AMD redisigned core before years end.

[HDTVMan]

I read somewhere earlier this year that AMD stole a critical engineer from Intel that he was to help redesign the Athlon 64 core to help it increase its speed on current cores. Basically helping AMD achieve 3ghz on 130nm processes like Intel is capable of doing. Something to the effect of moving specific pipelines and shortening certain paths would allow them to achieve similar speed to Intel without having to resort to die shrinks to get the speed increase.

I also heard someone say this was not Intel but IBM then someone else saying it was a software package AMD bought which was going to allow them to gain up to 15% speed improvement on existing manufacturing processes.

This sounded legit but I am no engineer on why this may or may not occur.

Basically because Intel can produce a 3+ ghz chip easily on 130nm and 3.73 on 90nm and yet AMD still cannot reach 3ghz on 90nm this made sense that a core redesign to help increase the speed was in top priority investigation by AMD to make it happen.

Anyone else have more information on this?

Shouldnt IBM be able to solve the 3ghz problem of the Athlon 64 even on 130nm? How come AMD hasnt reached that yet even on 90nm. Starting to sound like a broken record.

They cant bank on die shrinks every time they have to be looking into why they are still 25% slower clock on the same manufacturing process right?

 

[GuitarDaddy]

The difference in architecture makes it much more difficult for AMD to hit 3ghz than Intel. AMD has a higher IPC(instructions per clock cycle) approx. 9 vs Intels 6. Thats why an AMD chip running at 2.0ghz is equal in performance to an Intel 3.2ghz. So for AMD to hit 3ghz would be like Intel hitting 4.5ghz Thats the dirty little secret that all us AMD fans have known forever

My new San Diego 3700+ overclocked to 2.8ghz performs equally with Intels top chip overclocked to 4.2ghz🙂

 

[PerfeK]

Intel is "forcing" the high clocks with brute force and long pipelines. It really doesn't matter how many clocks per second it can do. It's all about instructions per clock. The ghz race is over.

 

[sangyup81]

The Pentium 4 is not a superior processor. It can have a 1ghz clock-speed advantage over an Athlon64 yet still lose in benchmarks.

 

[stevty2889]

The P4 is like drag racer, while the AMD is like a ferarri. Sure, the P4 has a higher speed, but you have to keep feeding it fuel in order for it to keep going, while the AMD might not be moving as fast, but it's much more efficient and doesn't have to stall out waiting for more fuel. They are completely differant designs, the A64 doesn't need to run at as high of a clock speed, because it has short wide pipelines allowing it to do more work at a lower clock speed. Why do you think Intel is going to ditch the Netburst architecture with Merom/Conroe?

 

[dmens]

Originally posted by: GuitarDaddy
AMD has a higher IPC(instructions per clock cycle) approx. 9 vs Intels 6.

holy jesus 9? where'd you get those numbers? is it another dirty little AMD fan secret?

 

[ahock]

Guys you didn't answer his question as to why AMD cant reach 3ghz. Ok we know that clock race is over. IPC wise AMD leads which makes it much efficient to Intel's P4. But why they cant reach the said clock speed?

Is there something on theor silicon process that prevents this from happenning? Or was it that once AMD match Intel's current speed, they will also hit 100+W power?

 

[dmens]

K8 was designed with more logic between each pipestage and hence cannot clock as high as P4. I doubt there is too much difference between the FO4 delays of gates between Intel and AMD's 90nm process. Intel might be a bit faster since their 90nm is more mature.

If K8 hit P4 frequencies, their power numbers would also blow through the roof. Can't tell without measurements if it is worse than P4. I'd be interested in the power draw of a 3.8 ghz K8. Any serious oc'ers out there with a big multimeter?

 

[miketheidiot]

Originally posted by: dmens

Originally posted by: GuitarDaddy
AMD has a higher IPC(instructions per clock cycle) approx. 9 vs Intels 6.

holy jesus 9? where'd you get those numbers? is it another dirty little AMD fan secret?

actually its 10 if i recall correctly from an arstechnica article a few months back.

