Dell exploiting that Core 2 Duo is a good chip, in demand, and people who shop Dell (for the most part) have no clue what the price could/should be.
Hence "Dell exploits computer shoppers"
PS, I got a reply from Charlie at The Inq...
I actually am looking at this, but getting people to talk cost is very
hard. If I can get enough, I plan on writing it. Thanks.
kgmbuoqJust FYI, some time ago when Google was talking about its server park it said it was using 2P machines since they offered the most bang for buck.
Using 4+P isn't reasnoable since they simply don't need to have so much power inside a single box. They'll just use lots of smaller boxes to get the same performance.
It's quite similar with most other applications. If you need some massive databse power then its cheaper to use a cluster of a few 2P machines.
Sharikou, which would be superior in overall performance?
fx-62 or the conroe x6800 ee?
The x6800 is much faster. But the FX62 is also cheaper.
I think AMD as put "all" their processors priced according to its performance vs Intel. At each aimed market. «- I put this just to make sure no one says that the 6600/6700 is faster and cheaper than the FX62.
Sharikou, which would be superior in overall performance?
FX62 is faster on floating point and Con XE 6800 is faster on integer. But both are more than 40% faster than 95% of Intel's chips, Pentium D and Pentium. That's why I said Conroe is the last straw that pushed DELL to AMD.
But both are more than 40% faster than 95% of Intel's chips, Pentium D and Pentium.
But that isn't 95%(run rate) is it? That number is based on a shrinking inventory that is currently flying out the doors at fire-sale prices. Current production is much more heavily biased towards Core Duo and Core2 Duo...
Kalle said: "Using 4+P isn't reasnoable since they simply don't need to have so much power inside a single box. They'll just use lots of smaller boxes to get the same performance."
And why is that? Because you speculate so? That's not nearly good enough!
As an estimation, before 2003, the year Opteron debut, Google already had $250 million worth of commodity servers. Per a paper on IEEE Micro about Google cluster, those servers were dual-CPU Xeon 2GHz.
- Of course they could not have been Opteron, which were just given birth a few months earlier.
- Of course they could not have been 4P because we all knew well Xeon don't scale well, price-performance wise, at 4P+.
- And they could not have been dual-core, nor using DDR2, not even SATA.
Maybe you ought to speculate that all these technologies above don't benefit Google? Nop you can't, can you?
Sharikou said...
"FX62 is faster on floating point and Con XE 6800 is faster on integer."
I know you like to link to SPEC.org, so I did it for you.
SPECfp_rate2000 Intel X6800 = 46.9
SPECfp_rate2000 AMD FX62 = 43.7
SPECint_rate2000 Intel X6800 = 61.8
SPECint_rate2000 AMD FX62 = 44.9
"FX62 is faster on floating point"
... unless you start using SIMD instructions, instead of the FPU stack that has been depricated with the coming of 64bit CPU's, of cource :)
I think AMD as put "all" their processors priced according to its performance vs Intel. At each aimed market. «- I put this just to make sure no one says that the 6600/6700 is faster and cheaper than the FX62.
No, i'm going to say E6400 ... :) just kidding. But really, E6600 alone is enough to beat FX62 in majority of the applications, including gaming.
SPECfp_rate2000 Intel X6800 = 46.9
SPECfp_rate2000 AMD FX62 = 43.7
The SpecFp_rate2000 high score for 1P 2 core AMD is 56.5
You have realize that those specint scores were done with Intel compiler, which cripples AMD processors. I bet if you use something like GCC, the results will be different.
"And why is that? Because you speculate so? That's not nearly good enough!"
Just take a look how much does a single 4P and comparable 2x2P machine cost and it should become more clear.
"You have realize that those specint scores were done with Intel compiler, which cripples AMD processors"
Not if you specify compiler flags as they did there.
Sharikou said...
"You have realize that those specint scores were done with Intel compiler, which cripples AMD processors."
Ok.
Could you or someone here explain the difference between SPECfp_rate_base2000 and SPECfp_rate2000, its to optimize, right?
And why Intel loses performance in SPECfp_rate2000?
Thanks
SPECfp_rate2000 Intel X6800 = 46.9
SPECfp_rate2000 AMD FX62 = 43.7
http://www.spec.org/cpu2000/results/res2006q3/cpu2000-20060815-07000.html
Sharikou you have made a mistake.
SPECfp_rate2000 opteron 2.6Ghz is 56.5
for amd fx62 is biger because amd fx62 is 2.8 GHz procesor.
lol
"FX62 is faster on floating point and Con XE 6800 is faster on integer. But both are more than 40% faster than 95% of Intel's chips, Pentium D and Pentium. That's why I said Conroe is the last straw that pushed DELL to AMD."
