This is why FB-DIMMs are a legitimate issue.

[suneohair]

just to show that multiprocessor optimization IS an issue, here is my experience using Handbrake.

It is supposed to be a multi-threaded wideo encoding program, and yet it didn't load the 4 cores of my MP. I assumed it was a memory latency problem.

But then I ran this test: I duplicated the Handbrake app and ran both applications at the same time encoding two videos (MPEG4) at the same time and I found out that both encoding processes were running at 115 fps, whereas one alone ran at 120 fps. The video processing power was now 230 fps, three times that of my PowerMac G5 (dual 2.5) and all four cores were loaded. So I assume it was a multiprocessor issue…

I love you. This is what I mean. The FB-DIMMs are an issue, that has been beaten to death. There is obviously a more important factor that is not being factored in. And that is the ability of applications to utilize the power of more than 2 cores/cpus.

I am awaiting the OPs return to comment on this matter.

 

[Lincoln]

Apple had no choice about the memory.

They wanted to have 2 processors in their pro desktops, the only chip that Intel have that can do this is the Xeon. The Xeon chip is designed as a server chip (primarily) as is the supporting chipset and memory. The latency of the FB Dimms isn't an issue with the sort of work that a server performs i.e. lots and lots of small tasks.

However, all of this doesn't change that fact that if you want the fastest Apple Desktop with the most expanadability the Mac Pro is your only choice.

Finally Cost. I configured a Mac Pro and 24" iMac as similary as possible:

• 2GB Ram
• 1x 500GB drive
• Bluetooth & Airport
• 7300GT -MP and 7600GT -iMac
• 2.0GHz Quad -MP amd 2.33Ghz -iMac
• 23" Apple Display - MP and 24" built in - iMac

Granted you could beef the MP up and areas and loose the wireless if thats what you want, but that would make them too different and could increase the cost of the MP. You may already have a display too.

The iMac costs £1,859 and the MP costs £2,658 a difference of £799! If the iMac provides what you need it is a bargain. But if you need more than 3GB Ram, 30" display support, more than 1 internal hard disk, 2 optical drives, etc, etc, then the iMac is no good!

 

[Trekkie]

I am sorry but what a useless test. The imac has a faster cpu. All the cores in the world aren't going to make for apps that are not multi core aware.

exactly, and the article even says it. The ones that won were CPU intesnive benchmarking, and the iMac has a 2.33GHz Core speed. Running that against a 2.0GHz core speed will mean it's faster, especially when the apps in question are not multi-threaded monster applications.

It's like drag racing school buses.

 

[devmage]

I read a lot about FB DIMMs as I went into this. They have some advantages and disadvantages. The Mac Pro as a whole is a very fast awesome machine I'm glad I ordered it. I got it yesterday

When I look at benchmarks and specs the things people say the FB latency effects is by a minor amount. The thing is still blazingly fast. It isn't like OMG my Amiga 500 is faster than this thing.

Plus if everything holds true, somewhere down the line we can all stick clovertowns in our Mac Pros and have 8 cores

The FB DIMM prices going up greatly distresses me. I was going to buy some after I sold my G5 and of course they have gone from $369 to $499. I guess I'll wait and see if they come down. Perhaps by the time I can make my sale.

 

[Trekkie]

The latency on an FB-DIMM is per channel depends on the number of channels provided by the system (Mac Pro has Four, at least the chipset does) Then, if I remember my numbers right.

Odd numbered DIMMs add 3ns

Even numbered DIMMs add 9ns.

Depending on the number of DIMMs per FBD channel, add it up, there is your memory latency.

This latency only comes into play when you're doing heavy memory operations. If your workload is in processor and the cache keeps it in there you're not going to be worrying about the latency of the memory subsystem.

 

[Sesshi]

*fchooom* (the sound of all that going straight over my head)

What is the exact disadvantage of FB-DIMM is non nerd speak? I've got the Dell "Mac Pro hit with the ugly stick, only better engineered and far better supported" Precision 690 with the dual 5160s and it's pretty darned fast in real-life usage.

And surely it's about usage. I use power-hungry apps or do concurrent work in multiple apps that takes good advantage of multiple cores. If you aren't a true power user you'd be far better served with the Core 2 Duo iMacs. Just as few people in the PC world buy the Precision 690s there's no need for the average Mac punter to buy a Pro.

 

[Lincoln]

To put it simply.

The Mac Pro could have been faster IF Intel had used a different memory technology (if possible). Even then it would depend on the Application being used. However, it would be debatable if this extra speed would be seen in real world apps.

