8 and 16 core cpus

C

citizen

Weaksauce
Joined
Jul 4, 2004
Messages
66
So i work on a charter boat that does private parties. Last week we did an event with a bunch of techies on it. I had overheard some of them talking about upgrading their companies cpus to dual core. So i came back around later and started to talk to the guys with some idle chit chat. Turns out one of them was a aprticle physicist at Intel. Seeing as how i was entertaining the idea of switching to a conroe i talked to him for a little bit and this is what i gathered. Firstly, Intel has working 8 and 16 core cpus ready to get shipped out for testing, but that they aren't going to release them for a while because they don't need to. It doesn't seem like they are gonna ramp core speeds like they used to, when its easier to just add more cores. Conroe production is going for well and he said they would probably have a surplus at launch. Lastly, that if I really wanted to o/c a conroe you need sub-zero cooling. Anyways that's it for now~
 
I kind of see this as the way things are going. Eventually it's going to be about how many cores you have and not so much about the clock speed.

Lets go programming!! What a pain in the ass it will be to optimize your program for 16 cores after you just finished optimizing for 8. :(
 
I bet he thought that some guy working on a boat wouldn't be able to spread this info around.
Good for you man. Right before the conversation ended you should have told him that you
were an AMD spy. :eek:
 
ThirtySixBelow said:
I kind of see this as the way things are going. Eventually it's going to be about how many cores you have and not so much about the clock speed.

Lets go programming!! What a pain in the ass it will be to optimize your program for 16 cores after you just finished optimizing for 8. :(
Click to expand...

I don't thibk it is that simple : you can't just trade single thread performance for throughoutput.If that was the case we would flock to buy SUNs Niagara 8 core beast.Some tasks are inherently serial and CMP+SMT won't help.

You need very good single thread perf and a lot of cores.
 
Did these guys offer sources for any of this, or just a bunch of pimply-faced youths doing the "I heard" stuff?
 
Well I did say that i was running an amd cpu. He said that they liked the competition from amd and that they had some great products. If my venice wasn't such a dud (2.3 on water) i wouldn't even be considering upgrading.
 
MrGuvernment said:
So tell that to the people getting 3-4GH on air ?


BS.
Click to expand...


yeah, i mean, that kind of undermines anything any of them told you.
 
MrGuvernment said:
So tell that to the people getting 3-4GH on air ?


BS.
Click to expand...

What a partical physicist may connotate as a good overclock and what you may connotate as a good overclock are probably not the same. I think he was refering more to the 5ghz range.
 
I bet Intel does have working 16-core parts, but they probably only run at 1Ghz or so per core. They won't release them until they can get that back up to a reasonable level.

As someone else said, you can't trade single-core performance for multiple cores.

Until, that is, we get reverse hyperthreading. Then things will be much different.
 
HOCP4ME said:
As someone else said, you can't trade single-core performance for multiple cores.
Click to expand...


Yes you can, many applications have made use of multiple cores since there creation, simply because you may not use more then 2 cores, doesnt mean the other %99.99999 percent of the world wont.

i would rather have 16x1ghz cores then 2x 3ghz cores.

One example for me is an Exchange server we run, i tried running it on a single 2ghz Northwood system - it wasnt responsive on doing anything, had a 74g ultra160HD and 1g of ram.

I then installed it on a dual PIII 733 system, 1g of ram again and the same 74G ultra160 drive and it is fast, responsive and never bottlenecks.

The issue comes with software and the operating system being properly coded.

Yes, reverse hyper-threading will be a fantastic feat if done, as then you dont need to worry about what software supports what :) and i cant wait to see it, if it comes out.
 
MrGuvernment said:
Yes you can, many applications have made use of multiple cores since there creation, simply because you may not use more then 2 cores, doesnt mean the other %99.99999 percent of the world wont.

i would rather have 16x1ghz cores then 2x 3ghz cores.

One example for me is an Exchange server we run, i tried running it on a single 2ghz Northwood system - it wasnt responsive on doing anything, had a 74g ultra160HD and 1g of ram.

