Why dual cores and not one single faster one?

[BDSM]

Both AMD and Intel are seemingly working hard on their dual core alternatives..

I'm wondering.. why don't they just make a bigger and better single processor instead? The dual cores will be big (about twice a single one).. Why don't they just slap on more execution units on as ingle chip instead? Wouldn't that be better knowing that dual setups only yield -5 to +70% extra processing power compared to a single processor system?

Wouldn't it be nice with an even more powerful fpu in the a64?.. mmmh. yummy

 

[DragonFire]

There going to dual cores because they cant seem to make cpus go any faster. They have both hit a speed wall. There hoping that with dual cores running at say 2.5Ghz they can give you the performance of a single cpu running at 5ghz. Intel is have major problems making 3.6/3.8ghz and will be lucky to get a 4.0ghz out. AMD on the other hand will have even more problems since they use shorter pipeline then Intel.

 

[Navid]

Pipelining!

There are times that the CPU has to wait if it was only working on one task, no matter how fast it was.
Like having multiple cash registers at a store. Multiple customers can go through the cash register line in parallel. If there was only one register, things would be much slower.

I could be wrong. This is just a guess!

 

[BDSM]

Hey there dragon. I know about the "speed wall" they've hit, but that's sheer frequency.


navid.. yes.. I think the same way you do, but today a cpu has many execution units inside.. so it already has more cash registers. I'm wondering why can't they just increase that number even further?


I don't really know anything about this stuff.. So that's why I ask.

btw.. are the forings acting up for you ppl as well or is it just me?

 

[DragonFire]

I could all be wrong about what Im going to say but here goes.....

Its not a matter of how many "cash registers" there are but how fast they are. If your comparing Intel to AMD, the AMD line might only take an avg of 3 mins per preson while the Intel line would take an avg of 6 mins per person. Because of how long the pipeline is. This is why Intel has 3.6Ghz and AMD doesnt. The shorter the pipeline the faster the work gets done. Adding more wont help at all.

If I understand your first post, your wondering why they dont just increase the size of the core of a single cpu since dual cores will in fact be bigger. The core itself in a dualcore cpu might look bigger but under it each core will in fact me the size on todays core if not smaller. Taking a AMD barton core and making it twice as big will only increase the length of the pipeline.

If I'm idiot and dont know what Im talking about, please someone correct me.

 

[3chordcharlie]

Increasing parallelization has worked really well for graphics cards, but most cpu tasks are harder to run in parallel which is why they have focused on increasing clock speed. Relaistically, dual core will be pretty similar to SMP, and for the mainstream consumer is probably not all that valuable unless they are die-hard multi-taskers, or want to become such. I would see more benefits to professional customers who will gain the ability to fit more processors on their video-editing (and similar) machines (effectively you could now fit 4 CPUs on a dual-socket board).

 

[lobadobadingdong]

The dual core chips won't take up twice the space, as they will be coming on a die shrink, in point of fact, the die space will be very close to the current size per cpu. The performance advantages will be realized best with parallelism (which is becoming more and more popular). From a price perspective the cpu's shouldn't cost much more (if any) than the current single core processors. The AMD dual cores will get a performance hand with this when running dual channel memory as their core won't be as memory bandwidth limited as the Shared memory buss used for the Intel dual core solutions. (esp. considering P4's only run well with a full data transfer of the fsb)

 

[BDSM]

I realize the dual core cores won't be much bigger than todays single core cores because the dual cores will be on a thinner process. Anyone knows that.

Someone said having more executionunits in a cpu won't help cuz they are often stalled waiting for other stuffs to complete.. Well. if ya have a single thread running it will be just as bad with a dual core system as on "my" super cpu.. no?

I bet a single core with as many transistors could perform on average much better than a dual core.

Come to think of it. They are probably gonna do the multi core thingie because it's much easier and faster (read cheaper) than designing a new super cpu.

 

[DragonFire]

Originally posted by: BDSM

Someone said having more executionunits in a cpu won't help cuz they are often stalled waiting for other stuffs to complete.. Well. if ya have a single thread running it will be just as bad with a dual core system as on "my" super cpu.. no?