Also, the iMac running at 2.33 Ghz is faster than the stock Mac Pro running at 2GHz (no surprise there) unless you use an application that can make use of all four processor cores. The cost of the iMac is less than the cost of the 2Ghz Mac Pro if you need a 23" Apple LCD with the Mac Pro and they have simlair configurations.

Finally, as soon as you put the ATi X1900 in the Mac Pro; the Mac Pro will be much faster then the iMac in any program that uses that card fully.

 

[Mackeyser]

This discussion is proof positive (to me, anyway) that Mac Pros are NOT going to be upgraded with Cloverton (Xeon 4 core chip) since they spec out at 2.4 GHz and internally, they are almost identical to the current Woodcrests, but use lower clock speeds.

Which means that unless you have apps that can max 8 procs, the Mac Pro at 2.66 or 3.0 GHz will run faster.

What I think we WILL see is an Xserve upgrade along with perhaps an Xserve Pro if you will which will have 2 or 4 chips. Cloverton FINALLY puts Apple in a position to make their servers the most powerful machines in their lineup.

I mean, while the 1U server is sweet, I think the Xserve has shown it is good enough to justify a bigger 2 or 4U big box. I mean a single 16 core box in a 4U box would be wicked sweet. Seriously, with an Xserve RAID, you could run 4 "servers" within one box easily.

Yeah, wicked sweet.

As for the Mac Pros, I think it makes no sense while they are still in backorder mode to switch processors. That would make the MacPros rev. A lose value immediately and create a TON of really upset buyers. Worse, those that bought the Cloverton Mac Pro would SEE that it is slower on just about everything...

 

[Sesshi]

To put it simply.

The Mac Pro could have been faster IF Intel had used a different memory technology (if possible). Even then it would depend on the Application being used. However, it would be debatable if this extra speed would be seen in real world apps.

Well everyone says this but noone explains why.
DDR2 is going to be faster? Why?

 

[amin]

Actually a 4x1GB with 4 dual rank DIMMs is optimal right now...

They are comparing to a PC with 2GB RAM, so they can't use a Mac Pro with 4GB RAM for comparison. A Mac Pro with 4x512MB will perform better and be cheaper than a Mac Pro with 2x1GB, so the former is what they should have used for the comparison IMO.

 

[snaf]

love FB Dimm

this is another good read about fb dimm tech

http://www.phoronix.com/scan.php?page=article&item=495&num=1

personaly i think the teething going on with fb dimm is understandable
but it seem to be the modules that are out that is causing latency issue
not so much the memory controller and fb dimm will support DDR3
so once production starts to ramp with other manufacturers they will start competing who will have the lowest latency dimm (well i hope so)

 

[Trekkie]

Well everyone says this but noone explains why.
DDR2 is going to be faster? Why?

No, that's why FB-DIMMs were developed. They couldn't get DDR2 to go any faster and have more than one or two DIMMs on the memory bus. So they developed the fully buffered technology to allow more DIMMs per memory port and increase in throughput.

Since then, things have changed.

The end result is that FB-DIMMs are the right technology for dual socket systems, and that as the review pointed out their technology had absolutely nothing to do with the Mac Pro being beat by the iMac

What the tests on the iMac 2.33GHz vs. a Mac Pro 2.0GHz proved is that if your workload is core intensive no matter how many you have you want the faster core. Something, BTW, that has been proved since the beginning of computing. Just because you have two things at 1.0GHz doesn't mean they're faster than one thing at 1.5GHz. The same applies here. You've got four at 2.0 and 2 at 2.33.

I work for IBM. I'm responsible for design inputs on our > 4 Way systems. We've been doing this for years. The Xeon MP has *always* had a slower front side bus, and a slower clock speed than the two way systems. There is always someone who will run FOlding @ home or some other CPU intensive 'benchmark' and then scream that their brand new $20,000 4-Socket computer is 'slow'

It's like drag racing a school bus.

The more processors are not to make one thing 'fast', they're there to do many things efficiently.

Think of it this way:

What the review did was drag race a mini-van vs. a Mustang and see how to get from point A to point B.

What it should have done was see how long it takes both to do a multi-core aware job, or say move 15 people 10 miles down the road. The mustang can only do four at a time, the minivan can do 8. So even though the mustang can go to 0-60 in 6 seconds and the Mini-Van takes 10 seconds...who wins the race of get 8 people down the road first? The Mustang or the Mini-Van?