AMD is also faster in memory bandwith:) What does it matter, integer/floating point? E6600 is generally better in gaming, encoding, synthetics than the FX62 the EX6800 even furthering the lead.
Intel could really hurt Dell, but they'd be shooting themselves in the foot.
DELL owns 20% of the PC market, Intel owns 73%. This means DELL sells 20/73 = 27% of Intel's production. Once DELL goes AMD 100%, Intel will be losing 27% of its revenue.
Steel Smack said...
"Although this isn't what Sharikou originally though the problem was, it does back up what he now believes, that the Sony batteries would be okay if Dell didn't screw something else up."
Does that also imply that the Apple laptops, not using Intel at the time, had a similar problem with there configuration, thus eliminating Intel from the problem?
"Dell is a known brand, Intel would have to reestablish that market share under a different PC manufaturer. Who is going to be willing to help them?"
Anyone who stands to get cheaper prices from Intel.
I am not saying this will happen or could, but speculating so here is an example...
Apple, who's machines are far better looking than anything Dell offers.
Keep in mind people like me or later generations have a completely different opinion of what style is, sleek and simple is beautiful, and thats just what Apple is.
How much does Dell spend on advertising/marketing, and what portion of there advertising/marketing is supported by Intel (due to the fact if you see Dell you see Intel)?
Lets just say 1/3, I feel it is probably more, but lets continue.
1/3 of the marketing... gone.
Now think about AMD, have you ever seen them on TV?
Do they have the resources to support 1/3 of Dells marketing?
AMD is only on the internet or in magazines, usually only tech related, the general public does not see these forms of marketing.
Intel throws a massive campaign of Apple and Intel commercials and all of a sudden (longer than that, I know) Apple is the new Dell.
Now honestly wouldn't you rather have an Apple (Mac OS X Leopard 64bit with bootcamp) then a Dell (Vista 64bit)?
And besides Dell is going 100% AMD anyways...lol
Sharikou said...
"Once DELL goes AMD 100%, Intel will be losing 27% of its revenue."
Thats assuming that Dell survives and maintains its 20%.
After making my previous post, you know the really long one, I started to research the whole Intel Apple relationship, rather than just speculate.
It would seem Apple is really looking to move in on Dell, HP and Microsoft.
Also maybe Intel did take your advice and realised Dell is in trouble, so they jump ship and buddy up to Apple :)
To be honest.. I doubt anyone reads your blog for facts. Especially not executives. You're not that bright... really. The reason many people read your blog is for a laugh or two. If you actually think you're pretty intelligent then head on over to the THG forums for a debate. My name is ElMoIsEviL, I'll debate you (and beat you). I'm not a fanboi either.. so goodluck.
Too bad I didn't catch this before. ElmoIsIdiot should be your name. You live off of some company that charges you the same as anyone else. You THG boys are the lowest form of life.
So says The Baron.
edward said: "Snoop filter can't help that (use the strongest "can't" there).
Besides, HyperTransport switch has snoop filter, too. Oh yes, it has."
Edward, I don't debate people who use an impolite tone. If you'd like to learn why I'm right, find some manners and solve your own contradiction first. If snoop filters don't work, why does Opteron use them?
anonymous said: "Second, calling Intel's old and slow FSB as a "win" performance-wise goes against the wisdom of almost everyone in the industry."
I never made a comment on the performance of the FSB, let alone call it a win. I said that Intel's centralized topology provides an advantage for I/O traffic. PCI traffic has a direct connection to memory through the chipset. Most people recognize this advantage when it comes to Opteron and the integrated memory controller (and for good reason). Snoop filters guarantee that most PCI-memory coherent traffic never hits the FSB. This was scientia's statement which I disagreed with: "Woodcrest has to communicate over the FSB for both I/O and cache coherency"
Kalle said: "Just take a look how much does a single 4P and comparable 2x2P machine cost and it should become more clear."
You just don't admit that you are wrong.
First, Opteron's good scalability means a 4P will perform better than 2x2P, that is, well worth its price. That is not true with Xeon, BTW, and that is the point.
Second, look at how much Google pays for electricity and you'll know why it will use scalable 4P instead of twice the number of 2P boxes.
Third, Google didn't go 4P because 1) Opteron was too new (in 2003) and 2) Xeon didn't scale. But there is no question that today it's better for Google to use scalable 4P boxes.
"Edward, I don't debate people who use an impolite tone. If you'd like to learn why I'm right, find some manners and solve your own contradiction first. If snoop filters don't work, why does Opteron use them?"
Again, Rediculous! So the problem is my tone, but not your materially wrong arguments? The fact is you don't know what you were talking about when you mention snoop filter; it has nothing to do with my tone. You just don't know what you are talking about - snoop filter is orthogonal to the bandwidth of IO.
