Athlon 64 X2 3600+ for $130
The 2GHZ, 2x256KB Athlon 64 X2 3600 PIB is being sold at $130 in some markets, $20 below the X2 3800+. Assuming the retailer gets 30% profit, AMD sold the chip for $100. There is no way Intel can compete against these CPUs at these prices.
Based on this projected AMD production schedule, we can estimate AMD's 3Q06 desktop revenue. Assume the ASP of X2 is $135, Athlon 64 is $75, Sempron is $55. We have
(2210+950)*135+(2250+1750+20)*75 + 4620*55 = 0.4266b + 0.3015b + 0.254b = 0.98 b
The overall AMD desktop ASP is $83. The total desktop market is 35 million units per quarter. Intel will take the remaining 23 million market. However, if Intel sells at the same ASP as AMD, then Intel's desktop revenue for 3Q06 will be $83 * 23 million = $1.9b, which will be a $1.4 billion drop from 2Q06. You can see there is no way for Intel not to BK as long as there is an oversupply. AMD has lower cost, so it can flood the market at will. I told Intel folks last year that their only way to survive AMD capacity expansion is to cut production and cause a price hike. But, these morons choose the exact opposite approach.
The $0.98 b revenue is enough for AMD to break even. Now, add the Turion and Opteron sales.
posted by Sharikou, Ph. D. @ 8/14/2006 12:36:00 AM
61 Comments:
Is AMD probing the market with this chip?
With success, we should see 4000+,4400+ and 4800+ versions.
Can you imagine 65nm X2 of that size?
AMD will be able to deliver, no... to FLOOD the market!
What really surprised me is how little the chip suffers from the smaller cache. And it overclocks pretty damn good too. No possible reason to purchase a single core anymore.
See www.legitreviews.com they have some good info up.
I think the AM2 Sempron 3000+ needs a little love too :)
Look at post #6
http://forum.lowyat.net/index.php?showtopic=323886
LOL..
More like there is NO WAY AMD can make money at these prices!
Figure 100 bucks to AMD, of which they need to fork over 5 bucks for packaging, 5-10 bucks for testing, 10 bucks to Charter, leaving them with 75 bucks for silicon die cost and their own lossesto pay for. Won't be any money to pay for their executives Ferraris at the expensive health clubs in slicon valley.
Compare this to INTEL selling their Celeron Ds and PentiumDs in their depreciated 90nm factories
More like there is NO WAY AMD can make money at these prices!
AMD's cost per CPU is $40. At $100, the gross margin is $60.
Sharikou, Ph. D said...
"AMD's cost per CPU is $40."
I have heard you state this before, could you post a link or show your source on this one.
Thanks.
Like his claimed PhD
"AMD's cost per CPU is $40. At $100, the gross margin is $60. "
Show me the proof Doctor!
"AMD's cost per CPU is $40. At $100, the gross margin is $60."
Any idea what it costs Intel to produce a CPU?
Intel can’t compete with this technology at this price. Intel’s only counter move would be to bomb the fab that produces them.
AMD's cost per CPU is $40. At $100, the gross margin is $60.
Is that factoring in the R&D costs? If the lion's share of cpu's are sold within two years of their birth, then you want to factor the R&D costs into the equation too.
I'm certain that they're going to make a even a small profit at these prices though, considering that now they compete with low-end Intel chips.
Forget about competing with Conroe though.
I have discussed the CPU costs of AMD and Intel before. AMD's per CPU cost is about $40. AMD's grosss margin was close to 60% and ASP was below $100. It's harder to compute Intel's cost per CPU because it's also making chipsets. But Intel's cost per CPU+chipset combo is close to $100.
There are people thinking that AMD will lose revenue because of the price drop. Wrong. AMD's revenue will increase substantially. Just go back and read a previous post about the ratios of X2 and Single cores.
Anyone here who post comments, especially ones like yours would be slammed for sources.
So, again, sources please.
I went ahead and took a look online and found something you will not like, and you may not even publish this one.
It would seem Intels cost per die is $40.
Granted this is a little older (9-14-2005), but from the article, the people were impressed by...
"Intel's being able to keep the manufacturing costs steady seems pretty impressive considering that fabrication and materials costs have risen during the same timeframe."
We have real numbers saying Intels cost per die is $40, now we need your real numbers.