The Mac Pro is *not* for everyone, Dual Socket Dual Core, or Dual socket eight core are not for everyone either. The more cores you put on a processor the slower it (initially) gets!

Look back at the short history of dual core processors already.

Pentium-D came in at 2.6, 2.8, and 3.0. I do not remember if there was a 3.2. Penitum was at 3.0, 3.2, 3.4, and 3.6....

Woodcrest is at 2.0, 2.66, 3.0. I would expect Clovertown to come in around significantly slower than the 3.0GHz. Then you can drag race a mini-van vs. a school bus. Because Woodcrest will be 'faster' at core intensive workloads will win again. You could run these exact same benchmarks and the 2x2.0GHz would probably win in several workloads agains a 2xWhatever.

So software guys are going to have to get better at writing multi-threaded applications before we can really enjoy all the cores Intel & AMD are cramming on their boxes.

 

[Sesshi]

Still it's not clear. You're mainly talking about multicore optimised apps, not the DDR2 or whatever else is out there vs FB-DIMM argument.

For me, in the way I work multicore clearly suits me the best. But I'm interested in what makes FB-DIMM inferior in any particular cases.

 

[Trekkie]

But I'm interested in what makes FB-DIMM inferior in any particular cases.

Nothing, the people that are complaining about it lack fundemental understanding of multi processor system design.

Just look at the systems, and you'll see the reason they use FB-DIMMs and not DDR2.

How many memory slots does the iMac or Mac Mini have...

How many memory slots does the Mac Pro have...

That's the reason. They didn't just put a couple of sockets in the iMac because of space, it's because they couldn't put more per channel on the chipset. DDR-2 maxes out at 2. Some designers have put four in, but they're using two DDR-2 memory channels to do it.

With FB-DIMM I can get more per channel, and have more channels to work with. Apple's implementation chose to do eight total, but you can do more (as noted in the Dell workstation comparisons)

For examples, look at Intel's system boards:

Quad DDR-2 Dimm Board
'Value' line of chipsets, again 4 max, two per channel

The Mac Pro's Chipset

Note the bandwidth available to the memory subsystems, compared to the previous.

Hope that helps in understanding that FB-DIMMs don't hurt anything in the right system. Only the price of them hurts our pocket books, and that doesn't look to be going down anytime soon since the market is bifurcated with DDR-2 and FB-DIMMs, and FB-DIMMs are DDR-2 memory with an extra buffer chip put on them, so DDR-2 will benefit from the volume and keep the price low, and FB-DIMM doesn't look to be doing so soon.

 

[MacsAttack]

Still it's not clear. You're mainly talking about multicore optimised apps, not the DDR2 or whatever else is out there vs FB-DIMM argument.

For me, in the way I work multicore clearly suits me the best. But I'm interested in what makes FB-DIMM inferior in any particular cases.

FB-DIMMs will give "poor" results in synthetic benchmarks designed to stress memory (i.e. something designed to highlight differences in memory latency).

The large cache size and improved fetch and pre-fetch logic Intel added to the Core 2 Duo signifiantly reduce the impact of memory latency - which is why the Core 2 Duo performs so well against the AMD processors which enjoy very low memory latency becasue they have the memory controler built into the CPU itself (almost as if Intel planned it that way... Oh wait...!)

 

[milozauckerman]

FB-DIMMs will give "poor" results in synthetic benchmarks designed to stress memory (i.e. something designed to highlight differences in memory latency).

That's not exactly true - FB-DIMMs will give 'poorer' results in any application that places heavy demands on memory calls - such as games, Aperture and Photoshop (among others), compared to a chipset/processor that uses DDR2.

The large cache size and improved fetch and pre-fetch logic Intel added to the Core 2 Duo signifiantly reduce the impact of memory latency - which is why the Core 2 Duo performs so well against the AMD processors which enjoy very low memory latency becasue they have the memory controler built into the CPU itself (almost as if Intel planned it that way... Oh wait...!)

Also not true - with Conroe and Merom (Core 2 Duo) and even the standard Core Duo line, the chips match or better AMD latency. Xeon, not so much.

And, again, the problems of FB-DIMMs are largely balanced out by bandwidth/FSB - but that places you in the situation of using a more expensive processing solution (and RAM) in order to match the performance of cheaper chip/RAM combinations.

 

[MacsAttack]

Thanks.

Nice to see you completly missed the point.

Thanks anyway.

 

[UnixMac]

I am sorry but what a useless test. The imac has a faster cpu. All the cores in the world aren't going to make for apps that are not multi core aware.

+1

It's a lame benchmark..