AMD quad to release at 2.8GHZ
Folks were dancing. There will be an AM2 quad at 2.9GHZ at launch.
The K10 will make all Intel CPUs sold in a narrow window of $65 to $125.
The following is a pricing list after K10 launch:
K10 quad (AM2) 2.9GHZ: $995
K10 quad 2.7GHZ: $795
K10 quad 2.5 GHZ: $595
K10 quad 2.3 GHZ: $395
K10 dual 2.9 GHZ: $695
K10 dual 2.7GHZ: $495
K10 dual 2.5 GHZ: $295
K10 dual 2.4GHZ: $195
Kentsfield 2.9GHZ: $125
Core 2 Duo 2.9GHZ: $105
X2 3GHZ: $101
Core 2 Duo 2.66GHZ: $110
X2 2GHZ: $65
Pentium D: $(-1)*
AMD FAB36 will be dedicated to K10 and Turion production. I expect dual core K10s be used for mobile. FAB30 will be dedicated to X2 production at 90nm, and it will become a 45nm FAB in 2Q08. Chartered should be converted to 65nm production soon.
Intel will pay $1 for each Pentium D sent to landfill. No one will buy a PC with Pentium D when Core 2 Duos will be dirt cheap. Remember, the CPU is just one component of a PC.
posted by Sharikou, Ph. D. @ 5/07/2007 08:06:00 AM
70 Comments:
Weren't the same folks dancing when the news about reverse hyperthreading appeared on inq?
zWOW!!! Barcerlona @ 2.8 GHZ. I am already picturing the bechmarks. Of course we shall wait and see what will happen.
Looks like these are high GHZ for AMD processors this year. The only question remains. Will the chips be available in high volume?
AMD isn't trying to beat Intel anymore, they aim to humiliate. I congratulate AMD on their unquestionable victory.
AMD isn't trying to beat Intel anymore, they aim to humiliate. I congratulate AMD on their unquestionable victory.
As do I. AMD will open a performance gap that Intel will never be able to close. As soon as K10 launches Intel will have to slash prices across the board. Core 2 Extreme will be sold for a mere $100. With such violent price cuts and AMD's ability to ramp up FAB36 to pump out million of K10 CPUs Intel's BK is inevitable.
Sharikou predicted that Intel will BK in 2Q08 and that is what will happen.
AMD is gonna do nothing but give those lemons to Dell 4free.
2.9 GHz on new superscalar architecture is impressive. Would be interesting to see how it can perform in various benchmarks/applications.
Intel may have to go beyond currently planned 3.33 GHz to counter this, and still, may not be able to match it in specfp_rate.
@penix
That all makes sense now. Before it didn't seem that way.
I would thoroughly like to see how the long waited Barcelona performs.
I hope all the Intel fanbois realise how Intel has been scamming folks with their crappy Core 2 sh*t and a new blue logo.
Whats a k10 and more importantly where can I buy it / read independent release reviews of it?
The K10 will make all Intel CPUs sold in a narrow window of $65 to $125.
I certainly hope AMD comes up with competetive product soon. It is really getting tiring watching you AMDROIDS make pathetic arguments.
AMD's utter and complete beating is getting boring.
Hurry up please AMD.
AMD talking out their a$$ again.
Fusion is a 2008 product. Whoops we mispoke, it's really a 2009 product.
Whoops again, we didn't really mean 2009, it will be somewhat later than that:
http://www.betanews.com/article/AMD_CPUGPU_Fusion_May_Be_More_Gradual_Past_2009/1178553756
Samsung launches Q1 UMPC.
Worlds largest electronics manufacturer says Intel CPU's are better than AMD.
http://www.dailytech.com/Samsung+Q1+Ultra+UMPC+Launches+at+799+USD/article7187.htm
AMD blows another marketing opportunity: Intel to be main sponsor of GDC 2007.
http://biz.gamedaily.com/industry/news/?id=16067
Worlds largest computer maker doesn't offer AMD cpu's on newly announced multimedia computers.
Hewlett-Packard uses Intel instead of AMD.
http://www.eetimes.com/news/semi/showArticle.jhtml?articleID=199300057
Fact: AMD sucks.
Oh, and Doctor? Notice how I post links to reputable sites, and not rumor blogs?
Maybe you should try the same.
“Bubba said...
Samsung launches Q1 UMPC.
Worlds largest electronics manufacturer says Intel CPU's are better than AMD.”
Bubba, are you talking about the toy computer combo calculator?
Heatsink failure causes AMD cpu to destroy itself, Intel designed to prevent failure.
http://www.tomshardware.com/2001/09/17/hot_spot/page2.html
Sharikou: The K10 will make all Intel CPUs sold in a narrow window of $65 to $125.
I dont know what will happen when barcelona releases but currently this is really true to what happened to K8 after Core 2 duo came out!!!
Werent you the idiot touting the cheaper price of its k8 as being good for AMD soo how is this bad for intel??
This comment has been removed by the author.
Heat; You make a good point. I wonder what the real effects of the K10 will be on the price war. Disregard the posted prices in the blog, they are most likely speculation, however I do agree Intel will have to drop there prices if they have nothing to compete with the K10. If the C2D becomes ultra affordable, GREAT.
So, until AMD fully tools to K10... what happens when they have to dump their existing product line?
Oh wow, yeah Intel is in the dumps when it gets rid of old chips, but nah AMD is perfectly fine when doing the same...
“Evil said...
So, until AMD fully tools to K10... what happens when they have to dump their existing product line?
Oh wow, yeah Intel is in the dumps when it gets rid of old chips, but nah AMD is perfectly fine when doing the same... ‘
You make a fair comparison, but you have to admit the X2 is lot more desirable than what Intel was dumping mostly on the third world, which was OK by me. X2 & CD2’s are gonna be cheap.
Ph"ony "D"eduction that is why he is Sharkou "PhD"
Let me do a bit of simple math and tell you all why AMD is going BK in 2008.
The "PhD" got the pricing about right. Now here is the laughable part that is why he is a "Ph"ony "D"octorate!
