Saturday, November 04, 2006

AMD 4x4 ready to frag Kentsfield with scalable architecture

Four 3GHZ cores+ four GPUs, AMD 4x4 will frag Kentsfield, which is just like two CPUs glued onto a slower bus. With Kentsfield, 4 cores share a 1033MHZ bus, each getting 266MHZ. We know this story ends in tears from 4P Xeon performance.

That's not all. By 2Q07, 4x4 owners will get an upgrade with the Rev H (aka, K8L) true quad. Then you have 8 Rev H cores, thermal managed independently. It will take Intel another 5 years to catch up, and AMD is a moving target.

Intel managed to squeeze the last bit of performance out of Bob Colwell's Pentium 3 microarchitecture, they simply ran out of brain juice in this game of innovation at the platform level. As I pointed out long time ago, communications is the key to high performance modern computing. AMD's core-core, core-memory, core-i/o, proc-proc communications are all Direct. Intel has only managed to establish core-core communication using shared cache at dual core level, it's about 5 years behind AMD.

We never get enough performance. I bought a scanner that can scan 20 pages per minute and create PDFs on the fly. However, if I turn on OCR, it slows down tremendously. Of course, this is mainly a software efficiency problem, but if you got 8 cores, each of them can OCR a separate page, and it should be able to match the scanning speed.

Of course, there is another possibility with Torrenza, instead of running OCR in software, we can have a Torrenza OCR chip, which does the most time consuming algorithms in hardware...

34 Comments:

Anonymous Anonymous said...

Prtender you are totally incoherent in typing and logic.

YOu have read the benchmarks and you can see with the very efficient code optimization of the Core2 and the large Cache the FSB isn't loaded for something like 90% of the applications. Thus the INTEL solution rocks. As to how you can predict the AMD 4x4 is fragging Kentsfield.. please show your data...

Its kind of like the bet I proposed to you a while back that you continue to ignore. I will give you 100:1 odds on your BK prediction... come now.. show you got a pair of balls and show your benchmarks and put your money where your blogg is!

10:07 AM, November 04, 2006  
Anonymous Anonymous said...

No, Altair, Agena, Angina, Vagina comes Q307 at best. First Barcelona chips don't even work.

What exactly is 4x4? Opteron DP without registered memory. Oh boy 4 GPUs, 2 on x16, the other on x8?... That would be over 700W on chips alone!

How in the world can you even scan 20 pages per minute, let alone 10?

10:16 AM, November 04, 2006  
Anonymous Anonymous said...

I just can't wait to see the AMD 4x4 results.

Kentsfield is a joke. 2 Dual Cores glued together. PATHETIC!!!

If they are the leading technology in the microprocessors (So they say) then why can't they engineer a native guad core instead???

They glued 2 Conroes for a reason, just so they can attract people that have no clue how the micro architecture works. "UUUUU, AAAAHHHH, FOUR CORES!"

Intel is simply behind AMD with its FSB as Mr. Sharikou mentioned.

10:38 AM, November 04, 2006  
Anonymous Anonymous said...

Just got my new C2D built and man it is faster than anything I have ever seen! And it is so cool I thought it did not turn on since I could not hear the fan running.

I have also seen a demo for the upcoming Kentsfield and that was another story. It was twice as fast as Conroe..

I used to be an AMD fanboy.. But I gotta say, Intel did a good job with Conroe.. AMD may catch up soon, but till I see it, touch it, and feel it, I will hold on believing

12:11 PM, November 04, 2006  
Anonymous Anonymous said...

It will take Intel another 5 years to catch up

Ehhh?? I'm pretty sure Intel should have an integrated memory controller sooner than five years from now. While Intel needs to catch up on that end, AMD has still yet to catch up with performance, and as you stated, it won't be able to until Q2/07. So, maybe AMD will take the crown back at that time... but then two quarters later Intel will come out with something new as well... Intel and AMD are both moving targets. It will get quite interesting once more specialized integrated cores start getting stuck into CPUs... like "Fusion" and whatever it is that Intel has planned. The details are unclear at this time, however...

12:42 PM, November 04, 2006  
Anonymous Anonymous said...

We know this story ends in tears from 4P Xeon performance.


HAHA. Youre actually comparing the netburst arch to the Netburst.
You are so dumb, if I were youre mother I'd hit you with an uglystick.

