OSNews

Well doesnt seem like amd is doing any better now are they? You haeve to consider that with the 'outdated' fsb stuff Intel has the perofrmance crown no matter what way you slice it...whereas AMD does have the integrated MMU and the Hypertransport. Kind of amazing to think what numbers the Intel architecture would be pusing out if it did go MMU and CSI...

Well doesnt seem like amd is doing any better now are they?
perhaps you missed http://www.sun.com/x64/benchmarks/ over 100 word record benchmarks for AMD cpus.

Edited 2006-12-17 21:26

Hey nice link man. Never saw that. Are there any Sun hardware running on Intel quad core processors? I would sure like to see the numbers from them!

No. Sun are an Opteron & SPARC-only shop.

As for the article, there's nothing unexpected in it - Intel craps on AMD for floating point performance; AMD craps on Intel for memory bandwidth. Clovertown doesn't change a thing - it just makes it more polar.

I have to agree about the FSB. IMHO this should be junked ASAP. The Memory to CPU bottlneck is Intel's biggest problem.
The raw CPU is fine but they are going to get hammered in late 2007 by the superior data shifting ability of AMD with their Hypertransport.
Server do a lot of bulk data shifting. At the present time, the biggest use of Quad Cores will be in the Server Market (Software Licenses permitting that is...) If you can't get the data in and out of the CPU quicker that your main competitor then you are going to have real problem.
It's ok for the likes of IBM & HP though as they have the resources to design good motherboards but for the rest of us, I see little or no gain over the current conroe CPU's.

I have to agree about the FSB. IMHO this should be junked ASAP. The Memory to CPU bottlneck is Intel's biggest problem. The raw CPU is fine but they are going to get hammered in late 2007 by the superior data shifting ability of AMD with their Hypertransport.

In reality AMD CPUs with superior data shifting ability get hammered by QX6700
http://www.hothardware.com/printarticle.aspx?articleid=911

In reality AMD CPUs with superior data shifting ability get hammered by QX6700
http://www.hothardware.com/printarticle.aspx?articleid=911

Quite frankly, considering that the Core Duos are a whole revision further on than AMD in terms of their die size and on 65nm, and AMD have reached their absolute limit with 90nm, those results are nothing short of an embarrassment for Intel.

From a multi-threaded application point of view, which is very relevant to servers and to this article, the Intels were taken to the cleaners:

http://www.hothardware.com/viewarticle.aspx?page=8&articleid=911
http://www.hothardware.com/viewarticle.aspx?page=10&articleid=911
http://www.hothardware.com/viewarticle.aspx?page=11&articleid=911

Basically, you have to buy a QX6700 to get anything that is significantly better than comparable AMDs, and you then have to ask yourself if it's actually worth the money to get it. I suppose that's the key question.

Additionally, Intel still don't have anything comparable to Hypertransport which works better as memory size grows which is important servers. That's why Intel have had to come up with FB-DIMMS, which AMD are somewhat less than keen on obviously. They also still don't have on-die memory controllers either.

As far as servers are concerned, AMD's architecture is just plain better. From a desktop and workstation (and servers - previous Xeons were terrible) point of view then Intel have improved significantly in performance from the rubbish they were producing before, but considering that Intel are still more expensive than AMD the question many people should ask is "Is it worth it?"

I have to admit your post sums it up quite well.

Intel do have a competitior to AMD's hypertransport, CSI, technologically there is nothing wrong with it, the problem is, however, its not going to be released until 2008, which quite frankly, is unacceptable given what changes AMD have announced for their future line up.

With that being said, one wonders why Intel feel the need to go with CSI - NIH syndrome I guess; because given the across-industry support, hypertransport would be a nice fit for what Intel needs.

With that being said, there was a good article on the merits and downside of a integrated memory controller - although I understand Intel wish to avoid the pitfalls of one vs. the other, their responsiveness, outside the laptop space, to AMD's product line up, has be quite pathetic to say the least.

Quite frankly, considering that the Core Duos are a whole revision further on than AMD in terms of their die size and on 65nm, and AMD have reached their absolute limit with 90nm, those results are nothing short of an embarrassment for Intel.

I fail to understand your point. How is Intel advantage with technological process is embarrasment for Intel?

Basically, you have to buy a QX6700 to get anything that is significantly better than comparable AMDs

Surely, you have to buy quad Intel cpu to get anything that is significatly better that quad AMD cpu. What did you expect?