AMD has at best 10,000 wafer start/month right now going to 20,000 next year. THat only assume they continue their big cash burn equiping the fab with expensive tools. Lets just pretend they are silly enough to continue to hold to their ramp.
10,000 wafers in Q3
15,000 wafers of 65nm in Q4
20,000 wafers in Q1'08
For a Barcebalogna 283mm^2 die that is about 205 die/wafer
Assuming that AMD as two steppings one a true quadcore and one a dual-core chop.
Assuming 50% yield to quadcore and 75% yield to dual core
Assume that they run a 50% split in their fab of quad-core to dual-core. Anything more then that would penalize their dual-core high volume running to a huge yield penalty for their quad0-core.
Q3: 5K wafers of Dual core, 5K wafers quad/core = 6.9MM die in Q3
Q4: 7.5K wafers of dual core, 7.5K wafers of quad core = 10.4 million die in Q4
Q1: 10K wafers of dual core, 10k wafers of quad-core = 13.8M in Q1'07
The estimated total x86 CPU volume is about 60million/quarter. That gives AMD 23% marketshare even if you use their Q1 results. Bottom line is that the new product brings them to Q4'06 pricing levels where they couldn't make money even then! This assumes some pretty agressive ramps of 65nm, agressive yield for quadcore, great speed on a new technology. Sorry even APM can't help ramps like this. they take time and volume none of which AMD has done on 65nm. Please adding charter won't help AMD. IN a price war by the time Charter takes its cut of profits AMD will be selling at a loss to Dell for every CPu from Charter...
The volumes at the top bin will be minscule a few thousand / week as they skim.
The bulk of the volume will be sold at the 300-500 dollar point for quad-core and the 200-400 price point.
Result is too little revenue to compensate for the huge cash burn!
INTEL will respond by dropping C2D to Celeron using their fully deprecicated factories selling them for 150 and less. The "Ph"ony is right. Its the natural evolutoin that C2D will become celeron. The top end C2D and quacores will drop into the 200-400 buck range. THen they will unlease Penrym at 400-1k buck range. Penrym will be far more competitive then INTEL's position with the Prescot Smithfield/Paxville offerings two years ago. Thus INTEL will continue to grow Market share and you'll soon see AMD dropping prices as faster Penryms and Nehalems appear in 2008.
The bottom line is AMD will only have betewen 10-15 million CPU capacity / quarter in 2008 while INTEL with 4 45nm factories and 4 65nm factories will be able to do about 300 million. THey will lower prices, still make between 1.5-2billion a quarter and AMD will end the year losing 200-300million a quarter.
In 2008 AMD will be another 4 billion in debt and won't be able to fund its 45nm technology, 45nm ramp. It will be carved up as it goes BK in 2008.
Sorry fanbois, AMD is finished! Do the financials and tell me where I'm wrong, LOL
In more interesting News, Jack Thompson is getting disbarred.
More trouble for Intel. They had better migrate to 32nm die shrink and bloat the cache even more to 16MB cache to keep up with the superior architecture of AMD native quad core.
Intel Hack Jobs
- hyperthread phony multiprocessing
- netburst low instructions per cycle
- front side bus bottleneck
- bloated cache to make up for architecture deficiences
- phony quad core
All Intel netburst CPU's were made extinct by AMD single core. AMD native quad core will extinct Intel Core 2 Dud, phony Quad and anything soon to be released.
Evil:You make a fair comparison, but you have to admit the X2 is lot more desirable than what Intel was dumping mostly on the third world, which was OK by me. X2 & CD2’s are gonna be cheap.
That would be true but what you fail to see is that intel has more than double the manufacturing capability of AMD and K10 more than likely will be a paper launch for the first quarter or soo not to mention there is no sight or mention of the desktop release yet.
If AMD's past is any indication AMD likes to jack the price of their chips up when they have a superior product unlike intel which had a full range of c2d at every price point.
If they release barcelona at a very high price point this will hurt them since intel is planning on having quad core to main stream by Q3 which means quad core for $250 to the mass market...
The current AMD X2 5000 stacks up quite nicely against E4400 now the price has come down.
Of course AMD are not making much money on it (or at least not as much as they would like) so I do have to wonder what is going to happen when FAB36 is making as many K10's as it can and FAB30, without the upgrade this year is making a few 90nm chips .. who makes all the 65nm chips ..Chartered? But they must make a profit as well so that takes money out of AMD's pocket. Maybe someone can explain whether FAB36 can do K10 and 65nm K8 also, I am not sure what the split is. Thanks.
It is good news indeed if K10 can be 2.9Ghz on launch, scaling has always been an issue with AMD new processes, but looks good this time. I doubt the price of the Intel Quad will go down to $125 though, wish it would, I would definietely get one then and give up desire for the extra grunt of the K10.
R600 and K10 both need to be good popular products for AMD to bottom out of their current downward trend. It's looking pretty rosy for K10, not so sure about R600. K10 is not going to bankrupt Intel but it should help stabilise AMD and hopefully give them enough cash to develope future generations in a timely manner.
bubba said...
Samsung launches Q1 UMPC.
Worlds largest electronics manufacturer says Intel CPU's are better than AMD.
I say the worlds largest electronics manufacturer will say anything to get cheaper parts.
I say bubba is a clueless FUD spreader!
--
Sharikou, what the hell are you doing letting bubba spread old 2001 articles that aren't even on subject?
Newcomers like 'NT78stonewobble' will read that article (not checking the date) and then will go tell 10 of his friends that he heard the K10 cracked due to heat!
It's Intel FUD, but it has worked for years!!
I couldn't tell you exactly by how much, but a price drop by Intel is inevitable. Eventually followed by AMD :(
But people, you've got to remember that Intel posted an increase in revenue in the last quarters, but this was NOT due to stronger sales.
This was due to the week U.S. dollar. Where 80% of U.S. companies, such as Intel which has massive amounts of sales outside the U.S., have artificially increased their sales in the last few quarters by 15% to 22%.
Many honest companies have put it out front in their statements, while most companies buried that information deep within their reports.