AMD has to go with 4x4 because if theyde try something as complicated as a QUAD-core on the K8 arch, it would end up with an processor with 220W power development.

12:44 PM, November 04, 2006  
Anonymous Anonymous said...

Didnt notice AMD had a shared cache 5 years ago

1:02 PM, November 04, 2006  
Blogger Sharikou, Ph. D. said...

How in the world can you even scan 20 pages per minute, let alone 10?


It's document scanner. The most useful equipment I have seen for a while. It is a small machine (like a fax machine), papers just fly through and two side got scanned the same time. PDFs are produced, then OCR can be used to make the PDFs searchble...

4x4 will frag Intel, that's for sure. Con X6800 had about 10% performance advantage over FX62...

1:48 PM, November 04, 2006  
Blogger Sharikou, Ph. D. said...

you can see the memory and the heat pipe block for one of the chipset components, probably the memory controller


He has X-ray eye that can see inside the CPU..

2:26 PM, November 04, 2006  
Blogger Scientia from AMDZone said...

This thread has been nearly useless so far. I can't believe the spin on both sides of the argument.

So, how many FX's are we talking about? AMD's original projection for Q4 was 20K units up from 15K in Q3. Now, AMD only had a single speed grade before but they are going to release 3 speed grades of FX for 4X4. So, let's be optimistic and assume that this triples the estimate from 15K in Q3 to 45K in Q4.

And, how many Kentsfields are we talking about? Let's say Kentsfield is 1% of production. Intel does about 3.5 Million chips per week. 13 weeks per quarter would be 45.5 Million and 1% would be 455K. So, I think it's safe to say that Kentsfield production will be much higher than FX production. I don't think FX can frag Kentsfield with 1/10th the volume.

It's also nice to see that people are still parroting the THG "tests". The level of testing at THG is so bad I don't understand why anyone knowledgable would repeat what they say. After the "server" comparison between an outdated 940 Opteron and a Woodcrest THG even got chewed up on their own Forumz. One of the authors even posted an apology claiming that it was really just a preview of the processors rather than a server comparison. This of course was nonsense since they had used the word server on nearly every page of the article. However, I guess that isn't quite as bad as the review Anandtech did where they used a dual Opteron "server" motherboard where only one chip had memory slots.

All I have to say is that if were actually true that Kentsfield could run full speed with a 1066Mhz FSB speed then Conroe should similarly run full speed with the FSB set at half, 533Mhz. Let's see some tests to prove this. We'll assume that the cache coherency traffic on Kentsfield is negligible and not worry about that. It shouldn't be that difficult to do but I'll bet that no one will touch this one; it's much easier to just keep quoting pointing to poor testing.

If testing actually mattered then testing from one website would never be considered solid. What would happen is that the tests from one site would get reinforced from testing from other sites and eventually you would build up a set of scores that were closely grouped and this would give confidence in the results. This is the way science works. You can publish any test results you want but until someone can duplicate them they don't mean much.

Kentsfield gives pretty good performance. The only problem is that the two separate dies present two separate loads on the FSB which makes it the same as dual socket sharing a bus. This is why Intel is having a hard time increasing FSB speed to 1333Mhz while AM2 and F can already handle the equivalent of 1600Mhz. Of course, it could be that the 64 bit quad bus is hitting limits that AMD's 128 bit dual bus isn't. If this is the case, I'm sure Intel will change this on their next socket.

I'm thinking though that a new socket won't show up until 2009. This will almost certainly be when Intel finally gets CSI out the door and gets a IMC. This shouldn't effect things too much except on 4-way. There will be some performance loss on quad core but not as bad as 4-way if Intel can get the FSB speed up to 1333Mhz. 4-way is a bit different because the northbridge itself will likely be the bottleneck. The problem for Intel is that they have to keep making the outdated Cedar Mill Xeon's until they release the Caneland northbridge. It should be interesting to see how they fit four FSB's plus at least four FBDIMM buses on the motherboard. I'm thinking more layers might be needed.

The biggest gain for C2D over K8 is on SSE. This gain will be gone when K8L dual core is released but this won't be until Q3 07. Until that time, nothing touches C2D on the desktop in SSE performance. At least on servers AMD will be back in the game one quarter earlier in Q2 07. However, that still leaves a lot of time for Intel to potentially take server share.