Here is the [LINK]
Then add super cool lookin fan/heatsink on 775 and all other stuff.
Those Intel coolers weigh about 2 pounds. I think that's some cost there.
OEM processors are sold without a heatsink. Retail processors sold with a heatsink are sold with a price premium (i.e.-HSF costs are passed on to the end consumer).
Additionally, Intel's production is 65nm now, which significantly reduces cost per die over 90nm.
Intel obviously is suffering from the war. Channel stuffing, 10,000+ layoffs, and blue light special on sale of assets and continue market share loss proves profits are down. On the other hand AMD has sold everything it can produce, what does that tell you?
How much proof do you need? Isn’t it evident that Intel has squandered it’s unquestionable past technological edge. Big-O-Chip is dieing, get over it. I agree this slow and painful death is hard to watch. Marketing will only get you so far.
Azary Omega said...
"Plus them are numbers based on intel numbers,"
Could you explain this, the article says...
"According to a new study by research firm In-Stat"
I do not see anywhere on the link that these are Intels numbers.
You also say...
"Take core duo and core 2 due in to the account as well and it might hit 80's."
In looking at the article and they are talking about 90nm, not 65nm which is CD and C2D.
I am trying not to speculate, I found numbers, and the whole point of my post was to get the Dr. to give us real numbers.
Is he just picking numbers to try and make his point, I do not know, but I would like links to his sources.
Also...
"Then add super cool lookin fan/heatsink on 775 and all other stuff."
Cost per die would not include the heatsink, hence cost per die.
Sharikou, Ph. D said...
"I think that's some cost there."
Do you think you could use your superior intelect and find some sources to back your claims?
You know, so us dumb people can learns sumthn'.
+Turion +Opteron, your so right I think that is your in the notebook buying mode one needs to consider Vista comaptability and at the 64bit end. Opteron is a no brainer I see growth from 25% of market to 75% this year alone...
Do you think you could use your superior intelect and find some sources to back your claims?
Well. I am not familiar with metal pricing, you can check out the price for copper and consider the fact those netburst heatsinks weigh 2 pounds.
"Do you think you could use your superior intelect and find some sources to back your claims?"
This was in regards to AMD $40 cost per die, sorry about that.
This was in regards to AMD $40 cost per die, sorry about that.
AMD sold 13 million CPUs in 1Q06. Cost was $553 million. Average cost $42. Please note this is averaged over all CPUs, X2, Sempron, Athlon64, Turion, Opteron.
Thank you.
We see with the relative good performance of the Athlon 64 X2 3600+ (with 256K cache per core) we can see that there is a lack of performance increase vs. cache size for AMD chip architecture.
The lack of performance increase demonstrates that the AMD cache algorithms do not scale with cache size.
This is one reason why AMD needs to rush to increasing the number of cores, either through "4x4" or "KL8".
Because with the original Opteron team gone, it looks like no one at AMD knows how to do real "rocket science" technology.
The cache algorithms should have been improved over time (and over all those steppings) but have not been. In fact, nothing much has been done to improve AMD's chip technology as AMD does not have the "know how" anymore.
As there is not much AMD can bring to bear from the technology side, AMD is fighting Intel on the pricing side, flooding the market with cheap chips.
In many ways, the roles of Intel and AMD have been reversed. Intel has the good technology and AMD has the low prices.
8-ball is on crack... can we smoke what you're smoking?
Opteron team gone? How do you know?
Source??! (besides pulling that out of your a$$)
=====
Let me start little spout that I'll keep repeating as long as 8-ball keeps posting without backup.
In other news, Intel has lost all of their top notch engineers... all ventured off to greener pasteurs as their stocks are as good as toilet paper. Israel engineers assume the future of Intel, as the current Intel designers know nothing expect "SHOW ME MORE CACHE!!!" :)
8-ball, you are kidding, right! Every box maker on the planet in trying to jump in bed with AMD because there technology sucks with small cache. Dell is leaving Intel after 22 years because AMD’s weak algorithm. IBM, Cray, HP, Dell, Sun and others are knowingly going to use an inferior processor because there cheap.
Due to energy concerns and scalability the FBS that needs the monster size cache just to compete is a dieing architecture.
I’m making an honest attempt to help you thru this friend. I have enjoyed a lot of your past posts. Nice try 8-Ball.
there is a lack of performance increase vs. cache size for AMD chip architecture.
