Linux Shootout: Opteron 150 vs. Xeon 3.6GHz Nocona

danalien writes "Anandtech with their previous review have stirred up a bit of controversy, and they've released their follow-up review where they pit AMD's Opteron 150 vs Intel's Xeon 3.6 Nocona (on linux)."

Comments

Memory

[lachlan76]

To be able to show the real potential of the Opteron, you need to have more than one processor.

This lets you take advantage of the on-die memory controller, by letting each processor do it's own memory work, rather than making the Northbrige do all the work.

If you want to use a single processor, you might as well use an FX-Whatever, since they are just an Opteron without MP capability and only one HT bus.

Re:Memory

[kasperd]

I still wish everybody would dump the backward compatability with the 8086 though. The CPU still has to bootstrap in 16bit real mode FFS!!!!

I agree. But at least AMD did something right when designing the AMD64 architecture. Virtual 86 mode and segmentation was eliminated from the 64 bit mode, but they still exist in 32 bit mode of course. Completely eliminating 8086 compatibility was not really so much of an option. Backward compatibility is part of the reason for the AMD64 success. But it would have been nice, if they had at least offered a way to boot the CPU in 64 bit mode. As it is now, the CPU boots in 16 bit real mode, then you switch to 32 bit protected mode, and only after that have been done you can switch the CPU into 64 bit mode.

Re:Memory

[ottffssent]

Actually, even without a NUMA-aware OS, the worst-case dual (and almost quad) memory latency is less than a Xeon's.

What really sets the Opteron apart in MP scenarios is the bandwidth. Each chip gets 6.4G/sec to memory: add more chips, get more bandwidth. The Xeon on the other hand has to share its 6.4G/sec with all the chips in the system, which severely limits its scaling. A quad Opteron has over 25G/sec of aggregate memory bandwidth, while a quad Xeon is stuck with its 6.4G shared 4 ways. That's half the bandwidth of a 400MHz P4 - no wonder the quad Xeons are often barely faster than the duals.

Add to this that cache snoops and other bus traffic all have to share the same FSB on the Xeon whereas on the Opteron local memory accesses don't touch the HT links at all. For a standard 2P system, this frees up 3.2G/sec of HT link bandwidth, and a NUMA-aware OS only increases the efficiency of the system.

Despite Intel's recent marketing push, they really don't have the best CPU, and don't have the best system either. There are still considerable advantages to choosing a Xeon system but these days they have little to do with the chip or the board and a lot to do with Intel backing. That's an advantage that will quickly evaporate as industry gets comfortable with non-Intel parts.

Difficult to trust?

[arose]

As we can see above, the difference between the two CPUs seems exaggerated and difficult to trust.

Maybe it's because Intel still makes processors for MHz and not performance? Maybe because unlike some comercial vendors the POV-Ray Team doesn't feel the need to make processor specific optimizations and leave that to job to the compiler (where it belongs)?

Re:Difficult to trust?

[ezzzD55J]

This review struck me as a bit clueless, or unfinished. The above quote is a good example of why i think so. They do some measurements, but aren't sure they're doing it fairly (compiler flags), and don't know what to do with the results. There is little in the way of analysis or conclusions. With the openssl measurements they don't even give any conclusions. But analysis and conclusions are the whole point of the review, and a remark like "As we can see above, the difference between the two CPUs seems exaggerated and difficult to trust." really devaluates this review - they're just showing us measurements they're not sure are correct ('difficult to trust') and let us figure out what they're worth?

Well, the conclusion that the opteron kicks the xeon's ass is pretty inescapable to me, finding out opteron is available and the xeon isn't quite yet and more expensive, really closes the deal to me. But the review isn't very scientific, and didn't go very deep.

Re:Difficult to trust?

[BlowChunx]

It says to me that compiler flags matter a lot! ...oops, maybe those Gentoo zealots are onto something!

Re:Difficult to trust?

[alienw]

That's because it is nearly impossible to do a scientific comparison of two different processors. Anyone who tells you otherwise is a moron.

You have to evaluate performance (possibly vs price) for your particular application. If you need a faster processor for Doom 3, look at Doom 3 benchmarks. If you need to encode video, look at video benchmarks. If you need to do integer computations, look at integer benchmarks. Xeons probably kick AMD's ass at some applications, and AMD might beat the Xeon at others. You can't just say that one is "better" than the other in general.

