Sign in to follow this  
Followers 0
dantm

AMD vs. Intel -- frequency/efficiency summary

20 posts in this topic

Hi,

I remember in the past I had found a very interesting article describing and comparing AMD and Intel in terms of their CPU frequencies, and efficiencies, but now I can't seem to find it.

Can anybody recommend an online source for a simple, easy to read summary on these topics? Ie. to answer why is it that AMD can be better for math operations at lower frequencies than a comparable Intel processor at higher frequencies, and why Intel can be better in various graphics programs/multimedia/etc.

Thank you very much!!

Share this post


Link to post
Share on other sites

Some of the most in-depth analyses are at:

http://arstechnica.com/articles/paedia/cpu.ars

and

http://www.lostcircuits.com/cpu/

That said, with dual-core much of it goes out the window. Everyone is still trying to sort out how to benchmark desktop multitasking in a meaningful way, and AMD also gains much more with dual core than Intel because they never had hyperthreading.

Share this post


Link to post
Share on other sites
Some of the most in-depth analyses are at:

http://arstechnica.com/articles/paedia/cpu.ars

and

http://www.lostcircuits.com/cpu/

That said, with dual-core much of it goes out the window.  Everyone is still trying to sort out how to benchmark desktop multitasking in a meaningful way, and AMD also gains much more with dual core than Intel because they never had hyperthreading.

203209[/snapback]

Actually, I'd tend to think that AMD gains much more from dual core than Intel because AMD has a crossbar switch on the chip that runs at core frequency, while Intel has to use their slow and overburdened FSB to communicate from 1 core to the other.

Edited by Trinary

Share this post


Link to post
Share on other sites
I dunno how well the P4's long pipes will lend to dual core. Only time will tell I guess.

203397[/snapback]

In much the same way they work in Dual Xeon setups. Same strengths and weaknesses as a single CPU pretty much. The 1CPU - 2CPU scaling factor seems to be a good bit higher for Opterons too.

Dave

Share this post


Link to post
Share on other sites

This is pretty academic since the products will not compete with each other for at least the next year. Intel can't make high speed dual-core chips until 0.65nm because of power and heat issues, and AMD won't make low-end dual-core chips because of a serious lack of manufacturing capacity.

So Intel's chips are all cheaper than AMD's chips, and actually priced about the same as high-speed single-core chips. AMD's chips are priced the way the fastest Opterons have always been. So AMD is much faster, just as you'd expect at a much higher price.

And as Xeon systems have shown, a single FSB is often not a big deal if you are only looking at two cores, especially as the AMD dual-core chips only have a shared memory controller anyway. After all, the Tiger K8W and Master2-FAR dual-Opteron boards don't lose much real-world performance by sharing only one memory controller.

The AMD will of course scale much better at 4 and 8 cores than any Xeon, though eg putting a pair of dual-cores on a single memory controller in the boards above is bound to hurt performance also. Think of it this way: if Intel uses a Hub, and AMD uses a switch, does it really matter if there are only two users on the network? It might start to when there are four.

Edited by bfg9000

Share this post


Link to post
Share on other sites
Think of it this way: if Intel uses a Hub, and AMD uses a switch, does it really matter if there are only two users on the network? 

I suggest that that such an analogy is incomplete; I suspect one would notice the difference between a 10Mb hub and a 100Mb switch, even with only two users.

Share this post


Link to post
Share on other sites

Would you notice a difference if they were both on 7.5Mb broadband? Maybe you could measure it...sometimes.

Both Intel's 800MHz FSB and AMD's memory controller can move some 10,000 Mbits/sec. Intel's architecture "needs" bandwidth, and AMD's provides lower ping...

Share this post


Link to post
Share on other sites

Upon further reflection, I think the hub/switch/network analogy is just a poor analogy for this case, as it does not account for other architectual differences (e.g. interprocessor communication on AMD does not consume FSB bandwidth). Perhaps a more accurate analogy would involve integrated video UMA (but that analogy is flawed as well, as video will constantly consume bandwidth, unlike interprocessor data exchange)...

Here's a comparison chart (biased, as it was prepared by AMD): Chart link

Share this post


Link to post
Share on other sites

It is a bit late, memory bandwidth for both is ~80,000Mbits/s no matter how fast the connection to that memory is. Kind of like how U320 is not faster than U160 on PCI 32/33 bus, because that is the bottleneck.

My point was that the single socket dual-core AMD has only one memory controller so it should perform like current dual-socket Opteron setups with one memory controller. And the Intel should perform similarly to how a Dual Prescott Xeon 1MB setup with 800MHz FSB performs already today.