That's a major architecural advantage for AMD64. Whenever you can do more with less die area, it's win. Cache is brute force approach. Look at Intel's Tulsa, it needs 16MB (or 24MB) of cache to achieve meaningful performance for MP servers. I would be surprised if Intel can get a good chip off 3 wafers with that kind of die size. CORE2 will suffer the same fate, eventually, Intel will have to add more cache to get some additional performance.
"That's a major architecural advantage for AMD64. Whenever you can do more with less die area, it's win. Cache is brute force approach. Look at Intel's Tulsa, it needs 16MB (or 24MB) of cache to achieve meaningful performance for MP servers. I would be surprised if Intel can get a good chip off 3 wafers with that kind of die size. CORE2 will suffer the same fate, eventually, Intel will have to add more cache to get some additional performance."
With all due respect, this view is nothing more than limited thinking.
Looking at the desktop processors in the market today, Intel is delivering roughly 15-20% more desktop performance than AMD.
This added performance translates into a lot of value for the customer, especially as there is essentially no price premium (purchase or energy consumption) for this added performance.
Intel's mid-range chip equals the performance of AMD's top of the line (and twice as expensive) chip. The customer is saving hundreds of dollars. And Intel is still making a profit.
With transistor density increasing, it makes sense to add the cache and give the customer the added performance. There is simply no reason to not use the added space and do something good for the customer.
Unless you are AMD and cannot make a cache system that scales.
The added cache on a chip will cost less upfront and far less energy over time vs. AMD's additional CPU ("4x4").
AMD is basically saying "If you want better performance, you have to buy an extra processor from us. And hope all your software takes full advantage of threads and is threaded properly. And put money aside for your higher energy bill."
This position is either driven by greed or incompetence, or perhaps both.
For most desktop customers, AMD's position is insanity.
Looking at the desktop processors in the market today, Intel is delivering roughly 15-20% more desktop performance than AMD.
15-20% is so minute. 4x4 on the other hand, will almost double performance. That's compelling. K8L will offer 60% more integer performance and 3x FP performance per core, all with 512KB cache.
...And intc has come up with another winner. Watch out AMD!
http://www.theinquirer.net/default.aspx?article=33660
4x4 is far from "compelling". It is marketing gimmick in which AMD offers a 2P motherboard to the desktop enthusiast crowd. The sad part is, if you go and spend money for a 4x4 setup, it will still be slower than a single socket Conroe on most applications. This is due to the simple fact that a great deal of software doesn't take advantage of the extra cores.
If the performance of 4 cores is something users will find "compelling", why aren't enthusiats buying 2P opteron and woodcrest systems which have been available for a long time?
sharikou,
Wrong as usual!!! It is not Core Duo problem after all..
http://www.theinquirer.net/default.aspx?article=33682
Any idea what it costs Intel to produce a CPU?
according to the doc:
$1000,0000 per chip
"But Intel's cost per CPU+chipset combo is close to $100."
It must be a very advance formula that your PhD sponsor or any of us simple people can not understand it.
LOL
8-Ball, says cache is king. What happens when you’re not running superPi and you have to use your computer for programs in the real world? Multi-tasking files that can’t be accommodated by your LARGE cache.
8-ball, how come Intel is doing away with large cache and fbs 2008 with there next architecture? 21st century is different, more processors is the game. Some CPU’s can relax while the co-processor efficiently works the task with others. I hope your going to participate in the 64-bit world, if not then keep playing old games my friend.
Btw.. Thrashing cache is a severe energy premium on-off-on-off-on-off (x? megs)
4X4 is not compelling and Hezbolla says they won the war in Lebanon. People say the stupidest things.
Err, because 2P Opteron and Woodcrest systems are expensive, and they aren't really suitable for desktop use (as these "enthusiasts" are wont to do)?
I have learned some amazing things here, and I want to thank Sharikou.
I have learned...
...that 20% performance superiority is everything, unless its by Intel's Conroe, then its nothing.
...that the supposed " netburst two pound copper heatsink" is ultra-expensive, even with Copper at 90 cents a pound.
...that Intel die + chipset costs are $100, even though market experts say $40 per CPU die.
...that Intel is "channel stuffing" while AMD is begging vendors to take more processors.