Re:Difficult to trust?

[drinkypoo]

Yes, yes we are. :)

You definitely have to be careful about which ones you pick, though, and if you're really worried about performance you have to do what they did in this review, and try different settings with different programs because different flags will produce different results on different programs.

In general I use -O3 on older processors, like pentium2 cores, and -O2 on newer ones. I don't know if it's still true but -O3 was known to cause problems (errors, not just a performance hit) with GCC3 and Athlon processors, as well as MIPS processors, though that may have changed in 3.3 or 3.4.

Remember, gentoo's strategy of compiling everything is about more than cflags, it's also about the USE flags that ensure that you get support for the things you want, and skip the support for things you don't need. There's more than one kind of optimization.

Re:Opteron cpu hacked

[Ianoo]

Microcode updates aren't permanent though - you need to reload them every time the machine boots. So clearly you would need to reload these "hacks" using a piece of software during the boot process.

Also, the article admits that it's "very unlikely" that a particular processor could be fried using a dodgy microcode update, so why even mention it? It would be much easier to write a BIOS flashing virus, I believe a few of these did exist at one point (although the old memory is failing). I doubt the hoops you'd need to jump through to write such a thing for Intel processors are no higher than for AMD processors, and as such, this is just FUD.

Re:Opteron cpu hacked

[arose]

It would be much easier to write a BIOS flashing virus, I believe a few of these did exist at one point

Chernobyl.

Hyperthreading is not good for these benchmarks

[DJStealth]

As hyperthreading cuts the L2 cache in HALF, it should be disabled before doing any of these benchmarks. Hyperthreading only seems to improve the multithreading ability. These benchmarks being run on a single process are not realistic.

Re:Hyperthreading is not good for these benchmarks

[Glock27]

As hyperthreading cuts the L2 cache in HALF, it should be disabled before doing any of these benchmarks. Hyperthreading only seems to improve the multithreading ability. These benchmarks being run on a single process are not realistic.

Since it isn't possible to dynamically turn hyperthreading on and off while the system is running, the benchmarks should be run in the mode most systems will use - with (highly touted) HYPErthreading turned on. After all, it is supposed to be a useful feature...

Personally, if I really need to run two threads with top efficiency, I'll invest in a dual 2xx Opteron box - with two real processors rather than an extra sorta-processor. That way if I have to pay a dual CPU license fee at least I'll get my money's worth. :-)

Why this is still a joke

[Anonymous Coward]

Why don't they just summarize each test/graph with the PERCENT difference?

Instead, we get gems like this:

This becomes our first real world test where we see Intel come out ahead. This coincides with what we saw on the previous page with the synthetic benchmark.

On a result where the Xeon is 55.894 and the Opteron is 56.26... Since when is half a percent significant?

And where is the kernel compile benchmark?

Re:AnandTech Biased

[swordgeek]

Anandtech isn't biased, it's incompetent.

Don't get me wrong--I've liked Anand and company since they first hit the internet. They don't generally have an axe to grind or an ego to boost (both of which TomsHardware suffers from terribly), but they don't have the slightest bloody clue about statistics, or significance.

Fun to read, and not consistently biased, but not a great source of actual benchmarking or review information. (techreport.com is better for that)

Re:Opteron cpu hacked

[Sunspire]

I'm not sure if other Linux distributions do this already, but at least Fedora Core 1 and 2 both come with a processor microcode update service that runs in the bootup sequence. It's even enabled by default out of the box.

Linux has for a long time already mostly ignored the system BIOS since they're notoriously broken because of legacy reasons. Supplying known good microcode is simply another step in eliminating variables that make system testing needlessly complex, I predict we'll see more developments along these lines in general.

Re:set-up benchmarks?

[Anonymous Coward]

I expect the OP was comparing Athlon XP and P4.. (note that there was no mention of A64) Now, the main difference between XP and A64 is memory bandwidth/latency and guess where it shows the most? Yup. Games.

Fair comparison?

[illumina us]

This is comparing Intel's latest chip to a very old Opteron.

First of all, AMD's Opteron 150 is the highest performing AMD workstation CPU money can buy

What about the AMD Opteron 850?