The AMD does have dedicated communication channels, and the number of memory controllers ramps with the number of sockets so it scales better. But even today, a 4P Opteron beats a 4P Xeon by a greater margin than seen in 2P systems. Intel keeps raising the FSB enough to well enough minimize this bottleneck for 2P systems, at least for most apps.

Share this post


Link to post
Share on other sites
It is a bit late, memory bandwidth for both is ~80,000Mbits/s no matter how fast the connection to that memory is.  Kind of like how U320 is not faster than U160 on PCI 32/33 bus, because that is the bottleneck.

Where does the 80 gbit/s come from?

Share this post


Link to post
Share on other sites
Think of it this way: if Intel uses a Hub, and AMD uses a switch, does it really matter if there are only two users on the network?  It might start to when there are four.

203581[/snapback]

It can matter with only two users, yes.

My point was that the single socket dual-core AMD has only one memory controller so it should perform like current dual-socket Opteron setups with one memory controller.

The difference is that the latency to the second memory controller is (probably) larger in case of different CPUs than in case of different cores on the same CPU.

Edited by Olaf van der Spek

Share this post


Link to post
Share on other sites

IIRC, the Techreport showed One socket, dual core outperforming two socket, single core by a small margin. The advantage of faster inter-core communication on the single socket system outweighed the two memory controllers on the two socket system.

Anyway, AMD architecture is less bandwidth dependant, so it can cope with halving the memory bandwidth per core much better than Intel can. If you consider a two socket, four core system, there's still twice the bandwidth per CPU as an 800 MHz FSB quad core Intel system, where all cores share the same FSB, whether it's two twin core sockets or four single core sockets.

Share this post


Link to post
Share on other sites
IIRC, the Techreport showed One socket, dual core outperforming two socket, single core by a small margin. The advantage of faster inter-core communication on the single socket system outweighed the two memory controllers on the two socket system.

203611[/snapback]

Just goes to show how little AMD needs that bandwidth, and how much it thrives on low latency. That's why I pointed out the small difference between Tiger and Thunder for most uses, despite Thunder having double the bandwidth. Olaf, I was just roughly translating 6.4GB/s to bits for the network analogy.

Share this post


Link to post
Share on other sites
IIRC, the Techreport showed One socket, dual core outperforming two socket, single core by a small margin. The advantage of faster inter-core communication on the single socket system outweighed the two memory controllers on the two socket system.

AMD implemented a System Request Interface (SRI) between cores on the same cpu-package. This basically allows much tighter inter-core communication (as Spod mentioned above), but a large amount of the benefit comes from the fact that the SRI significantly increases the speed of the transfers required for cache-coherency. This is why a single dual-core chip is able to beat two single-core's, even though the dual-core has half memory bandwidth per core.

Regardless, Intel's chips must go over the slow FSB to deal with cache-coherency, and that's why they're much slower and show per-cpu returns that diminish much more quickly than those of AMD's chips.

ELiTe

Edited by ELiTe

Share this post


Link to post
Share on other sites
That'd be 51,2 gbit/s then.

203689[/snapback]

Make up your mind, Olaf! Is it 51.0 gbit/s or is it 2.0 gbit/s? ;)

Share this post


Link to post
Share on other sites
That'd be 51,2 gbit/s then.

203689[/snapback]

Make up your mind, Olaf! Is it 51.0 gbit/s or is it 2.0 gbit/s? ;)

203728[/snapback]

Ah, sorry. 51.2 gbit/s. ;->

Share this post


Link to post
Share on other sites

Personally, I don't think the frequency 'effieciency' matters much. What's more important in my opinion is price/performance. I don't personally think the users should care if their computer is 1ghz or 10ghz. If Intel is requires 10ghz to do what AMD can do with 1ghz does it matter? IMHO no unless you have to pay more for the 10ghz Intel then the 1ghz AMD. This is the key reason why AMD IMHO is the better choice for most people. Regardless of the frequency 'effieciency' for most users AMD have for quite a while been the best price/performance. Of course, one of the problems is you can't come up with a universal price/performance since the diff between the AMD and Intel architecture means it depends on what app your talking about.

One efficiency that does matter IMHO is the power efficiency. I.e. performance/power usage. AMD appear to have this over Intel at the moment as well.

Of course, the difference between the architectures and Intel's frequency 'effiecy' is one of the key reasons for some of these problems but IMHO, the more important issues are the above not the frequency efficiency

Share this post


Link to post
Share on other sites

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!


Register a new account

Sign in

Already have an account? Sign in here.


Sign In Now
Sign in to follow this  
Followers 0