, and you then have to ask yourself if it's actually worth the money to get it. I suppose that's the key question.

Sure it is, considering that quad AMD offerings are slower, more expensive (look at the Xeon vs Opteron or QX6700 vs FX series prices), and hotter (look at the QuadFX power consumption - twice as comparable Intel).

Additionally, Intel still don't have anything comparable to Hypertransport which works better as memory size grows which is important servers.

In theory, yes. In practice, as shown in multiple reviews, Core2-based platform is clearly ahead of Opteron or QuadFX one. In addition, there are NUMA coherency issues which requires fine-tuned VM and can significantly affect performance.

but considering that Intel are still more expensive than AMD the question many people should ask is "Is it worth it?"

It's cheaper or comparable. Check the prices.

How is Intel advantage with technological process is embarrasment for Intel?

Basically, because if their architectures really are that much better, moving to 65nm should have left AMD way, way, way behind on everything. Sadly for Intel, it hasn't, and their cheaper CPUs are not showing as much as they should be. You're going to have to buy the top end Core to get anything that is head and shoulders above an Opteron or even an Athlon.

Surely, you have to buy quad Intel cpu to get anything that is significatly better that quad AMD cpu.

Nope. Intel's CPUs are based on a 65nm process. They should be able to do several fold the work of what the AMDs can do.

In practice, as shown in multiple reviews, Core2-based platform is clearly ahead of Opteron or QuadFX one.

The reviews simply don't show that, including the one quoted. That was the point.

In addition, there are NUMA coherency issues which requires fine-tuned VM and can significantly affect performance.

Mmmmmm. No.

It's cheaper or comparable. Check the prices.

The day you can show me a chip that Intel produces that is as cheap as the Sempron, Athlon or Opteron is the day I might consider an Intel CPU. As it is, Intels are still ridiculously overpriced.

Basically, because if their architectures really are that much better, moving to 65nm should have left AMD way, way, way behind on everything.

And it actually did.

Nope. Intel's CPUs are based on a 65nm process. They should be able to do several fold the work of what the AMDs can do.

??? Smaller technological process doesnt improve performance directly, but allows higher frequencies and lower TDP.

The day you can show me a chip that Intel produces that is as cheap as the Sempron,

$40 Sempron, $40 Celeron. Who cares? Anyway, talk is about quad core not low-end stuff.

Athlon or Opteron is the day I might consider an Intel CPU. As it is, Intels are still ridiculously overpriced.

OK, let's look at the facts:

E6400 box $215
X2 4200+ box $215
(Intel is faster and cooler)

Opteron 285 $630
Xeon 5140 $470
(Intel is faster and cooler)

Opteron 2220 $750
Xeon 5150 $670
(Intel is faster and cooler)

QX6700 2.66Ghz $1000
2 x FX-74 3Ghz = $1000
(Intel is faster and twice as cooler)

As it is, Intels are still ridiculously overpriced.

:)


In practice, as shown in multiple reviews, Core2-based platform is clearly ahead of Opteron or QuadFX one.

The reviews simply don't show that, including the one quoted. That was the point.

Maybe you need to re-read the reviews?

http://enthusiast.hardocp.com/article.html?art=MTIzMywxLCxoZW50aHVz...
http://www.anandtech.com/printarticle.aspx?i=2772

In addition, there are NUMA coherency issues which requires fine-tuned VM and can significantly affect performance.
Mmmmmm. No.


Mmmmmm. Yes. When CPUs operate within the same address space (NUMA) each CPU should address the memory of other process via HT bus to other socket. So if the data is located in the memory of other CPU, the latency of the corresponding operation will be significantly higher than in case the data were in the processor’s own memory. The more sockets you add, the worse situation becomes, because it falls back to being limited by the number of HT connections and you have to multi-jump between sockets. So with multi-threaded loads where all threads work the same data (most cases) AMD NUMA implementation wouldnt perform much better than FSB because there is only one pipe to the memory on both the NUMA and the FSB model.

And it actually did.

Whatever. You can spend over the odds to get what you believe to be a significant performance increase. Even the reviews quoted say otherwise. Like I said - you need to buy a QX6700 to really get ahead of any of the AMDs. Not convincing when I'm looking for a good server CPU at the lowest price.

Smaller technological process doesnt improve performance directly, but allows higher frequencies and lower TDP.