...That its all about power vs. performance, until AMD starts getting its ass kicked, then 4x4 the power whore and her screaming cooling fans are the new AMD answer.
...That AMD fanboys cant admit it takes four Athlon cores to (maybe) be faster than two Conroe cores.
.that 20% performance superiority is everything, unless its by Intel's Conroe, then its nothing.
Dude, AMD maintained a 90% performance advantage over Intel on servers and a 40% performance lead over Intel on desktops for a couple of years.
Some idiot, clearly another highly trained and educted pretender like the said PhD editor said
"The transistor gates have to stay the same, or go BIGGER in going beyond 65nm to 45nm to 32nm as there is this thing called a physical barrier of controlling atoms. (Scaling as we knew it before, is done! END OF MOORE'S LAW!!) Dimensions are being controlled in the order of nanometers and angstrom."
I will tell you know that the same was sad 20 years ago about 0.5um, and 10 years ago about .1micron.
Today gate lengths in production are 50nm and less at 65nm. At 45nm it will scale again. 10nm gates have already been demonstrated. With new tools ability to create more abrupt junction has continued to enable scaling.
Do some research befor you inhale the farts out of some other idots backside.
Moore's law will continue.. It won't be stopped by physics but economics.
AMD has already been stopped by the economics. They can't afford their own silicon R&D and are starting to farm capacity out. Next step, won't be a manufacture at all.
Hector made the wrong bet buying ATI. He should have taken the debt a few years earlier to put more manufacturing so AMD could have grown to 30% market share 3 years ago. They waited to long and Conroe , Merom, Woodcreat just slammed the door on their dick.
Doctor Doctor
Things are getting boring here.. hee hee
Not much BS to spout with INTEL cleaning up left and right.
Man even Tom's hardware anniversary PC is powered by an INTEL chip. Imagine that. The ultimate gaming machine is an INTEL machine...
LOL
LOL I don't know if Sharikou will censor this or not, but funny reading:
http://forumz.tomshardware.com/hardware/AMD-official-test-X2-E6300-ftopict195851.html
Is AMD so desperate they have to fix results now?
"15-20% is so minute. 4x4 on the other hand, will almost double performance. That's compelling. K8L will offer 60% more integer performance and 3x FP performance per core, all with 512KB cache."
Statements about 4x4 offering "almost double performance" are simply NOT true.
I have a quad-core 2P Opteron system and it is not double the performance on any application that I normally use. Or even have on the machine.
It will be faster for a very small subset of applications -- video encoders, web servers, etc -- i.e. applications that have data sets that can be fit to parallel algorithms.
For most things, having faster instruction processing and a hefty load of cache is what counts. And this is what Core 2 delivers.
For most of what I do, day in and day out, I would gladly trade these four 2Ghz cores for two 3Ghz cores and $1000 to put towards faster I/O, faster graphics, etc.
Going to Intel's 3Ghz core would give the customer approximately 3.6Ghz worth of AMD core speed. So it is really 7.2Ghz (standardized) split into two speedy cores vs. 8Ghz split into four slow cores.
For almost all CPU-intensive apps, the Intel approach will deliver real value in time saving to the customer.
4x4 is a deeply flawed decision by AMD. It is 2P marchitecture for the enthusiast market. It offers nothing better than existing 2P systems except a smaller standard motherboard (there are already a number of ATX 2P Opteron systems, but many EATX as well).
With the smaller motherboard, you give up RAM capacity, which makes the entire 4x4 platform incredibly retarded. 4 cores and no RAM! This is a dumb move as "mega tasking" will need plenty of RAM. And those people doing the "mega tasking" won't want random bit errors creeping into their "mega tasks"... but without ECC RAM... you get data corruption on those "mega tasks".
If someone is really a "mega tasker", then the smart move is to get a real 2P system, not some dipshit marketing hack for "enthusiasts". AMD is doing the market a disservice with the very idea of "4x4".
Yeah, down the road... a year or more... if AMD delivers on the vaporware that is KL8... and KL8 in turn delivers market-leading performance, then that is true value delivered to the customer. Let's look at it when it ships, okay?
However, for today, Intel is delivering performance to the customer and AMD is delivering empty promises.
Gross margin ignores capital costs to produce the chip including R&D, fabs, overhead as well as operational support costs like management salaries etc. Poor measure of profitability.