Plus, I didn't say "Opteron," I said "HyperTransport" that has snoop filter. (Yes, Opteron uses ccHT, but so can any other chip as long as its maker license the technology from AMD.) And why/when does HyperTransport utilize snoop filters, you ask? When there are more than 8 Opteron processors it needs to connect. Snoop filter helps reducing cache snooping traffic; while Woodcrest needs it for more than 2P, Opteron needs it only after 8P.
When you don't know what you are talking about, you don't. I really can't find a different tone to state that. What you said were just rediculous. And if you don't like this tone, don't talk rediculously, okay? Thanks!
Edward said: "Again, Rediculous!...snoop filter is orthogonal to the bandwidth of IO
You'd think someone so fond of saying ridiculous over and over would learn how how to spell it right. Maybe that's the redneck version.
The only orthogonal things are your comments to my post. You seem to think that I/O traffic is immune to snooping. Well then, better watch out next time you send your comments over Ethernet and your text is still in the cache. Perhaps that's why we end up with this stream of garbage from you.
Well then, better watch out next time you send your comments over Ethernet and your text is still in the cache. Perhaps that's why we end up with this stream of garbage from you.
LOL. This was just too funny!! Probably Edward won't understand what you are saying... Keep up the good sarcasm...
"You'd think someone so fond of saying ridiculous over and over would learn how how to spell it right. Maybe that's the redneck version."
What you were saying were clearly wrong. Snooping protocols and snoop filters do not act on I/O directly, but cache and main memory. I am not fond of saying ridiculous (which I sometimes mixed up the 'i' and 'e' keys on the keyboard), but in this case, what you said WAS ridiculous.
"You seem to think that I/O traffic is immune to snooping. Well then, better watch out next time you send your comments over Ethernet and your text is still in the cache. Perhaps that's why we end up with this stream of garbage from you."
No, the snooping that the snoop filters try to filter out has nothing to do with bad guys snooping your sensitive data. The data were never snooped, but their 'states' were. I see that you simply don't understand such microarchitecture detail, and I don't blame you. But please stop spreading misinformation that you picked up but not digested from those amateur websites.
Snoop filter improves multiprocessor performance while maintaining cache coherency. Anything that touches local cache (which is everything that goes into the CPU core) AND is shared by multiple processors/cores could be affected by snoop filters.
Those unfortunately don't include I/O, which is rarely shared among multiple processors.
I appreciate that you corrected my spelling. I'd appreciate more if you'd simply admit wrong when you are wrong. (What if I've kept spinning that "rediculous" is the right word? Wouldn't I've made myself seem really ridiculous?)
"LOL. This was just too funny!! Probably Edward won't understand what you are saying... Keep up the good sarcasm..."
Yes, I don't understand how people could speak so blatantly on topics that they don't know.
You do think the snoop filter could stop bad Ethernet traffic, don't you? I pity you. What you need is beyond my persuasion; what you need is some good education.
I'm terribly sorry.
"First, Opteron's good scalability means a 4P will perform better than 2x2P, that is, well worth its price. That is not true with Xeon, BTW, and that is the point."
Google has hundreds of thousands of 2P machines. Its software is way more parallelized than anything ever created before. Even assuming they get roughly the same performance than any other web/database cluster out there they don't see much of a performance loss with using 2x 2P instead of 1x4P.
One interesting link for you to read:
Google buys 10,000+ Dual Opteron Servers?
Dated tuesday, March 21, 2006
Edward said: "Snoop filter improves multiprocessor performance while maintaining cache coherency. Anything that touches local cache (which is everything that goes into the CPU core) AND is shared by multiple processors/cores could be affected by snoop filters.
Those unfortunately don't include I/O, which is rarely shared among multiple processors."
Well, this is where you are wrong, Edward. A cache coherence protocol is defined between all agents who read/write into the coherent address space. That includes processors and I/O agents. I/O agents can also have caches of their own if that helps you understand better. Picture the Intel chipset as having three buses (two FSB and one PCIE) with two Woodcrests and a third processor which is there to move data back and forth between PCI and memory. It so happens that this third "I/O processor" has direct access to memory and the snoop filters help it go full speed without having to snoop the two FSBs everytime. By contrast, with Hypertransport, I/O data from the southbridge always has to cross a number of links (sometimes more than one if the I/O buffer is in a distant memory) and has to be snooped as well (per chance one of the CPU cache holds a copy). The vast majority of I/O traffic never leads to a hit in the processor caches, which makes snooping a complete overhead. That's why snoop filters work very well for I/O.
"A cache coherence protocol is defined between all agents who read/write into the coherent address space. That includes processors and I/O agents."
No, if an "agent" as you spoke of doens't keep a local copy (cache) of data which could be in modified or exclusive states somewhere else, then it doesn't need run cache coherency protocol.