Well yer (in theory - just ask any Macbook or laptop owner ;-))- and they still have nothing that will move right ahead of an AMD performance-wise apart from the QX6700. And guess what? It ain't cheaper or even level.

Additionally, over the years AMD have quoted a maximum TDP and Intel a typical TDP - and they still do. You'll find various differences around those figures with various benchmarks and reviews. A Core processor still eats over 20 watts at idle.

$40 Sempron, $40 Celeron. Who cares?

Intel simply have nothing to compare with the Sempron, or even the Athlons at the low-end.

E6400 box $215
X2 4200+ box $215
(Intel is faster and cooler)

Opteron 285 $630
Xeon 5140 $470
(Intel is faster and cooler)........

And so on and so forth. I'm just wondering why you're having to pull woefully slanted comparisons out of your ass and why you're having to buy two FX-74s :-).

Let's have a fair comparison of price. Here in the UK (prices pulled off nearest site and both with tax just to make it a bit realistic for someone to buy):

AMD Socket F Opteron 2216 2.4GHZ: £276.99 (545.596 USD)
Intel Xeon 5148 Active 2.33GHz Socket 771 4MB: £341.99 (673.663 USD)

I'm being fair and nice by picking that as well, and comparing both clock speed and the fact that both are Dual Core. Depending on what I pick Intels are at least twice the price the higher the spec I go for, and no Intel anywhere gets down to the price of that Opteron, combined with its power.

This price trend has been reflected for years now, all the way through Pentium 4s, Athlons, Xeons and Opterons. There's a reason why gamers largely pick AMD and why Opteron became so popular - they're very fast and just plain cheaper. The only difference is that Intel have now got their finger out and are producing the kind of performance they really should have been pushing out years ago for the ridiculous prices they charge. Not exactly headlines.

Mmmmmm. Yes. When CPUs operate within the same address space (NUMA) each CPU should address the memory of other process via HT bus to other socket. So if the data is located in the memory of other CPU, the latency...

I won't dignify it with a response, since it's pulled out of Intel's marketing literature.

Suffice it to say that after years of usage, Opteron's on-die memory controller and large memory usage is another area that has left Intel behind. FSB is widely known to suck so much ass it isn't believable, particularly when compared with symmetric NUMA which is important here. If you don't know that by now, well, can't help you mate....... Might I suggest Google?

Like I said - if I want a server to use large amounts of memory, to be good and performant at handling multi threading and processing and to be cheap then it simply isn't an Intel. Even now that Intel have shot some bolts by going with a 65nm die and pointing to every review under the sun it hasn't changed a thing, or any of their shortcomings.

I also don't really want a processor that eats over 20 watts at idle either, compared to around 7 watts for an AMD Brisbane processor. However, both companies produce a load of rubbish on an awful lot of TDP figures.

Edited 2006-12-19 22:45

and the fsb and non die mounted mmu still bites them in the a$$.

Well, if by "bittes in the ass" you mean significant performance lead over AMD offerings...

Even with the best of those numbers its only a 25% improvement

Yes, and it shows that octal-core Xeon setup is not limited by FSB.

and none of the benchmarks were typical uses for a server. Start a few large transfers of data from disk to ram or disk to network and then do the same benchmarks and see if the benchmarks still hold up.

http://www.anandtech.com/printarticle.aspx?i=2772
Intel is absolute performance leader in almost all benchmarks.

where are server type benchmarks?

those are all toy benchmarks where the main loop in the code fits in l2 if not l1 cache, and the total working set is only 4GB, so nothing really taxing to anything but the cpu.

where are server type benchmarks?
Well, if SPEC / SSL / Apache/PHP/MySQL / PostgreSQL / Java Webserving are not "server type benchmarks", I think you need to explain what do you mean by this term.

those are all toy benchmarks where the main loop in the code fits in l2 if not l1 cache

PostgreSQL or Apache core fits in 1-4Mb or 128Kb L1 cache? Man, I wish it were true!

Anyway, code data size has little to do with memory bandwith and latency limitations. So what's your point?

, and the total working set is only 4GB, so nothing really taxing to anything but the cpu.

I wonder, did you actually read the article? They tested 64bit systems ONLY running 64-bit Linux. WHere did you get this "4GB working set"?

Edited 2006-12-18 05:30

I think you need to scroll past the LMBench benchmarks and look at the server benchmarks in the 2nd half of the document.