So if Intel drops their prices, plus the U.S. dollar picks up in the near future, you can probably expect Intel to have 30%+ revenue drop.
What are you talking about? The UMPC article is current.
AMD BK Q2'08.
AMW:R600 and K10 both need to be good popular products for AMD to bottom out of their current downward trend. It's looking pretty rosy for K10, not so sure about R600. K10 is not going to bankrupt Intel but it should help stabilise AMD and hopefully give them enough cash to develope future generations in a timely manner.
Things looked pretty rosy for the R600 as well......till the actual benchmarks came out!!
Lets hope Barcelona is not another one of AMD's recent fiascos but they win the "Powerpoint Company of the Year" hands down something AMD fanbois can be proud of.
@Jeach...
"Newcomers like 'NT78stonewobble' will read that article (not checking the date) and then will go tell 10 of his friends that he heard the K10 cracked due to heat!"
Well please answer me, instead of jumping to false conclusions.
Where can I buy the k10 and / or read independent reviews of it?
wakakakakakaka.. QC AMD 2.9Gzh on 4x4 MB will be easily frags intel QX6800 by 100% at the same clock speed. Funny..
sharikou
"The K10 will make all Intel CPUs sold in a narrow window of $65 to $125. "
Considering that K10 will not be a volume product for several quarters, does that mean that far inverior K8 will be sold at considerably lower proce levels than C2? Somewhere around $30-60, perhaps?
mi7chy
"- hyperthread phony multiprocessing"
Free 20-30% performance boost in many situations.
"- netburst low instructions per cycle"
They did a too optimistic prediction on improving transistor leakage.
"- bloated cache to make up for architecture deficiences"
CPU is designed to be maximally optimal, having bigger cache is part of it. You could say the same about AMD that it needs low-latency RAM to hide its inefficiency.
"- phony quad core"
But cheap and works well in most cases. AMD will have similar thing soon also.
wakakakakakaka.. QC AMD 2.9Gzh on 4x4 MB will be easily frags intel QX6800 by 100% at the same clock speed. Funny..
Wow! Really?! You mean two AMD quad cores CPUs in a 4x4 motherboard will be faster than one Intel quad core CPU?! Who would have ever guessed that?!
Budda,
Worlds largest computer maker doesn't offer AMD cpu's on newly announced multimedia computers.
Hewlett-Packard uses Intel instead of AMD.
Fact: AMD sucks.
Well the Core 2 Duo was beat up pretty hard by AMD Athlon X2.
Oh, and Doctor? Notice how I post links to reputable sites, and not rumor blogs?
Maybe you should try the same.
Here is a link from one good reputable site that has done RealWorld benchmarks of one Core 2 Duo 6300 VS X2 4800+ and the Core 2 Duo was beat up pretty hard.
“Game benchmarking was also heavily weighted toward the AMD configuration, as it won five out of the six real-world tests.”
HP Intel VS HP AMD
Wow! Really?! You mean two AMD quad cores CPUs in a 4x4 motherboard will be faster than one Intel quad core CPU?! Who would have ever guessed that?!
You should to be more specific while saying the FASTER word..
FASTER = 50% + 50%
= 100%
LOOOOOL..
"- hyperthread phony multiprocessing"
Free 20-30% performance boost in many situations.
Bollocks. That kind of performance boost, if it exists, is very rare. In my previous job, we had to turn off hyperthreading to increase performance and in some cases, to get stability. You will find plenty of reports like these on the Linux kernel mailing list, the centos mailing list and other similar places. hyperthreading is a joke. ho ho, you are learning how to deal with clueless managers. I hope you never find yourself explaining your false reports to a competent manager.
According to recent tests done by Toms Hardware, AMD processors consume much less power than the Core 2 Duo even at 90nm VS 65nm.
The results are impressive: although AMD's Athlon 64 X2 5600+ is a 90 nm part rated at 89 W, and the Core 2 Duo E6400 is a modern 65 nm product rated at 65 W, the AMD system required clearly less energy over a period of 60 or 90 minutes.
This another real world test, much like the one done by shaky extreme that I have already posted.
According to recent tests done by Toms Hardware, AMD processors consume much less power than the Core 2 Duo even at 90nm VS 65nm.
That's one of the proofs how obtuse are those core2 stuffs
wakakakka... funny.. funny.. very funny..
Ah, quoting from Tom's Hardware?
I thought Intel owned them? And kept money in Tom's pocket.
Oh I see, when they are showing Intel superior to AMD they are being paid by Intel, but when they show AMD superior to Intel, its OK to quote from them?
What a bunch of hypocrites.
aguia
"Well the Core 2 Duo was beat up pretty hard by AMD Athlon X2."
Are we comparing CPUs or IGPs?
cristopher
"Bollocks. That kind of performance boost, if it exists, is very rare. In my previous job, we had to turn off hyperthreading to increase performance and in some cases, to get stability"
Well, my Java applications ran considerably faster with HT enabled. I know that some server applications do not benefit from it, thats why I didn't say it improves speed everywhere.
Ah, quoting from Tom's Hardware?
clearly, the AMD test machine required less energy, as its idle power draw is considerably lower (58 W versus Intel's 84 W). Most office or multimedia PCs typically spend most of their time running idle or almost idle, which explains AMD's dominance, even with an aged 90 nm processor.
Are we comparing CPUs or IGPs?
Both. You know computers need both to work.
pezal, I know when it isn't spelled out for you by Ph(ake)d, you have problems with basic reading skills, but as I said:
Ah, quoting from Tom's Hardware?
I thought Intel owned them? And kept money in Tom's pocket.
Oh I see, when they are showing Intel superior to AMD they are being paid by Intel, but when they show AMD superior to Intel, its OK to quote from them?
What a bunch of hypocrites.
Silly little fanboi's can't have it both ways. Either Tom's is an accurate source of information, or they are not... take your pick.
pezel:wakakakakakaka.. QC AMD 2.9Gzh on 4x4 MB will be easily frags intel QX6800 by 100% at the same clock speed. Funny.
Will it come with its own nuclear reactor to power it wakakakakakak??