I then installed it on a dual PIII 733 system, 1g of ram again and the same 74G ultra160 drive and it is fast, responsive and never bottlenecks.

The issue comes with software and the operating system being properly coded.

Yes, reverse hyper-threading will be a fantastic feat if done, as then you dont need to worry about what software supports what :) and i cant wait to see it, if it comes out.
Click to expand...

Yeah but how many ways can you split a register move, load, combine register and store the results. R/HT isn't going to be as great as everyone thinks because a lot of single threaded apps are still going to have to be optimized to expect being split up across mulitple cores. Devs might as well code for multiple core to start with.
 
tdg said:
Did these guys offer sources for any of this, or just a bunch of pimply-faced youths doing the "I heard" stuff?
Click to expand...

Did you miss the part about the guy being a particle physicist?

What the hell is particle physicist anyway haha. Don't tell me to look it up I probably wouldn't understand anyway. :p
 
deeznuts said:
Did you miss the part about the guy being a particle physicist?

What the hell is particle physicist anyway haha. Don't tell me to look it up I probably wouldn't understand anyway. :p
Click to expand...

Someone who knows a hell of a lot about the physics of subatomic particles. Like electron flow in the core of a CPU.
 
Slightly off topic, but I just wanted to respond to something that was said earlier....


NEC has had a reverse compiler for YEARS.. RHT is nothing new. It does work, and it works well.

I personally dont think that scaling the number of cores is very efficiant... First of all how much redundant hardware will there be? Do you really need to have seperate front ends on each core? Or why not just use a specialized core that handles all of the decoding, then use general purpose cores that execute?

I REALLY dont think symmetric milti-cores are going to be the answer. This is just another mistake Intel is making, that will set the industry back ANOTHER five years.

Just like Clock speed wasnt the answer... This isnt either.... It is always one extreme or the next, but never a solid middle ground. That is Intel's problem.
 
duby229 said:
Slightly off topic, but I just wanted to respond to something that was said earlier....


NEC has had a reverse compiler for YEARS.. RHT is nothing new. It does work, and it works well.

I personally dont think that scaling the number of cores is very efficiant... First of all how much redundant hardware will there be? Do you really need to have seperate front ends on each core? Or why not just use a specialized core that handles all of the decoding, then use general purpose cores that execute?

I REALLY dont think symmetric milti-cores are going to be the answer. This is just another mistake Intel is making, that will set the industry back ANOTHER five years.

Just like Clock speed wasnt the answer... This isnt either.... It is always one extreme or the next, but never a solid middle ground. That is Intel's problem.
Click to expand...

Well if Intel's making a mistake, so is AMD, IBM, and others.
 
MrGuvernment said:
So tell that to the people getting 3-4GH on air ?


BS.
Click to expand...

I can tell you from my own experience that in order to get a Conroe fully stable @ higher FSB you do need sub zero cooling.

I got up to 340FSB on my X6800 with a stock multi, but I needed 1.55V and it wasnt fully stable. The temps were getting high even tho Im on water so I decided to stop.

On the other hand I got to 314FSB on stock voltage, but after that its really not worth it as high voltage adjustments do not gain that much as far as FSB goes.

I kept mine @ 325FSB with 1.46V going through the chip since I like the temp that its at.
 
ThirtySixBelow said:
... What a pain in the ass it will be to optimize your program for 16 cores after you just finished optimizing for 8. :(
Click to expand...
It depends on the code. Some stuff is amenable to be split up into as many subtasks as you need, and so it is effectively optimized for as many cores you can throw at it. For instance, the image processing code in Paint.NET is written to use 1 thread for every core/processor in the system. I've seen it benchmarked (with PdnBench) on a quad dual core Opteron (eight cores) and we had a performance gain of 7.1x if I remember right.

Other stuff, like maybe VirtualDub, has a thread for each stage of its pipelined task and will thus hit a performance wall when you reach the number of processors matching the number of pipeline stages. (maybe stage 1 is "fetch data", stage 2 is "decompress", stage 3 is "resize to lower resolution", stage 4 is "compress to some other format", stage 5 is "write data to disk").
 
You must log in or register to reply here.
Back
Top