I'm not sure where the ridiculous 220 watt number came from. I'm guessing that this individual has somehow overlooked the fact that Intel was unable to produce a full speed quad Kentsfield for this very reason. Kentsfield only clocks 2.66Ghz because the power draw is unacceptable at 2.93Ghz. AMD could hit 130 watts with the 65 watt chips if they stayed down one clock.

Amazingly though, the reports are that AMD will release 90nm chips as high as 3.2Ghz. I'm not sure how they'll manage to do this. Also, I'm not sure how their 65nm tech compares as there have not yet been any reports of 65nm chips clocked higher than 2.9Ghz. I guess we'll see.

3:03 PM, November 04, 2006  
Anonymous Anonymous said...

Please tell me where you see the fx62 trailing the x6800 in all benchmarks by an average of 10%. x6800 on average beats the fx series by atleast 20% on average in all benchmarks which is a bigger whooping than the fx gave to the netburst.

To the anonymous person who said kentfield was a joke then pray tell what is 4x4. A dual socket k8 to keep up with a single socket solution......and you say intel is trailing behind......pathetic indeed.

And sharikou your articles continues to live upto the reputation of utter garbage which you have earned over the years. Thank you for the laughter!!

3:17 PM, November 04, 2006  
Anonymous Anonymous said...

"If they are the leading technology in the microprocessors (So they say) then why can't they engineer a native guad core instead???"

Because they can and it is much, much more simple and cheaper. Also for lots of problems dual-chip is good enough. Dual-socket of cource has its merits in some situations but it is also quite a bit more expensive.

"Intel is simply behind AMD with its FSB as Mr. Sharikou mentioned."

Are they behind? Yes
Does it matter for home user? No.

"4x4 will frag Intel, that's for sure"

In a little while ago power usage was a big deal for you. How is it now?

As for IMC, Intel doesn't have it yet and even then its CPU's are rather fast. If AMD gets so much power from its IMC then just imagine what would it do with Intel CPU's.

3:23 PM, November 04, 2006  
Anonymous Anonymous said...

With just 1066 and 266 MHz per core...
A 3GHz Quad is equivalent to two 3GHz 5160s in threaded programs and slaughters it in single threaded.
http://www.hardwarezone.com/articles/view.php?cid=2&id=2102

4:25 PM, November 04, 2006  
Anonymous Anonymous said...

Scientia i keep on hearing you blabbing about about how these professional review sites are not doing the rite reviews. You have yet to show any application that mainstream consumers use that will hamper the FSB of kentfield. Could you

A) produce some evidence to back your claims instead of coming up with things out of ur a$$

B) do the kentfield tests yourself and show the professional site how proper benchmarking is done.

Mind you if you choose option B it has been tried before with sharikou when the woodcrests came in (still waiting on those results pretender) and he shut up after he could not back up any of his claims about the supposed benchmarking he was going to do.

4:40 PM, November 04, 2006  
Anonymous Anonymous said...

Great. I can just take my current Core 2 rig and pop a Kentsfield Quad core right in there. 130W of power is more than I would like, but it still beats AMD's 250W of power!

The FSB is not a bottleneck either, we'll borrow some text from hardwarezone.com:-

Throughout out benchmarks though, the age-old question of whether the current 1066MHz FSB is sufficient has been answered. Looking at all the applications that do take advantage of all four cores (eg. SPEC CPU2000, Lightwave 3D, Cinebench), performance increase is linear with the number of cores available and doesn't seem hampered at all by bandwidth bottlenecks, so it'll probably still be a while before Intel will bump desktop FSB speed to 1333MHz.


Who cares about 4 x 4? They need TWO processors to match ONE processor by Intel. That's pathetic.

You need an expensive 4 x 4 motherboard that has two Nvidia northbridges on it, makes for some nice high prices eh?

About the only interesting thing is that the board supports 4 GPUs. But that is still useless since SLI supports only TWO video cards. The exception of course is Quad SLI, but even with quad SLI you're still only running two video cards.

Quad core performance at half the power use of AMD's awful solution. I'd say AMD is fragged!

5:48 PM, November 04, 2006  
Anonymous Anonymous said...