While we are on the subject of quoting tomshardware....here is what they say in the conclusion of that link......
They don't have a chance against the Core 2 Duo, however, if you are looking to encode or transcode digital audio or video. If you are willing to spend more than $200 on the processor, Core 2 Duo still is the only reasonable choice.
Evil dont get frustrated lack of reading comprehension and crying like lil girls is the cornerstone of what it takes to be an AMD fanboi.
heat,
why not, post the rest of the conclusion:
In other words: AMD still provides a great value, but mostly in the mainstream. This conclusion might be even more interesting for people that have to purchase hundreds or even thousands of systems: clearly, the AMD test machine required less energy, as its idle power draw is considerably lower (58 W versus Intel's 84 W). Most office or multimedia PCs typically spend most of their time running idle or almost idle, which explains AMD's dominance, even with an aged 90 nm processor.
Aguia....so i am guessing that you think 4x4 was a dumb idea and are admitting intel's superiority because of the huge energy difference between the two machines........i am sure the rules dont change with you when the company you dont favor does the same thing you are touting............
bubba
"Whoops again, we didn't really mean 2009, it will be somewhat later than that:"
Unfortunately, Richard (from AMD) never said that. The beyond 2009 comment is the conclusion of Scott, the author. Nice try though.
"Worlds largest electronics manufacturer says Intel CPU's are better than AMD."
Didn't you forget to mention that the processor they used only clocks between 600Mhz and 800Mhz? This would be down in AMD's Geode LX range.
"Worlds largest computer maker doesn't offer AMD cpu's on newly announced multimedia computers."
Curious, their current Media Center m8000 series is offered with both AMD and Intel processors.
lex
AMD's volume for 2007 will be about 80 Million. I'm still baffled at how you came to the conclusion that AMD would be $4 Billion greater in debt. AMD's losses should be stopped by Q3 but will probably less in Q2. An additional $200-$300 Million is a better estimate.
I have to admit that I am completely stunned though that you believe that AMD will halt installation of new tooling. AMD will begin installing 45nm tooling in FAB 36 plus 300mm tooling in FAB 30 very soon. Some of the money to pay for this comes from grants from Dresden and some comes from the sale of 200mm tooling.
Half quad core is a ridiculous estimate. Intel will only be 6% quad core on the desktop. Add this 6% plus the server quads and less than 10% is more like it.
heat
"That would be true but what you fail to see is that intel has more than double the manufacturing capability of AMD and K10 more than likely will be a paper launch for the first quarter or soo not to mention there is no sight or mention of the desktop release yet."
Double? Triple is closer. K10 won't be a paper launch but it will be server only at launch. I assume you aren't counting the FX chips as desktop which are Q3. The regular desktop release is Q4 07.
amw
"Maybe someone can explain whether FAB36 can do K10 and 65nm K8 also"
Yes, it can. FAB 36 will make all of the 65nm chips and the first 45nm chips in 2008. FAB 30 won't start using 300mm wafers until early 2008.
K10 won't be a paper launch
Wanna bet? AMD has said so themselves.
No bubba apparently AMD contact scientia before making such decisions so he would no more than Henri.....
Half quad core is a ridiculous estimate. Intel will only be 6% quad core on the desktop. Add this 6% plus the server quads and less than 10% is more like it.
And Core 2 Duo will eat the dual cores for breakfast. The advantage of K10 is not so much in the core. Afterall, it is more of the same core. The real performance comes from the 4x4 crossbar, and increased bandwidth.
1333 MHz is plenty of bus for dual cores, as we all have seen (except for extremely memory intensive apps).
If K10 launches at 2.9 GHz, it may cause some problem for Intel in high-end DP and high-end 4P problems. You and other fanboys were plainly sounding ridiculous claiming that K10 will be a conroe killer at 2.5GHz.
And Scientia, when it comes to ridiculous, you yourself are not above others. Go back to your bog and read the posts where you claim Intel won't be able to launch Penryn at over 3 GHz. Guess what, Intel has already publicly demoed Penryn at 3.33 GHz.
The truth is, now Barcelona is looking a lot more competitive, and will have performance in the ballpark range of where Intel can take C2D. That will probably force Intel to release higher clocked C2Ds. C2D is not clocked higher because Intel does not need to. It is a chip that will easily go to 3.2+ GHz.
I have a C2D E6700. Guess what, I undervolt it to 1.22 V, and then overclock it to 3.2 GHz. And the system does not crash, ever! And if a CPU can give 20% overclock when undervolted by 5%, it is definitely underspecced.
Now I expect Intel to release Penryn at 3.5 to 3.6 GHz. That wills still be well within current TDP of 120 W for high-end QC. However, even at that frequency, K10 might be able to outperform it in specfp_rate.
ho ho said of hyperthreading: Well, my Java applications ran considerably faster with HT enabled. I know that some server applications do not benefit from it, thats why I didn't say it improves speed everywhere.
But you claimed 20-30% increase in performance in many situations which is blatantly false. I seriously doubt your claim of getting 20-30% boost in performance of all your java applications due to hyperthreading.
Some people have had to turn off hyperthreading to avoid java exceptions besides to gain performance. On the desktop side of things, turning OFF hyperthreading usually meant better performance and stability. Server software like databases on hyperthreading aware operating systems absolutely hate hyperthreading.
Hyperthreading is good for video encoding and edge cases in java base software.
20-30% boost in performance in many situations due to hyperthreading? Bah!
Now I expect Intel to release Penryn at 3.5 to 3.6 GHz. That wills still be well within current TDP of 120 W for high-end QC. However, even at that frequency, K10 might be able to outperform it in specfp_rate.
You should consider that the minimum. Intel has a 3Ghz Quad core now. They have further room to increase the clockspeed if the need arises at 65nm. Penryn will scale a lot higher than 3.33Ghz, that's for sure. It wouldn't suprise me if Penryn breaks the old Netburst clockspeed records.
I have a C2D E6700. Guess what, I undervolt it to 1.22 V, and then overclock it to 3.2 GHz. And the system does not crash, ever! And if a CPU can give 20% overclock when undervolted by 5%, it is definitely underspecced.