 

[Aenslead]

a.- Yes, AMD got an ex-Intel engineer.
b.- no, he did not help get AMD CPUs @ 3.0Ghz. AMD64 and NetBurst are COMPLETLY DIFERENT THINGS.
c.- yes, IBM's silicon on insulator helped quite a bit with thermal loss. A smaller process is NECESARY for increased clock on AMD processors.

 

[Zebo]

Sounds Like Intel is looking for AMD engineers😉

"AMD has a few good CPU architects that we would definitely go after. They have some good people, and we are always pursuing good talent. We have hired some in the past. One of them was Jay Pickett, who was one of the lead designers of the K8."

http://www.tomshardware.com/business/20050506/gelsinger-05.html

Also debunks your "AMD stole a critical engineer from Intel that he was to help redesign the Athlon 64 core "... Intel stole him for AMD😛

 

[CrispyFried]

Originally posted by: ahock
Guys you didn't answer his question as to why AMD cant reach 3ghz. Ok we know that clock race is over. IPC wise AMD leads which makes it much efficient to Intel's P4. But why they cant reach the said clock speed?

Is there something on theor silicon process that prevents this from happenning? Or was it that once AMD match Intel's current speed, they will also hit 100+W power?

I thought it has something to do with how may layers were in the CPU. Intel has more "insulating" layers so layers next to each other dont bleed signals. Kinda like high UDMA mode drive cables have 80 conducters, not 40, to prevent cross talk.

Cant remember where I read that though.

 

[GRIdpOOL]

The question was answered. It is the pipelines. Longer pipeline architecture of Intel allows higher clocks because less instructions are handled on each cycle. It is a trade-off. AMD chose the other route. They stuck to a better architecture that allowed more IPC rather than crank the wattage up and stretch pipelines. AMD's pipelines will grow and clock speeds will increase in the future, but that isn't the direction they took. They made the right decision, but in the clock speed race, it looked bad for PR. We all knew it was just a PR show by Intel.

 

[dmens]

Originally posted by: miketheidiot
actually its 10 if i recall correctly from an arstechnica article a few months back.

even under ideal (read: impossibly optimistic) conditions, no x86 uproc can sustain a thru-put of 10 macro-ops per cycle.

ars is just flat out wrong sometimes

 

[Aenslead]

Originally posted by: GRIdpOOL
WANG 386sx 16Mhz 4MB Windows 3.11

OMFG!!!!! that has got to be the BEST computer I've ever seen!!! Please, post some benchmarks!! do something!! How much did it cost!? How do you cool it?!

 

[Duvie]

It is 9...read the top of the article on the changes of K8 versus the K7 which had 9 IPC then..K8 has enhancements above that....http://www.cpuid.org/K8/index.php

 

[HDTVMan]

I can understand the efficiency or process part however this still doesnt make sense to me unless the cores truely do not run at clock speed.

If I understand correctly the fastest Intel gates can operate at 3.2ghz on 130nm and 3.73 on 90nm. Is this only the memory controller or the CPU core?

Why cant amd gates operate at this speed? As I think about it what part keeps the AMD switches/gates from running beyond 3ghz even on 90nm?

If intel can make gates that switch on/off at 3.73ghz AMD should eventually be able to have switches operate at the same speed. Right?

Now if its because its feeding the IPC which is running above clock speed then it makes sense that if Intel is doing 6 per clock then amd is doing 9 per clock then that means to me around 2.5ghz AMD its IPC is running near 3.7?

Am I right or is this something I should put in with Infinite smallness?

 

[Duvie]

"Now if its because its feeding the IPC which is running above clock speed then it makes sense that if Intel is doing 6 per clock then amd is doing 9 per clock then that means to me around 2.5ghz AMD its IPC is running near 3.7? "

Ding Ding Ding Ding!!!! AMD has one more of each integer unit and fpu unit.....That article pretty much breaks it down...

i read a good technical one about 1 week ago that borke it down even better for a very clear understanding...remember pipeline length will add penalties and so it isn't likeIntel can just say hey lets do 9 as well...