This why everyone doesnt already have a dual setup.

I bet a single core with as many transistors could perform on average much better than a dual core.

Come to think of it. They are probably gonna do the multi core thingie because it's much easier and faster (read cheaper) than designing a new super cpu.

Execpt they cant make a cpu with that many transistors using current tech....

No, there doing it because they are slowly realizing that they cant just keep increasing the speed. Even Intel is learning that speed isnt everything. So now it looks like there going to work on making cpus that can do more work in less time and dualcores are "supposed" to help do that.

 

[Geomagick]

I suspect the reason for going to dual cores is cost related. To design a new chip from the ground up to run in a highly parallel nature is going to cost a lot more in R&D than adapting a current core to run side by side with another. In the case of AMD it should be even better to do this because of the Hypertransport system.

It will be interesting to see just how intel manage to combine their P4 chip onto a single package given the penalty you get when you have dual Xeon's.

 

[Sahrin]

Dual Cores are a solution to every problem that befalls CPU manufacturers today. Process technology, thermal dissipation, gate leakage...it all. At least, that's what the CPU manufacturer's would have you believe. The truth is it's a band-aid, just like Hyperthreading is. It's getting harder and more expensive to make a faster processor. Putting more throughput in a CPU doesn't necessarily change things. If there are 2 math operations to do and you've got the choice of having 2 Arithmetic Logic Units or 5 ALU's, you're obviously going to choose 2. 5 is a waste. The idea is to make the Executions Units you have go faster. A processor with 1 ALU, 1 FPU, 1 VU all running at 50Ghz is faster than any processor tiwh 100 of each running at 100Mhz. Most tasks aren't huge throughput ones, even now we are seeing that generally, encoding stresses CPU's and pretty much it's the only thing. What helps encoding the most? Multi-threading. Gaming won't work faster or better, that's GPU dependent. Content Creation won't. It's a band-aid, that's all. A sloppy one.

 

[DonPMitchell]

Most of the tricks you see in modern CPUs go way back to Seymour Cray - pipelining and multiple functional units. The CDC machines did this stuff in the 1960s.

The analogy to waiting in line for a cash register is not accurate. You are actually doing a little part of the cash transaction every time you take a step forward. So maybe it takes you 30 steps to get through the line, but some person finishes every step. There are advantages to breaking the task into more steps, but tinier faster steps.

Of course its not that simple, long pipelines are more vulnerable to interruption, bad branch decisions, etc. The engineers at Intel and AMD are experimenting and exploring the design space, the way competing companies should. You know how tricky it is to benchmark systems, so imagine you are trying to judge a design before it is even built! Someone had to build the P4 and see if it was a win. Even if it isn't, Intel tried something new.

 

[quickford]

The reason dual processors haven't become mainstream is because most programs don't take advantage of it. If the CPU companies make dual-core processors which allow anyone to have multiple cores regardless of special motherboards, then I would expect to see software companies begin to optimize their programs for the dual-core technology. A program designed for that *would* run on a dual 2.5Ghz system as though it were a single 5Ghz system, other than being limited by the video card.

At least, that's how I've always understood it.

 

[DAPUNISHER]

Good thread :thumbsup: Here's hoping pm will find the time to duck in and explain the reasoning.

 

[Falloutboy]

my guess whats going to happen is as time goes on with these multicore setups with each redesign they'll start to share parts till its more like 1 CPU with multiple pipelines. So right now amd cpu's are going to share the memory controller. maybe in the next revision they'll find out a way fo them to share the same cache without hurting performance. this all will save transistors and make it possible to put 4 cores on a die and so on

 

[Bateluer]

And once you see Dual core chips on the market, you'll see software written to take advantage of it.

 

[clarkey01]

" AMD on the other hand will have even more problems since they use shorter pipeline then Intel"


Hitting 3Ghz for AMD is a lot more achievable then Intel Hitting 4.4Ghz any time in the next 12 months.