What you can do with you CPU and with your motherboard is one thing, what every one can do with every CPU and every Motherboard, is other thing. Don’t mix up things.
Hyperthreading is All Hype
One of many Intel hack jobs such as Core 2 Dud and Phony Quad.
Hyperthreading Performance Degradation
http://blogs.msdn.com/slavao/archive/2005/11/12/492119.aspx
Hyperthreading Insecurity
http://news.zdnet.co.uk/hardware/0,1000000091,39198583,00.htm
christopher
"I seriously doubt your claim of getting 20-30% boost in performance of all your java applications due to hyperthreading"
I did in all the apps I could meaningfully benchmark. Those were the ones dealing with lots of number crunching, I couldn't time GUI applications that well. Also just to be sure, those apps were multithreaded, singlethread apps can never get a boost from HT.
"Some people have had to turn off hyperthreading to avoid java exceptions besides to gain performance"
This is not possible. Only reason I can think of why HT can cause exceptions is that the code they were running was buggy. If you claim otherwise then could you link to some articles talking about the same problems?
"On the desktop side of things, turning OFF hyperthreading usually meant better performance and stability"
Yet again I say that HT had no effect on stability. I personally experienced considerably better interactivity with HT enabled since when some program was using 100% of one virtual core the other one could run other tasks (GUI) at the same time. Without HT the PC would simply be unresponsive the whole time.
"Server software like databases on hyperthreading aware operating systems absolutely hate hyperthreading"
From what I know that mostly came from bad task scheduling and replay system built into earlier P4's. Replay system problems are fixed in Prescott. In 2P+ machines it could happen that when you have less threads running than there are virtual cores it could happen that you have several threads running on one CPU and other one be completely idle.
"Hyperthreading is good for video encoding and edge cases in java base software."
Why is Java so much different in any other language? Btw, when Intel did its ray tracing benchmarks a couple of years ago they too used HT since they reported higher performance running heavily optimized C++ code:
Ray tracing goes mainstream
From Figure 6. you can see that we get a >25% overall performance improvement across the board when hyper threading is turned on
Btw, what other applications are know to loose performance on HT-enabled machines? So far it seems as only SQL Server has issues.
"20-30% boost in performance in many situations due to hyperthreading? Bah!"
Is your claim of HT causing instability any better?
aguia
"What you can do with you CPU and with your motherboard is one thing, what every one can do with every CPU and every Motherboard, is other thing. Don’t mix up things."
Have you heard of speed binning?
mi7chy
"Hyperthreading Performance Degradation"
His theory seems to be kind of off there. When thread scans through memory once then it isn't usually kept in caches. Also it shows that MSSQL is badly programmed, such memory scanning shouldn't even try to cache the data and work directly on RAM by skiping all the levels of cache and streaming the data through the CPU.
Also he sais that main problem is that second virtual core trashes the shared cache. Well, doesn't the same thing happen with Core2 too? If it does then where are all those reports of massively reduced performance on Core2?
There are some conflicting reports there also:
"I ran a controlled application test on a 4-way (8 logical) server with HT disabled/enabled at the BIOS level and observed 15-20% improvement with HT enabled"
"Hyperthreading Insecurity"
Did you actually read the thing? It requires the server to be already infected with backdoors/trojans. If it is then there are lots of other ways to steal data that are considerably simplier.
Also the same mechanism works on every single computer that caches or pages memory, that includes pretty much every single server and PC there is.
In short my conclusion is that MS SQL server is kind of buggy and doesn't work well on earlier P4's with HT enabled. It should work fine with Prescott and newer Netbursts. The security flaw is not a problem.
Also just to be sure, those apps were multithreaded, singlethread apps can never get a boost from HT.
And single threaded java apps suffer a performance degradation at that. multithreaded java apps have to be very multithreaded to see the 20-30% boost.
"Some people have had to turn off hyperthreading to avoid java exceptions besides to gain performance"
This is not possible. Only reason I can think of why HT can cause exceptions is that the code they were running was buggy. If you claim otherwise then could you link to some articles talking about the same problems?
Articles? Sorry, no articles. Experience and similiar reports on the centos mailing list (not in regard to java though). java exceptions were also hearsay...cannot say whether it is because of the code or because of the underlying OS on which the jvm is running.
From what I know that mostly came from bad task scheduling and replay system built into earlier P4's. Replay system problems are fixed in Prescott. In 2P+ machines it could happen that when you have less threads running than there are virtual cores it could happen that you have several threads running on one CPU and other one be completely idle.
Too bad I never got any Prescott's to play with in my side of things then.
Why is Java so much different in any other language?
Are you rhetorically asking why java can get 30% boost with hyperthreading in heavily multithreaded java apps?
what other applications are know to loose performance on HT-enabled machines? So far it seems as only SQL Server has issues.
DB2, postgresql, mysql and on Linux at that too. Anything on Windows gets coloured by Windows' great stability...
"20-30% boost in performance in many situations due to hyperthreading? Bah!"
Is your claim of HT causing instability any better?
Sorry, I trust what I see and likewise what others on the centos list report.
HT is only useful in specific cases and not generally. As you say, Intel had to fix their HT of which finally appears in Prescott but that only helps those specific cases.
http://www.matrixlist.com/pipermail/pc_support/2005-January/000126.html
A 'boost' in a multi-threaded Java application? And your going to tell me that you ran it using Sun's JVM right?
I don't know about now, but a few years back Sun's JVM only used Green Threads. That means that it has a one-to-many threading model. Which means that at any one time, the application only runs in one process, simulating parallelism.
This is why your capable of running multi-threaded Java applications on DOS. But yet DOS is not threaded or a multi-processing OS.
Unless Sun made improvements to its threading engine or you were using an other JVM, this is total bullshit...
Unless Sun made improvements to its threading engine or you were using an other JVM, this is total bullshit...
There is a huge improvement in performance (load up time don't count) with the latest 5.0 java platform from Sun and its jvm (1.5.x) and native thread support has been available at least on Linux since 1.3.x.
christopher
"And single threaded java apps suffer a performance degradation at that."