Dude, AMD maintained a 90% performance advantage over Intel on servers and a 40% performance lead over Intel on desktops for a couple of years.
Haha, this is simply a lie. Not that I'm surprised since you are simply a complete liar.
Please post comprehensive benchmarks showing AMD's supposed 90% lead in servers. You're a liar, and I know this, so please don't post a link to a single SuperDev9964 benchmark that shows a 90% lead - trusted, real benchmarks comprehensive in nature, please, you liar.
"With transistor density increasing, it makes sense to add the cache and give the customer the added performance. There is simply no reason to not use the added space and do something good for the customer."
"dumbm0fo said: 8-ball... you should do yourself a favor, and just shut up now before you continue spouting worthless, wrong, misleading information over and over again... To prove that you TRULY have no clue what's going on in the industry!!"
Take a better look at what I said, dumbamdm0fo.
Transistor density is increasing. Witness the move from 130nm to 90nm to 65nm to 45nm, and so on.
This is not "google", this is the bedrock of the semiconductor industry.
If you want to blast me with retarded nonsense, you've done a good job. All you showed me was that dumbamdm0fo doesn't get it.
AMD's chips are architected from the position of a 130nm process (the original Opteron process).
Intel's Core 2 chips are architected from the position of a 65nm process Unsurprisingly, with the additional transistor density, Intel's chips have more cache than chips designed for a 130nm process.
Intel's current Core 2 based chips deliver more performance (and performance per watt) to the customer. They are modern chips. They support the apps that exist today and offer great clock speed and great value.
When AMD puts out K8L next year on 65nm with a lot more cache (including a big L3 cache), are you going to say "told you so, cache is good"??? Are you going to be a good AMD fanboi and flip-flop?
All I see from AMD fanbois is people who cannot think at all. Stone cold fucking retards.
"Intel is delivering roughly 15-20% more desktop performance than AMD."
... but producing only 15-20% of them. LOL.
And now for some lighter side of
Procesors,
http://www.bbspot.com/
News/2000/5/
amd_moron.html
I have a quad-core 2P Opteron system and it is not double the performance on any application that I normally use. Or even have on the machine.
Why do you even HAVE a 2P system if you can't get any use from it then?
It will be faster for a very small subset of applications -- video encoders, web servers, etc -- i.e. applications that have data sets that can be fit to parallel algorithms.
ie. applications that take alot of processing power. For most mundane single threaded applications, it doesn't matter how fast your single core is anyways. Most applications that require alot of number crunching or repetitive tasks ARE parallel (or are being run side by side). If you don't run these kinds of tasks then of course you don't need 2+ CPUs, many people do.
Just because you don't need or want 4x4 doesn't make your opinion fact.
LOL I don't know if Sharikou will censor this or not, but funny reading:
http://forumz.tomshardware.com/hardware/AMD-official-test-X2-E6300-ftopict195851.html
Is AMD so desperate they have to fix results now?
Haha, that is the best thread ever! AMD deciding to do what INTEL has done for years. It's funny to see Intel fanboys whine about something that their "beloved" Intel does.
As one guy in the thread put it "...this topic should be named "Is AMD trying intel-like bench propaganda?"
And the best thing is, that's a link to Tom's Hardware Forums, the most biased forums on the net. Yes, even more so than this blog. At least this blog brings out the Intel fanboys that blindly follow Intel, and will defend Intel to the death.
Oh, and by the way, the reason Sharikou hasn't posted why Intel chips are so expensive to make is because Intel doesn't just make processors. It's a lot easier to find the cost of something when a companies only cost is directly related to that product.
I would bet Intel's average CPU cost is around $60, but that's without a memory controller, so AMD's motherboards are going to be slightly less expensive directly because of that, so you have to factor that into the expense of Intel. Hence Sharikou's guesstimate of $100 for CPU + Chipset. I don't think it's quite $100, but the point is AMD does things more efficiently, they HAVE to, they don't have the resources Intel does, and they are fighting a company that doesn't play fair, well at least they didn't play fair. Now that the suit is on, Intel isn't holding their marketshare anymore, strange, isn't it...
Screw-Ball-Idiot said:
"Transistor density is increasing. Witness the move from 130nm to 90nm to 65nm to 45nm, and so on.