 

[HDTVMan]

Thanks Duvie I will dig a little deeper now that I have somewhat of an idea what is going on.

I am trying to find the article on the proposed speed increase by redesign but it looks like I understand why this might not help. But then it may by decreasing other possible latencies around the IPC and possible current leakage or something else I heard inductance bit jumping something or other.

 

[Intelia]

Now wait a min . you guys !
1) First of all this pipe line thing. You talk like its new but its old Both Intel and Amd engineers understood along time ago how there pipes work and the +- of both. Intel choose to go to long pipes (netburst ) while AMD choose the Throughput and simpler method. all here have heard the old saying (Keep IT Simple STUPID) at present time this seems to to have woked.
2) Intel seems to have solved some problems with the 6XXXX series and If they get 65 nano. process better and Intel can infact scale better with the Pressler sigle core brother Smithfield (thats a rifle that help tame the wild west)( Prescott). Than it is possible that netburst may be here longer than what people are saying. I don"t know. I do know smithfield will have shorter pipes than current Prescotts but I haven't found what yet. It well also use some PM Arch. I know some of what is being changed but not all. So right now I going to stay neutral and go on I wait and see mold.
3) I don't care how good of an engineer you are you can't change the laws of physics (Right now Amd doesn't have that kind of tech)
Whether or not Intel can scale backwards we well see but I think that is a good possiability. Netburst does seem to have hit a wall. The arch. and engineering that went into netburst no matter what happens is going to serve Intel well for a lot of years.
4) Pci and PCI E good examlpe. The tech. been around since the 70's but there was no need for it. Now that bandthwidth has created a bottle neck enter PCI-E an old but unnessary Arch. (at the time ) to the rescue .
5) 2 months ago I would have said netburst and presscott Arch . has died ! But right now I am thinking I have buried it to soon.

 

[coldpower27]

Well Intel is going to keep around the NetBurst architecture around for awhile at least on the 65nm process, with the Presler and Cedar Mill cores, though they are also working on thier new High IPC architecture with Merom and Conroe. They will exists side by side and Intel will decide if they think NetBurst later on is worth keeping around along side their new architecture. For the moment, it looks like the high IPC way is the way to go, but that could change depending on how well NetBurst scales versus the Conroe and Merom architecture.

The NetBurst architecture has served Intel well over the times it was alive making Intel lots of money, Prescott has done this really well being built on 90nm with 300mm wafer processing.

Oh Intelia, Presler's single core brother is called Cedar Mill not Smithfield, the Smithfield we already currently have with the P4 EE 840. and we will soon also have it's mainstream derivatives P4 820, 830, and 840. Though I apologize if your refering to something else.

 

[Duvie]

Originally posted by: Intelia
Now wait a min . you guys !
1) First of all this pipe line thing. You talk like its new but its old Both Intel and Amd engineers understood along time ago how there pipes work and the +- of both. Intel choose to go to long pipes (netburst ) while AMD choose the Throughput and simpler method. all here have heard the old saying (Keep IT Simple STUPID) at present time this seems to to have woked.
2) Intel seems to have solved some problems with the 6XXXX series and If they get 65 nano. process better and Intel can infact scale better with the Pressler sigle core brother Smithfield (thats a rifle that help tame the wild west)( Prescott). Than it is possible that netburst may be here longer than what people are saying. I don"t know. I do know smithfield will have shorter pipes than current Prescotts but I haven't found what yet. It well also use some PM Arch. I know some of what is being changed but not all. So right now I going to stay neutral and go on I wait and see mold.
3) I don't care how good of an engineer you are you can't change the laws of physics (Right now Amd doesn't have that kind of tech)
Whether or not Intel can scale backwards we well see but I think that is a good possiability. Netburst does seem to have hit a wall. The arch. and engineering that went into netburst no matter what happens is going to serve Intel well for a lot of years.
4) Pci and PCI E good examlpe. The tech. been around since the 70's but there was no need for it. Now that bandthwidth has created a bottle neck enter PCI-E an old but unnessary Arch. (at the time ) to the rescue .
5) 2 months ago I would have said netburst and presscott Arch . has died ! But right now I am thinking I have buried it to soon.