All I have to say is that if were actually true that Kentsfield could run full speed with a 1066Mhz FSB speed then Conroe should similarly run full speed with the FSB set at half, 533Mhz. Let's see some tests to prove this. We'll assume that the cache coherency traffic on Kentsfield is negligible and not worry about that. It shouldn't be that difficult to do but I'll bet that no one will touch this one; it's much easier to just keep quoting pointing to poor testing.

I really don't understand this part...
How cache coherency traffic on Kentsfield is negligible?
1066MHz for memory access shared between four cores is enoght, but when coherency comes in play...
Every time a core request memory the chipset MUST check for it in every cache, it is a huge problem if all cores use the same channel for transmiting data.
1 cache - ok
2 caches - slower, but acceptable (kentsfield)
4 caches - f***ing slow (old Xeons MP)

6:02 PM, November 04, 2006  
Anonymous Anonymous said...

HWZ ia a known pumper. Feed them the right amt of moolah, and they'll sing whatever you want.

As for power consumption, Intel's TDPs are measured differently from AMD's. Intel uses "typical" while AMD uses a factory diagnostic that enables all the transistors in the processor (never gonna happen in real life but in engineering, it is known as worst case scenario that you have to plan for). Expect a 20-30% difference for the two.

An Intel rated 65W part ran at about 5W less than a 89W AMD part in some of my tests.

Bear in mind, also that the AMD cores have their own memory controllers (budgeted in the 89/65/35W) whereas Intel's CPUs still need to account for the extra 20+W from the Northbridge.

So how hot is the Kentsfield again?

http://www.theinquirer.net/default.aspx?article=35506

"Start first with the default settings of 2.66GHz/FSB 1066 operation, the system ran up to 72C CPU temperature within few seconds after entering BIOS setup, all that with extremely high fan noise, comparable to a belching public bus departing from a stop.

The system, nevertheless, went on and handled the WinXP installation all the way without a hitch. After installing all the usual drivers and utilities, I ran Sandra 2007SP1, and right after starting the CPU benchmark, a nasty Intel Monitor message propped up "CPU temperature 81C, exceeding the set limit" - uuuh, this was awful, I never had such message on default CPU speed, even with Pentium XE965."

Stock CPU speeds and voltages and Intel's own hardware monitor tells you it is overheating!!!

6:26 PM, November 04, 2006  
Anonymous Anonymous said...

Kentsfield already feeling the FSB limitation crunch.
Why would anyone shell out a good deal of money on a crappy processor
that can't do efficiently what it is supposed to(multitasking).
So eat the dust fanboys

http://www.xbitlabs.com/
articles/cpu/display/
core2extreme-qx6700_10.html

7:47 PM, November 04, 2006  
Anonymous Anonymous said...

I just can't wait to see the AMD 4x4 results.

Kentsfield is a joke. 2 Dual Cores glued together. PATHETIC!!!

If they are the leading technology in the microprocessors (So they say) then why can't they engineer a native guad core instead???

They glued 2 Conroes for a reason, just so they can attract people that have no clue how the micro architecture works. "UUUUU, AAAAHHHH, FOUR CORES!"

Intel is simply behind AMD with its FSB as Mr. Sharikou mentioned.


Wait. The ironing is delicious.

Intel's solution (4 cores on one physical chip) is a hack, while AMD's "solution" of using two whole damned processors and a specialized new motherboard and specialized new chips is _elegant_?

I've seen every brand of idiocy conceivable on this board now.

8:13 PM, November 04, 2006  
Anonymous Anonymous said...

Whats the power consumer Pretender..

What is AMD's power please pray tell us you PhD?

8:33 PM, November 04, 2006  
Anonymous Anonymous said...

Some vulture wrote:

"Throughout out benchmarks though, the age-old question of whether the current 1066MHz FSB is sufficient has been answered. Looking at all the applications that do take advantage of all four cores (eg. SPEC CPU2000, Lightwave 3D, Cinebench), performance increase is linear with the number of cores available and doesn't seem hampered at all by bandwidth bottlenecks, so it'll probably still be a while before Intel will bump desktop FSB speed to 1333MHz."


OH PLEEEEASE spare us the BS.
Toms is a GUTTER SITE WITH GUTTER REVIEWS.
Its paid intel buttsniffer.
Will go to any lows when the payday nears.

TOMS DO YOU HAVE ANY SHREDD OF HONOR LEFT?