This is not "google", this is the bedrock of the semiconductor industry.
If you want to blast me with retarded nonsense, you've done a good job. All you showed me was that dumbamdm0fo doesn't get it."
It's not nonsense, Intel's transistors in their cache are getting BIGGER compared to 90nm!! Their Logic is going BIGGER as well. You TRULY have NO CLUE what you're talking about. Are you IN the industry? Do you actually even understand Device Physics? Do you understand the concept of reverse engineering to prove it?
If not, then SHUT THE FUCK UP already and you're demonstrating your Googling skills and basing all of your knowledge off a textbook that is ancient game back in the 0.18um or 0.13um days...
To the dude who made the comment of 10nm gates, you are VERY wrong. FINFET's maybe, but the transistor gates are scaling... FACT! Go to IEDM, VLSI conference, you'll get an idea.
Paper launches, Conference papers tauting that shit is all Bullshit! Whether's its e-beam processed, ONE working bitcell, etc...
So, to SCREW-BALL with no clue, take your GED Education and keep "trying" to prove you're smart...
-SmartM0F0
"The transistor gates have to stay the same, or go BIGGER in going beyond 65nm to 45nm to 32nm as there is this thing called a physical barrier of controlling atoms."
Interesting information, though I highly doubt that 'thing' is called a "physical barrier of controlling atoms". ;-)
Also, it doesn't seem likely that Intel's 65nm transistors are bigger than its 90nm ones. Going beyond 65nm, to 45nm or smaller, it might be. But comparing 65nm P4 and 90nm P4, the 65nm one seems clearly smaller.
I for one would like to know detail on how the scale-down works (not) below 65nm. Does Intel's vertical gate has anything to do with this, too (beside reducing body current, of course)?
"ie. applications that take alot of processing power. For most mundane single threaded applications, it doesn't matter how fast your single core is anyways. Most applications that require alot of number crunching or repetitive tasks ARE parallel (or are being run side by side). If you don't run these kinds of tasks then of course you don't need 2+ CPUs, many people do.
Just because you don't need or want 4x4 doesn't make your opinion fact."
I have quite a few 2P machines. Some are servers, some are workstations. Many of these machines were bought before dual core, so the only way to get 2 cores was to go 2P.
In the area of software development, we've not seen any great performance boost with dual core 2P.
That includes video editing and video processing (which does see a good boost from 1 core to 2 cores, but not from 2 cores to 4 cores).
You can look at the fairly recent Xbitlabs tests that show most things today only scale to 2 threads, not 4.
There is no general statement "most number-crunching tasks" can be parallelized. This is not true.
We even have some 1P DC Dells that perform very similarly to DC 2P machines. Yes, there is a small degree of extra "smoothness" on the DC 2P machine, but the performance is very similar. It is not worth the extra cost of the DC 2P system.
It would seem that I am offering an educated opinion and you are offering nothing but your own "what you want to be true".
4x4 is a bad move by AMD. You look at a modern ATX size 2P Opteron board from Asus and you wonder "why is AMD even spending time on this 4x4 marchitecture nonsense?".
AMD needs to spend time and energy on things that really matter. Putting out a 4x4 straw man a year ahead (vaporware) is not going to help them or anyone else.
The entire idea of DC 2P is to move to a new performance level. So far, there are limitations beyond the processors that prevent any but a few from reaching this new level.
We even have some 1P DC Dells that perform very similarly to DC 2P machines. Yes, there is a small degree of extra "smoothness" on the DC 2P machine, but the performance is very similar. It is not worth the extra cost of the DC 2P system.
DELL users have low IQ, that's proven again. You have to understand that Intel's FSB is the bottle neck. You can have 4P DC xeon Tulsa, but all 8 cores share 667MHZ FSB.
DELL users have low IQ, that's proven again.
That must be why AMD fanbois have been SOOOOO excited about the prospect of landing the Dell account for so long. Congratulations- you get a whack at the user base you so desire!
I believe the word you are looking for is "Touche" :)
shakirou said...
You have to understand that Intel's FSB is the bottle neck. You can have 4P DC xeon Tulsa, but all 8 cores share 667MHZ FSB.
we know that intel fsb is a bottleneck in the current intel design. but according to you intel chips aren't usable due to this bottleneck. it seems all websites that do benchmarks come to a different conclusion showing current intel chips supperior to amd.