The shorter pipeline helps with the high IPC, so a lower number of Mhz are needed to match and excel performance of an Intel product.

 

[clarkey01]

But you dont know, prescott might start to shine come 4.4Ghz.

 

[VIAN]

I alwasy saw dual cores as being compared to VPU pipelines.

You will double your power.

 

[DAPUNISHER]

Originally posted by: clarkey01
But you dont know, prescott might start to shine come 4.4Ghz.

More like glow*from the heat* than shine

 

[clarkey01]

[Q} "More like glow*from the heat* than shine "[/quote]


lol true DAPUNISHER, but being on the side that I am, being pro AMD that is, I still try not to count out Intel.

I have my old slot 1 PII 350 Mhz with an Intel FAB man doll on my desk, soon we all need suits like the fab folk to work the Intel chips ( they'll be that hot).

 

[Wahsapa]

Read this this and this

infact, read all of these

its really hard and takes a long time for ANYBODY to put out a 'super cpu' its just not something easy to do. amd's a64, ibm's ppc970 and the pentium m are good steps in the right direction(baby steps at best) but its gonna take a lot of developing, and even then. if they start on it right NOW you wont see it for a long time(read years)

 

[BDSM]

Thanx for all the replies guys!

I have two other lil questions 4 ya:

1. Why can gpu's only achieve maybe 500 mhz or so at a 0.13 micron process while the p4 does six times that with ease?

2. Why can't they put both the cpu and gpu on the same chip?.. preferably with a gig of super fast gddr 3 @ 1 ghz .. Just imagine the sandra scores with a 256 bit bus.. woohoo

Anyway.. such chip would prolly have 1500 pins or so.. but it's doable I'm sure. except maybe the gpu couldn't handle that speed (maybe?).

Anyway.. What kind of performance do you think such a system could yield?

Cheers!

 

[DAPUNISHER]

Originally posted by: BDSM
Thanx for all the replies guys!

I have two other lil questions 4 ya:

1. Why can gpu's only achieve maybe 500 mhz or so at a 0.13 micron process while the p4 does six times that with ease?

2. Why can't they put both the cpu and gpu on the same chip?.. preferably with a gig of super fast gddr 3 @ 1 ghz .. Just imagine the sandra scores with a 256 bit bus.. woohoo

Anyway.. such chip would prolly have 1500 pins or so.. but it's doable I'm sure. except maybe the gpu couldn't handle that speed (maybe?).

Anyway.. What kind of performance do you think such a system could yield?

Cheers!

I'm not qualifuied to address the technical aspects, I can give an example and practical reasons for not to combine CPU&GPU though. First, heat. Add a nv40 core to a prescott and you will need water cooling minimum. Then you have the quandry of who's GPU to put on your CPU, after all there are several companies in the graphics game. Imagine trying to design every new CPU and be able to integrate either a basic 2D through 3D workstation capable GPU on to it depending on the client's needs I'm certain there are many more reasons this is impractical from biz. and tech. perspectives as well.

Why CPUs ramp better is probably a sheer architectural difference between them, but others have to address that.

 

[BDSM]

Dap.. yes if we took a nv40 and slapped it together with a prescott @ 3 ghz+ it could probably melt glass.

BUT.. Maybe we wouldn't need 16 pipelines if we were at 3.4 ghz.. if we had only 4 that would (according to my very scienific calculations about equal 16 pipes at 850 mhz (twice what we have today) and we would only have used a fraction of the die space too.

Then lets integrate a decent soundchip, good nic and a hdd controller as well

Just imagine a chip that included all that.. No slow buses. everything ON chip!. Plus everything would run inside the gfx memory (super fast gddr3 ofcourse)(which would also serve as main memory).

I remember using a very simple 1 chip computer back in school. It was from the early 80's. It didn't have a nic and no sound.. but it worked. And it was tiny. Surely they could make 1 chip puters today that rocked if they wanted to!

I know.. these are all wet fantasies for a nerd like me.. But nerds have the right to have those just like everyone else