9:05 PM, November 04, 2006  
Anonymous Anonymous said...

sharikou Said "4x4 will frag Intel, that's for sure. Con X6800 had about 10% performance advantage over FX62... "

Are you for real dude? You finally caved on CONROE fragging the FX62.. What kind of drugs are on these days..

You realize this statement invalidates every blog in which you claimed that Intel is engaged in a conspiracy, rigged benchmarks from Anand and company, and every argument you made about fragging Intel C2D..

Next thing you know, you come clean and invalidate other lies such as you PhD, that you actually went to school, you are employed, and do not forget the battery exploding lie!!!!!!!

11:13 PM, November 04, 2006  
Blogger Fujiyama said...

How you can predict the AMD 4x4 is fragging Kentsfield..please show your data...

If you look at Anand benchmark
here
you will see that FX-62 scores 3.31 in rendering. Another 200MHz should give about 0.25 point. (3.0 GHz FX fastest model)

Now:
(3.31+0.25)*2 = 7.12
Intel's quad-core scores 6.59

Why we can scale this way?
- Because each FX CPU has indepenend memory access
- Because every 200MHz in X2 line allowed to increase ~.25 in benchmark.

Of course it doesn't mean that 4x4 is going to win in every benchmark. But any useful software for quad-core these days is rendering and video encoding. DivX or any other encoding results should give even better results for 4x4 platform.
This is also proof that FSB-based architecture is not quite...uhm...well designed.

AMD will the winner! On Wednesday!

1:04 AM, November 05, 2006  
Anonymous Anonymous said...

AMD: "Kentsfield isn't really a native quadcore CPU! BTW, here's our 4X4 solution to compete with it. No, it's not quadcore too".

3:38 AM, November 05, 2006  
Anonymous Anonymous said...

Why Kentsfield really doesn't need superfast FSB is really simple: vast majority of all applications rarely need lots of bandwidth.

Let's simplify and say every thread needs maximum amount of memory bandwidth 5% of the time. With single core CPU this is simple as it can have full FSB at that 5% of the time any time it needs it.

With dualcore there are two threads that both need maximum bandwidth 5% of the time. As long as these times are not cuncurrent they both have almost zero slowdown caused by sharing the bus. There is roughly 1 in 20 chanche of these 5% overlapping.

With four cores it is simply 20% of the time when one of the cores need the full bandwidth. Sure it can happen that all the four cores want to have that bandwidth at the same time but that happens rather rarely. Anyone with basic math knowledge can tell you that.

In real world things are of cource more complicated. Almost no programs need as much bandwidth as any today's CPU can provide. Most of them need constant but little bandwidth, something in the order from few tens to couple hundred MB/s with rare spikes of few GB/s bandwidth needs.

Most programs benefit much more from the lower latency of ram. Intel can't provide that low latency for as long as it doesn't have IMC.


"AMD 4x4 ready to frag Kentsfield with scalable architecture"

Can you give any predictions on how much wil 4x4 be faster of Kentsfield? 3dmark 2006 CPU benches should be quite good and nicely measurable. Of cource you can always give some of your own suggested benches.

There is also Cinebench but that isn't really a good example of a fast ray-tracer. Seems as that one was programmed by people who can't do concurrent programming that well since dualcores get much less than 100% speedup. This is much better one and really there things scale almost linearly with CPU speed and core count. Though you could say it is "C2D optimized" since it uses SSE2 exclusevly but scaling is still measurable. I know that 3.1GHz e6300 was just a little bit more than twice as slow as 3.25GHz Kentsfield, thus nearly 100% efficiency in scaling. You can see some of the results down that thread.

"AMD's core-core, core-memory, core-i/o, proc-proc communications are all Direct"

And currently still AMD gets beaten by ancient Intel technology. What does that show?

10:51 AM, November 05, 2006  
Anonymous Anonymous said...

I find it odd that Intellers keep saying FSB doesn't matter.

It does.

Even on my old 955 Extreme Edition, getting the 1066MHz FSB made all the difference vs a similarly clocked Pentium 4 on 800MHz FSB in real world applications, as well as in memory access.

And this is JUST on a single core.

I am pretty sure there were tonnes of benchmarks that demonstrated this when the 1066MHz FSB CPUs were first launched.