How many times do I have to say that? No, they do not. I never saw such degradation and believe me, I have tried it with and without HT.
"multithreaded java apps have to be very multithreaded to see the 20-30% boost."
Is two threads on two virtual CPUs enough to be very multithreaded?
"Articles? Sorry, no articles. Experience and similiar reports on the centos mailing list (not in regard to java though)."
Links to those ML posts would be good enough also.
"java exceptions were also hearsay...cannot say whether it is because of the code or because of the underlying OS on which the jvm is running."
So from those options you choose that CPU is to be blamed.
"Are you rhetorically asking why java can get 30% boost with hyperthreading in heavily multithreaded java apps?"
No, I'm asking why you think that it is easier to get performance increase on HT with Java than with any other language.
"DB2, postgresql, mysql and on Linux at that too"
What do you mean by "linux"? My Gentoo box compiled stuff considerably faster with HT enabled.
"Anything on Windows gets coloured by Windows' great stability"
You are joking, right?
"Sorry, I trust what I see and likewise what others on the centos list report. "
Links to those ML posts, please.
"http://www.matrixlist.com/pipermail/pc_support/2005-January/000126.html"
Too much false information there.
jeach!
"A 'boost' in a multi-threaded Java application? And your going to tell me that you ran it using Sun's JVM right?"
Yes, starting from 1.5 betas. I haven't tried the scaling with 1.4 or earlier versions.
"That means that it has a one-to-many threading model"
No, this is not what it means. Java can use green threads in places where there are no multitasking operating system but if there is it uses regular OS level threads.
"This is why your capable of running multi-threaded Java applications on DOS"
There is a DOS version of Sun Java? Where can I get it?
Something makes me think you are pretty much clueless when it comes to Java or programming in general.
How many times do I have to say that? No, they do not. I never saw such degradation and believe me, I have tried it with and without HT.
Wonderful. Even when Intel themselves have documentation out to inform developers how to work around the performance degradation problems?
"multithreaded java apps have to be very multithreaded to see the 20-30% boost."
Is two threads on two virtual CPUs enough to be very multithreaded?
Heh, depends on the dumb OS. If the OS cannot handle the 'virtual cpus' then you better hope there is a work around or turn hyperthreading off.
"java exceptions were also hearsay...cannot say whether it is because of the code or because of the underlying OS on which the jvm is running."
So from those options you choose that CPU is to be blamed.
When you have to work around the enabling of hyperthreading whether for stability or for performance, yes, the cpu is to blame since its implementation is obviously flawed and this can be seen in the fact that Intel engineers have to put out documentation on how to work around the problems.
No, I'm asking why you think that it is easier to get performance increase on HT with Java than with any other language.
The programming language makes no difference. The important thing to gain performance is making it multithreaded. Any multithreaded application has a chance to get performance so long as they do not overload the shared cache (guess why databases do not benefit).
"DB2, postgresql, mysql and on Linux at that too"
What do you mean by "linux"? My Gentoo box compiled stuff considerably faster with HT enabled.
What do you think 'on linux' meant? db2, postgresql and mysql running on Linux (which became a very hyperthreading aware OS) do not like hyperthreading. That is my point.
"http://www.matrixlist.com/pipermail/pc_support/2005-January/000126.html"
Too much false information there.
I am sure you have the qualifications, knowledge, inside information and experience to diss a proven electrical and electronics engineer who has been in this field for over two decades.
Fantastic ho ho. You are defending a feature that only applies to the flawed Netburst architecture because it enables putting stuff into the pipeline, which would otherwise be empty and thus doing nothing, and was implemented poorly as you clearly know. The implementation 'fix' in Prescott is also a laugh and hyperthreading does not exist on the Pentium-M, Core Duo and Core2Duo because it does not offer any benefit at all. A true Intel fundamentalist who blinds himself to every engineering flaw made by Intel besides their criminal business practices.
Do continue to defend and promote technology that Intel have themselves scrapped.
christopher
"Even when Intel themselves have documentation out to inform developers how to work around the performance degradation problems?"
Was that for < Prescott or >=Prescott? Also a link to that document would be good.
If the OS cannot handle the 'virtual cpus' then you better hope there is a work around or turn hyperthreading off."
Every single OS can use those CPUs without problems and with only single physical CPU there is no difference on what OS you are using. In my case I was using XP Pro SP2 and Linux with 2.6.9 kernel.
"When you have to work around the enabling of hyperthreading whether for stability or for performance, yes, the cpu is to blame since its implementation is obviously flawed and this can be seen in the fact that Intel engineers have to put out documentation on how to work around the problems."
You yourself said it wasn't clear what to blame, the code, the VM, the OS or CPU. How did they eliminate all the other things and found out that the CPU was the problem?
"Any multithreaded application has a chance to get performance so long as they do not overload the shared cache (guess why databases do not benefit)."
Then why did you say that specifically Java can benefit from HT?
"What do you think 'on linux' meant?"
I was just a bit confused since you made it sound as Linux has something to do with it and OS itself looses performance when running on HT CPU. Thanks for clearing that up.
"db2, postgresql and mysql"
Can you give links to cases where those applications have problems with HT CPUs?
When he sais things like "Intel uses this because they have yet to build an x86 CPU with register renaming, out-of-order execution and other goodies, unlike AMD" it makes me doubt of his knowledge about CPU internals.
Also:
"HyperThreading is _not_applicable_ to newer CPUs -- only old CPU designs like Intel's 12-year old 7-issue i686 (Pentium Pro)"
What about Power series, especially Cell an Xenon? Are they some decades old CPUs that are not exactly capable?
"You are defending a feature that only applies to the flawed Netburst architecture because it enables putting stuff into the pipeline, which would otherwise be empty and thus doing nothing, and was implemented poorly as you clearly know"
I'm not defending it, I'm just trying to fix the FUD that is spreaded about it. The first version was kind of bad but the updated one was considerably better.