1) Tell us something we didn't already know...

2) It doens't appear to have been that great if they still cant get much better then dual 3.2ghz core to the market..Whether it is 5xx series or 6xx series they still comsume way too much power and are hotter on average then AMD cpus....You always talk about the 65nm as the saving grace of INtel...Wha happened to the 90nm process?? Short attention span do you have (yoda voice)??? After that monumental dud I think we shall all wait and see...Also dont forget AMD will have a 65nm and if it goes as smoothly as their 90nm did they really have little to worry about...

3) AMD doesn't need to...They are far from that wall at this point ad it is clear dual cores will be the new standard in the future anyways...

4) the bottleneck is called software genius!!! This will be the bottleneck that will erar its ugly head single-handedly the most in the nest few years...

5) I would unbury it too so it could decompose faster.....Again genius all the brightest things in Intels future have nothing to do with netburst or anything alng the lines of prescott. You must clearly have the head planted up your arse to think anyting along those lines...

 

[stevty2889]

Originally posted by: Intelia
2) Intel seems to have solved some problems with the 6XXXX series and If they get 65 nano. process better and Intel can infact scale better with the Pressler sigle core brother Smithfield (thats a rifle that help tame the wild west)( Prescott). Than it is possible that netburst may be here longer than what people are saying. I don"t know. I do know smithfield will have shorter pipes than current Prescotts but I haven't found what yet. It well also use some PM Arch. I know some of what is being changed but not all. So right now I going to stay neutral and go on I wait and see mold.

5) 2 months ago I would have said netburst and presscott Arch . has died ! But right now I am thinking I have buried it to soon.

2. Smithfield is the dual core prescott, cedarmill is the single core version of Presler. It won't have shorter pipelines, it's not a redesign, it's prescott shrunk to 65nm, with improved strained silicon. It will run cooler, and have less leakage current, but it won't have a shorter pipeline. They aren't repeating their previous mistake, they made too many changes to prescott rather than just shrinking northwood to 90nm that got them unexpected results.

5. Netburst will most likely die with Presler and Cedarmill, the Merom based cpu's will be the future.

 

[sangyup81]

Originally posted by: Intelia
5) 2 months ago I would have said netburst and presscott Arch . has died ! But right now I am thinking I have buried it to soon.

Intel plans to start using the architecture of the Pentium M in the future. Netburst, for all intents and purposes, is dead.

 

[IntelUser2000]

"Your Athlon = 9 IPC versus P4 = 6 IPC is not quite correct.

The Intel and AMD chips both break the x86 instruction set into smaller chunks for internal operatioin, and it is these so called "mirco-ops" that are quoted as the IPC versus full x86 instructions.

The size of the micro-ops (ie how many micro-oops you have to execute to equal a full x86 instruction) is not the same, so the 9 vs 6 is really NOT an apples to apples comparision.

So the other thing that greatgoogilymoogily forgot is that to get to the 9 IPC, it has to embed more override codes to get the x86-64 instructions in. That means you have to bloat the code to use 64-bits. That means you have to do more ops to get the same effect. If the overhead is only 33% then IPC 6 = IPC 9. What you really need to optimize is IPC * Total # of instructions."

I got this from a quote somewhere. In reality, x86 CPUs like Pentium 4 and Athlon have IPC of 3-4, the CPUs with highest IPC being Itanium, which is fully taken advantage of, it has IPC of 6-8, plus some other architectural advantages.