The "cache" of an I/O device is so far away from the processor core/cache, and the aggregated I/O bandwidth is so much slower than the memory bandwidth, that it is never efficient to snoop the memory traffic, even with snoop filters. While I/O device could have cache and use DMA, the processor always write-through any I/O traffic.
FYI, none of single-processor CPUs run snoop protocols. Do you suggest that they don't work with I/O devices that have caches?
"I/O agents can also have caches of their own if that helps you understand better."
See above.
"It so happens that this third "I/O processor" has direct access to memory and the snoop filters help it go full speed without having to snoop the two FSBs everytime."
And this "third I/O processor" is just the Blackford northbridge, right? Just to let you know, the "snoop filter" running on the northbridge doesn't improve I/O performance, rather it reduces how I/O traffic interfere with that between the processors and the memory.
The reason that such filter can reduce performance degradation is because there was (more severe) degradation for the FSB approach in the first place.
Please don't be confused by those advertising information. Just because you read it somewhere and it sounds logical, doesn't mean it's correct.
"Google has hundreds of thousands of 2P machines."
True.
"Its software is way more parallelized than anything ever created before."
Very likely.
"Even assuming they get roughly the same performance than any other web/database cluster out there they don't see much of a performance loss with using 2x 2P instead of 1x4P."
For the same electricity bill, Google will be served much better by half the number of 4P boxes.
Even for pure parallel performance, 5k 4P boxes will outperform 10k 2P boxes, simply because the former redeuces communication overhead at least 50% (could be way more depending on the algorithm).
FYI, none of single-processor CPUs run snoop protocols.
Edward, you really love to sound like an expert, but every statement you make is so wrong that it's pointless to address more than the little gem above.
Take the task of editing and saving a file. How do you suppose portions of your file which are modified in the processor cache make it to the disk controller during the DMA transfer (memory->PCI)? It's done by snooping and it happens in your desktop/laptop whether you know it or not. Doesn't even matter if the controller has or hasn't a cache (he's just a reader in this instance). When data is transfered to/from a coherent memory location (by a processor or during a DMA), all caching agents must be snooped. And get it out of your mind that snooping is done only by processors.
Peace out. Feel free to have the last (wrong) word on the topic.
"Take the task of editing and saving a file. How do you suppose portions of your file which are modified in the processor cache make it to the disk controller during the DMA transfer (memory->PCI)?"
Have you ever done any low-level programming? You do know that it is required to add a flush() call after you've written to the output stream, don't you? That call will flush the cache, and write through any memory data to the disk.
No snooping will save you if you forgot to do that.
You can have all kinds of imagination in your head. You have have all types of wrong ideas about what is right or wrong. But at the end of the day, what won't work simply won't.
Even for pure parallel performance, 5k 4P boxes will outperform 10k 2P boxes, simply because the former redeuces communication overhead at least 50% (could be way more depending on the algorithm).
who told you that? 5k 4p box will have advantage of its own, and so do 10k 2p system. Some apps run better in one of these system. There is no clear winner as such. I'd say google is better off with more 2p system. Nowsaday, load balancing is done at the gateway, no much communication overhead as what you would claim. you might be good in computing, but trust me, not networking.
Has anyone considered that the smallest form factor for a 4P box is 4U? Google can stick four 1U 2P boxes in that space. Cutting the footprint in half must have some advantages for Google, no?
"5k 4p box will have advantage of its own, and so do 10k 2p system. Some apps run better in one of these system."
If the FSB becomes the bottleneck, then 10k 2p might outperforms 5k 4p. That is not the case with Opteron, though.
Mind to give us an example of application where it runs better on 10k 2p than 5k 4p (suppose FSB/memory bandwidth is not a limiting factor, as in Opteron systems).
"Nowsaday, load balancing is done at the gateway, no much communication overhead as what you would claim."
Even if you have perfect load balancing, to run the load balancing itself on a 10k 2p system you have twice as many boxes to balance; that means twice the computation efforts and communication for load balancing itself.
Do know that load balancing routines can easily use up a significant portion of overall computing power (15%+).
" you might be good in computing, but trust me, not networking."
And why is that? Because I know that load balancing is actually the main source of communication overhead? Or because I'm not convinced that twice the number of nodes, with 1.5x communication hop lengths in average, would perform better?
"Has anyone considered that the smallest form factor for a 4P box is 4U? Google can stick four 1U 2P boxes in that space."
Of course the smallest form factor for a Intel Inside 4P box is 4U. But maybe you didn't notice this?
And allow me to mind you that the price of such a 4P/2U box is lower than two 2P/1U boxes (just calculate both on the MonarchComputer site).
Post a Comment
<< Home