"The implementation 'fix' in Prescott is also a laugh and hyperthreading does not exist on the Pentium-M, Core Duo and Core2Duo because it does not offer any benefit at all"
What about Nehalem, a CPU that is considerably more efficient than Core2. Would it benefit from simultaneous multithreading?
Btw, I'm still waiting for links to those mailing list threads. Without them it is kind of difficult to believe your claims.
When he sais things like "Intel uses this because they have yet to build an x86 CPU with register renaming, out-of-order execution and other goodies, unlike AMD" it makes me doubt of his knowledge about CPU internals.
When you get yourself an electrical and computer engineering degree and also directly consult for Boeing Integrate Defense Systems, NASA/USAF, EMC, Walt Disney, CMP Media or equivalent, feel free to diss him. You rate yourself too high.
ho ho: Btw, I'm still waiting for links to those mailing list threads. Without them it is kind of difficult to believe your claims.
I don't care whether YOU believe. This is to show that I am not just bad mouthing Intel Netburst heaters for no reason. Since you do not appear to want to do your own testing (20 machines you say in your company? oh you don't have the auhority? too bad.) you can very well do your own search for those who have had problems running db2, postgresql, mysql with hyperthreading on and likewise the Intel documentation.
I am not going to waste my time digging those up for you or for others. You can all very well either do your own tests or your own research. All review sites should be taken with a pinch of salt.
http://lists.centos.org/pipermail/centos/2005-January/044343.html
http://lists.centos.org/pipermail/centos/2005-January/044344.html
http://lists.centos.org/pipermail/centos/2006-January/059274.html
http://lists.centos.org/pipermail/centos/2006-January/059276.html
http://lists.centos.org/pipermail/centos/2006-January/059249.html
http://lists.centos.org/pipermail/centos/2006-June/065380.html
http://lists.centos.org/pipermail/centos/2006-September/070230.html
http://lists.centos.org/pipermail/centos/2006-September/070233.html
Bye Intel fundamentalist.
christopher
"When you get yourself an electrical and computer engineering degree and also directly consult for Boeing Integrate Defense Systems, NASA/USAF, EMC, Walt Disney, CMP Media or equivalent, feel free to diss him"
Since when does a degree give you the right to spread false information?
Btw I do have a degree in CS, I graduated last year.
"20 machines you say in your company?"
More like 50 desktops and 20 servers. It is a small web app development company so we should know when some DB is getting bad performance due to the things you describe. We use only Linux and don't use MSSQL, perhaps thats why we haven't seen any problems.
"oh you don't have the auhority?"
I do, though we have upgraded from those HT P4s some time ago, we are now on C2, X2 and some Opterons.
"you can very well do your own search for those who have had problems running db2, postgresql, mysql with hyperthreading on and likewise the Intel documentation."
Unfortunately I can't because I don't have access to those CPUs any more. I used to own one a few years ago but don't any more.
Now to the links. I read all the threads from start to finish, not just the postings uyou linked to.
1) Basolutely no information of why exactly he thinks what he sais. I personally have used HT CPU's with Linux for almost two years and had zero problems.
2) no valuable information there on the current topic. Also covered in 1).
3) They were describing problems with 2.6.9 series kernel. I've used the same version on my box for several months and had exactly zero problems. I've ran all sorts of different stuff from simple games to Oracle, mysql and apache. Last two were running constantly because I was hosting some stuff, mysql was also used by my media player. Also see 4).
4) Reading the repilies I saw this:
"While I haven't actually measured the performance of a machine that has it enabled vs. disabled, I have noticed that the same machine with hyperthreading enabled responds much faster to SSH logins etc. when the machine is under heavy loads. Given that I've had no problems with using HT, I'm happy to have it."
"My experience is pretty much the same, the increase in performance varies - usually around 10-30% (depends on how well optimized programs are), but it greatly improves responsiveness (the simple fact that when one program is using up all of the cpu there is still that 20% left for other tasks - and 20% of 2GHz CPU is still quite a decent amount of horsepower)..."
Also found this:
"Disabling HT only solves it on single processor systems."
That implies it is not exactly HT problem but SMP problem in general.
"In my case it fixed our dual processor systems (Dell PE2650) but to be honest I was never quite sure if it was an OS patch or HT'ing"
Emphasis mine. Note that it is likely that kernel had the SMP fixes.
"Did't solve for me on a Dual Xeon. The problem was reduced, happens less frequently, but still happens :("
Again, not fixed, only reduced by turning off HT. Does that mean all the Intel 2P+ boxes are flawed and do not work correctly with more than one CPU?
"I disabled hyperthreading... it is still hanging."
That means HT was not the problem.
And finally response from the original author:
"We've narrowed this down to the serial port. Whenever we have anything plugged into the serial port, the system hangs."
"All our CentOS 4.1 systems have this problem [with serial ports] -- regardless of hardware. Disabling mgetty allows us to use the systems after plugging in the serial connection."
In the end it seems as HT was not the problem for the original poster and others reported that disabling it did not solve their problems. Kernel fix might have done it for one person.
5) covered in 4.
6) Replies:
"I have a x86-64 system with one Xeon(TM) CPU 3.00GHz running CentOS 4 with 2.6.9-34.0.1.ELsmp (hyperthreading) as a production PostgreSQL 8.x database. No probs so far."
"I run a dual xeon machine with HT enabled, and see no issues. Good machine."
"Four x86_64 systems, all dual Xeon with hyperthreading enabled: one dedicated MySQL server, two web servers, one combined web + MySQL server, running CentOS 4.3 (2 systems) and Debian (2 systems) One x86 system, also dual Xeon with hyperthreading enabled, web server (Debian).
Everything has worked flawlessly so far."
Seems as the one who had problems couldn't prove it but several others were reporting their production servers worked just fine.
7) Yet again, no real information there. Only seems as the kernel has a buggy SMP support. People are just stating that it can cause problems in database environments but they provide no proof. Also, see 6) about DB performance in real-world situations.
8) covered in 7
So in conclusion it seems as disabling HT does not help lots of people meaning that SMP support in general is kind of buggy. It works perfectly fine for most people so it can't be a general problem. Nobody gave any proof of their claims that HT can lower performance* and they only said it might. My guess is they have read that article about MSSQL and think that it works the same on every other DB also.