And like Scientia suggested. Run a C2D 6800XE on 533MHz FSB and see if the system will run just as fast. If Intel's claims that the Kentsfield at 1066MHz FSB takes up soo little utilization, a simple test like this should immediately verify the results.

5:18 PM, November 05, 2006  
Anonymous Anonymous said...

"Even on my old 955 Extreme Edition, getting the 1066MHz FSB made all the difference vs a similarly clocked Pentium 4 on 800MHz FSB in real world applications, as well as in memory access.

And this is JUST on a single core"

"And like Scientia suggested. Run a C2D 6800XE on 533MHz FSB and see if the system will run just as fast. If Intel's claims that the Kentsfield at 1066MHz FSB takes up soo little utilization, a simple test like this should immediately verify the results."


Scientia has yet to back up his claim with any benchmarks. If the FSB is the limiting factor no application that the mainstream uses has yet to slow it down.

IMC has its advantages but you dont hear sharikou or scientia saying that if you dont get low latency ram you will take a sizeable performance hit whereas core 2 duo will pretty much perform the same even with generic ram.

And if the FSB was aging like sharikou and scientia will lead you to believe then intel wont have beat down on AMD's high end like they did using "ancient and less elegant" technology.

Even back in the x2 versus smithfield days amd's solution didnt give as much of a whooping as intel's core 2 gives to amd's fx.

10:07 PM, November 05, 2006  
Anonymous Anonymous said...

I've seen some 4x4 results and I was not that impressed, however I do not remember that there was encoding benches in those results. It certainly did not do as well with rendering compared to Kentsfield.

10:49 PM, November 05, 2006  
Anonymous Anonymous said...

"And like Scientia suggested. Run a C2D 6800XE on 533MHz FSB and see if the system will run just as fast."

Of cource it won't because you add quite a bit to the memory latency. The main benefit of faster FSB is smaller memory latency, not more bandwidth. Sure, more bandwidth helps too but only on rare cases. Faster latency on the other hand helps almost everywhere. Just remember how nicely K8 works.

11:42 PM, November 05, 2006  
Anonymous Anonymous said...

Well all this bs so amd 4x4 will it have any effect on playing an online game such as world of warcraft no it only sees one core.

Amd could have 50 cores it would have no impact on gaming performance with games such as these.

I have been building and testing machines for over 20 years NOTHING and i repeat NOTHING amd has is even close to core 2.

I have done real world benchmarks using games and 3d rendering programs and amd has nothing that can touch it atm.

4x4 who is going to dish out that much money for a system, not home users even gamers wont pay that much for a system where to have 2 gigs of ram you need to buy 4 gigs, have you seen the price of ram!.

Most of our customers would not even buy a quad core, most games and all online games are single core only.

11:15 AM, November 06, 2006  
Anonymous Anonymous said...

"most games and all online games are single core only"

For how long?

Also keep in mind that games like Unreal Tournament 2007 and Crysis are coming.

12:48 AM, November 07, 2006  
Anonymous Anonymous said...

I still do not get AMD's 4x4 platform when AMD has been heralding lower power consumption. 4x4 is definately going to use much more power than the Intel Core 2 Quad. The smartest thing AMD could have done to save face is what Intel did, a non-native quad core by dropping 2 dies on a package. I am just not impressed with 4x4 on paper, I think it will turn out to be a error in judgement on AMD's part.

12:36 PM, November 07, 2006  
Anonymous Anonymous said...

4x4 was an AM2 platform. Then AM3. Now Socket F only. Didn't require coherent HT busses or ECC RAM. Now it does. Didn't require "Special Processors" (Opteron) now it seems they are specifically creating new A64 chips (In this case the FX line) with the HT speech required to fuel a dual socket configuration.

So to sum it up. You will need an FX processor which are quite expensive. an E-ATX case + E-ATX Motherboard (Also more expensive than ATX cases and mobos) and you get lower-performing processors...

All I have to say now is who do they plan to market these to? People with no common sense?.

9:41 AM, November 09, 2006  
Anonymous Anonymous said...

http://xbitlabs.com/articles/cpu/display/amd-quad-fx.html

Not looking good for QuadFX and no fragging is being done except maybe Kentsfield on QuadFX apart when you get to running 5 programs at once ( and how often does that happen for the ethusiast ?

10:47 PM, November 29, 2006  

Post a Comment

<< Home