I would bet Intel's average CPU cost is around $60, but that's without a memory controller
But of course having the memory controller on board results in:
1) increased die size, which has a resultant decrease in yield. No matter the yield, increased die size drops it due to the nature of defect density.
2) increased complexity of test, which impacts cost/die
3) substantially increased pin count- tester cost is heavily driven by pin count, and cost/pin testing increases with pin count (not a linear function). The cost of testing a discrete CPU @775 pins + a northbridge at 400-600 pins is probably (but I readily admit no hard data to support this) no more expensive, and likely less expensive than testing a 1207 pin part.
4)Reduced assembly yield due to tighter bump pitch and increased exposure to defectivity/failure at attach.
So- it has its advantages and disadvantages. Costwise, I doubt it is a win for AMD on that front. My bet is that die cost for 90nm process as a function of unit area is fairly similar for the 2 companies- within a few percent. For 65nm, Intel probably saw early yields that result in a minimal to no cost benefit over 90nm. Based on previous ramps, and the fact that the 90-->65nm transition is considered to be an easy one in the industry (from a process perspective, not design), I would postulate they are seeing cost savings already, and will be at mature yields by year's end. I readily admit that this is speculation (unlike others), but it is speculation based on 17 years (and still counting) in semiconductor manufacturing. Blast away.
That must be why AMD fanbois have been SOOOOO excited about the prospect of landing the Dell account for so long.
I pointed out long ago that DELL will BK unless it goes AMD. AMD is the future and will own 80% of the PC market soon. AMD doesn't need DELL. DELL needs AMD to survive. DELL is just a screw driver co which can be replaced by another screw driver co overnite.
"I for one would like to know detail on how the scale-down works (not) below 65nm. Does Intel's vertical gate has anything to do with this, too"
Vertical gate, trigate, FINFET, etc... relies on the fact that the channel area is effectively increased because instead of only controlling "one side" of the silicon in a conventional planar transistor you now control "2 or 3 sides" (I'm using the word "side" liberally here). This in turn allows more drive current through the channel and thus higher frequency (or lower power). What double/trigates will buy you is that you can potentially use this benefit to counteract the increase in offstate leakage that comes from gate length scaling.
One of the issue scaling down beyond 65nm is you need to also scale the gate dielectric thickness to minimize so-called "short channel effects" like Vt rolloff, etc. This means switching to a metal gate (to eliminate Poly Si gate depletion and/or moving to a higher K dielectric so you can scale the gate oxide from an electrical thickness perspective). Both of these signifcantly effect the front end integration (and in some cases add considerable cost to wafer cost)
Another issue is cost of litho - it is not clear immersion will be ready in time and if you need to do tricks like double exposure (2 litho steps to effectively generate 1 patterned layer for the critical steps); the cost effecitveness of scaling will be reduced.
Then there are the other traditional scaling issues; RC scaling in the interconnect area, metal reliability (EM) and of course general yield.
This is an oversimplification of the issues, but hopefully provides some insight into the major ones...
"It's not nonsense, Intel's transistors in their cache are getting BIGGER compared to 90nm!!"
Huh? Have you looked at SRAM cell size between 90nm and 65nm (SRAM cell = 6 transistors). As the overall cell size is decreasing, I'm not sure how Intel (or AMD or IBM or TSMC or...) can do this if the transistor is getting bigger.
Please define what you mean by bigger are you talking Lg (gate length) or W (gate width) or something else? If you are referring to the end chip size this is obviously due to higher amount of overall cache.
"4x4 on the other hand, will almost double performance."****
*****(the fine print with the appropriate legal disclaimers)
1. Source is Inquirer? (with no mention of SW or actual benchmark data.)
2. Benefit may be on only certain SW.
3. "Almost double" will remain unquantified and is subject to change at the whim of the poster when the actual data comes out ("See I EXACTLY predicted the 4x4 benefit")
4. Reverse hyperthreading required, oh wait, that's a separate Inquirer story from another "reliable" inside source! :)
8-ball, how come Intel is doing away with large cache and fbs 2008 with there next architecture?
Smoking dope? Intel is doing away with FSB, not with large cache. Int 2008, they will have 8 megs of it.
Per your logic, it was a mistake to put any cache on the processor.
Post a Comment
<< Home