*) I'm not saying it never reduces performance, I'm only saying that there are rare cases when that happens and it was not the case in any of the links you gave. So far I know only about MSSQL and that is mostly because of badly designed software. If they would fix that problem their DB would work faster on every CPU, no matter if it has shared caches or not.
Basically those links did nothing else but proved my point that HT is not to be blamed most of the time, certainly not in the cases you linked to. So can you give me any links to threads that really were solved by turning off HT? Though with single CPU it might not show much when the SMP support is buggy. I'm not sure why the kernel had a buggy SMP support because as I've said I've used HT CPUs with Linux starting from early days of 2.6 up to around 2.9 when I upgraded my last HT capable CPU and never saw any problems. Perhaps I was just lucky during those few years.
So, will you try again and give me some real proof or will you simply continue to give links proving my point?
Since when does a degree give you the right to spread false information?
Btw I do have a degree in CS, I graduated last year.
CS != Electrical & Computer Engineering. The guy also has words against AMD design and regularly has to amend his comments (on the spot, not after being 'exposed') so if he said anything wrong, it would just be another mistake given his tendency to make really long posts.
So, will you try again and give me some real proof or will you simply continue to give links proving my point?
Bah! Read them however you like. Hyperthreading 'technology' was nothing but a hack for Netburst. Linux has worked around its quirks just like they worked around quirks for via chipsets besides Intel fixing their hack job as you have clearly stated.
You find reports of both good and bad experiences of turning on hyperthreading. The truth is, all reports are true. The good ones are more recent. Just about the same period when you mucked about with your Netburst. My isp job was from 2002 so I saw all the bad stuff that you did not. End of story. Geez.
That is why PIII were my friends for them servers that tackled 200 million smtp transactions daily. A 4GB dual Northwood Xeon box did not process much than the dual PIII 800Mhz with 1GB of RAM. The PIII 800 boxes would take 800 concurrent connections but I could not push the dual Xeon box over 1000 concurrent connections due to the poor syscall performance (which was later improved in the Prescott too) and the number of transactions handled was not good enough compared to the PIII boxes to justify buying more of these boxes as mail servers even though it had a 3ware RAID10 array whereas the PIII boxes only had two IDE disks running software raid. I was given 20 dual PIII Compaq DL380 boxes that had been sitting down doing nothing for their previous owners instead.
Sorry for calling you an Intel fundamentalist. Forgot to take your circumstances into account. You would not have seen the bad P4 days.
cristopher
"CS != Electrical & Computer Engineering"
CPU architecture was part of my education, I've designed a simple ALU myself, not to mention all the theoretical knowledge I've gathered over the years by myself. You can be quite sure I've got more knowledge about internal workings of a CPU than most programmers twice my age.
"if he said anything wrong, it would just be another mistake given his tendency to make really long posts."
Are you saying that the part of his talk I quoted before was just a typo from his part? Funny, it was almost at the beginning of the post and most of the rest was based on it. Kind of like this one (offensive).
"Bah! Read them however you like."
I read them exactly as they were. I also pasted the replies other people gave me and did not make up any information on my own. All UI said was directly based on those replies.
You, on the other hand, managed to read some problems with HT out of those very same posts. How did you manage it is beyond my comperhension. One possibility is that you just searched for "HT problems" and posted whatever you found without bothering to read them yourself.
Feel free to give your own interpretation of those threads if you feel that I misunderstood them or presented them wrong.
"You find reports of both good and bad experiences of turning on hyperthreading"
From the bad performance examples I've seen one about MSSQL and nothing else so far. Other things I've seen are just some people telling that you might loose performance on servers by turning on HT but nobody seems to have made public benchmarks.
Also I'll repeat that this explanation of MSSQL performance loss is very questionable. If it were true then Core2 and Barcelona both should have considerable performance loss compared to X2 and even P4D.
"besides Intel fixing their hack job as you have clearly stated"
Intel just made HT on Prescott even better than it was on Northwood. On Northwood there were more situations where HT did not improve performance than there was on Prescott.
"The truth is, all reports are true"
Perhaps so but so far I have seen a single report about reduced performance and lots about improvement.
"Just about the same period when you mucked about with your Netburst"
I had 2.8GHz HT Northwood for over two years before I upgraded to 3.2GHz HT Prescott*, neither had any problems.
*) I basically exchanged the CPU for free, had I have to buy it I'd have gone with K8.
"A 4GB dual Northwood Xeon box did not process much than the dual PIII 800Mhz with 1GB of RAM"
Don't take this personally but this is a problem between keyboard and chair.
"Sorry for calling you an Intel fundamentalist."
I wouldn't call myself a fundamentalist, I'm just trying to fix some of the lies and false statements people make. Suprisingly there aren't many about AMD that are not fixed but there are lots about Intel.
Sometimes truth can hurt. Sharikou said he would ban me once before, scientia has done it to some extent on his blog and I was also harrassed on rubyworks forums. I won't be describing the details about those cases here but I'll say that what I said was quite similar to what I've been talking here. Unfortunately there are many people who simply can't handle the truth.
"You would not have seen the bad P4 days."
When were those bad days? Perhaps around H203-H106? That was the time I used 2.8GHz HT Northwood, 3.2GHz HT Prescott and 2.8GHz P4D 920 in my PC. Would any of these CPUs be from the bad days?
I'd be interested to hear your comment about Nehalem and if HT is a thing that might be useful in a more efficient CPU than Netburst. A little while ago you said that it wouldn't give any benefit on Core2.
Ahh.... a dead thread. Here it is August of 2007 and we can confidently say Sharikou was 100% wrong.
2.9Ghz K10's? hahahaha what a joke... try 2.0 and try 2.0 in SEPTEMBER (if you are freakin' lucky to have a deal with one of the few OEM's to actually get them).
Intel did cut prices... like they were always going to since Intel does regular price cuts. However, Intel still has higher prices than AMD... because people are willing to pay more for quality.
Post a Comment
<< Home