Anandtech Mac Pro Review, pt. 2

[jroad]

I didn't see this posted here, so:
Apple's Mac Pro - A True PowerMac Successor

 

[ksz]

Another great article from AnandTech. What concerns me most is the "absolutely horrid" performance of FB-DIMM memory with Intel's 5000X chipset. Conroe (Core 2) with standard DDR2 should have 58% better memory read performance and 29% better write performance.

It's not Apple's fault, but FB-DIMMs absolutely kill memory latency; even running in quad channel mode, the FB-DIMM equipped Mac Pro takes 45% more time to access memory than our DDR2 equipped test bed at the same memory frequency. Things don't get any prettier when we look at memory bandwidth either.

Additionally, when populating 4 memory slots to gain maximum performance from quad-channel mode, the average improvement over dual-channel is only 3.2%.

One of our biggest concern about the Mac Pro is that users who don't need 8 memory slots or four cores would be better off if Apple released a single socket Core 2 based Mac tower. The memory performance of FBD on the Intel 5000X chipset is absolutely horrid and there's nothing you can do about it unless you switch entirely to an all serial interface or go back to using regular DDR2 memory.

 

[quruli]

Workstation/Server memory has always lagged in performance to desktop memory. This is nothing new.

While unfortunate that Intel choose the FBDIMMs, hopefully if Intel decided to continue using this memory the performance will increase.

 

[ksz]

Workstation/Server memory has always lagged in performance to desktop memory.

Can you give some recent examples?

While unfortunate that Intel choose the FBDIMMs, hopefully if Intel decided to continue using this memory the performance will increase.

I hope so. At this time, this is of no consolation. FB-DIMM prices are high, power dissipation is high, availability is low, and performance is awful.

 

[Chone]

Can you give some recent examples?


I hope so. At this time, this is of no consolation. FB-DIMM prices are high, power dissipation is high, availability is low, and performance is awful.

I think it has to do with the ram being buffered and ECC, I think that whats make it slower, I can't rememeber how the whole thing went but it was something like that.

 

[DavidC1]

Another great article from AnandTech. What concerns me most is the "absolutely horrid" performance of FB-DIMM memory with Intel's 5000X chipset. Conroe (Core 2) with standard DDR2 should have 58% better memory read performance and 29% better write performance.

FB-DIMMs are one of the faults. The chipset is another. Though for Xeon workstations, 5000X is one of the better chipsets.

Can you give some recent examples?

http://www.tomshardware.com/2004/05...han_one_intel_p4_extreme_platform/page16.html

You can see from here: http://www.anandtech.com/mb/showdoc.aspx?i=2813

Pentium 4 at 3.06GHz is sometimes 20% faster than Single CPU Xeon 3.06GHz.

http://www.2cpu.com/articles/44_4.html

 

[tobyg]

There seems to be something people here just don't get. The Mac Pro is a PROfessional workstation. Professional workstations have always used ECC/Registered ram, if they were actually a true 'workstation'. I'm not talking a home PC or a crappy PC the receptionist uses. I'm talking the workstation the CAD users in back use, or the workstation the doctors use to run medical simulations. ECC/Registered ram helps verify the data going in and out of the ram is 100% good and correct. You have to give up some speed in order to make sure your data is 100% perfect. I have no problem with that sacrafice. The problem here is, most people here won't be using Mac Pro's for this sort of work or need this level of protection.

This is where the entire argument about the mid-level tower comes in. I personally have a feeling apple will no doubt have to figure something out with regards to this. Apple has ventured into new waters. They have PC converts coming over who expect expandability and performance. They want to game on these machines and use them as their primary desktops. They want the fastest box they can get, and don't care necessarily about the precision and data protection the Mac Pro offers. Just look at all of the PC vendors who sell machines with 2 drive RAID-0 setups. One drive fails, you lose everything. This would never be done on a professional CAD workstation (except maybe for a scratch/temporary disk just to help speed). The Mac Pro machines just don't fit this need.

Personally I welcome ECC/Registered ram. No, my data isn't worth a million dollars. But even so, my data is important to me. I store all of my data at home on a RAID 5 setup, and I back it up to tape every other night. Knowing the memory is protected helps me feel better, even at the expense of performance.

If you want a gaming machine, the Mac Pro may not be for you.

 

[ksz]

Apple's workstations have never required ECC memory. Non-ECC memory is extremely reliable and is in widespread use. The decision to use FB-DIMMs in the Mac Pro wasn't really a choice. Only Woodcrest chipsets support dual sockets at this time and those chipsets require FB-DIMMs. As mentioned on AnandTech, the fault does not lie with Apple, but with Intel.

Unlike you, I don't welcome the use of high-priced, high-power, but lower performing memory. Unfortunately Apple had no choice, but I suspect this is something that may well change next year.

I store all of my data at home on a RAID 5 setup, and I back it up to tape every other night. Knowing the memory is protected helps me feel better, even at the expense of performance.

Also, lack of hardware RAID suggests that the Mac Pro is not designed for server environments. For that you'll need to attach an external XServe RAID.

 

[tobyg]

Apple's workstations have never required ECC memory. Non-ECC memory is extremely reliable and is in widespread use. The decision to use FB-DIMMs in the Mac Pro wasn't really a choice. Only Woodcrest chipsets support dual sockets at this time and those chipsets require FB-DIMMs. As mentioned on AnandTech, the fault does not lie with Apple, but with Intel.

Unlike you, I don't welcome the use of high-priced, high-power, but lower performing memory. Unfortunately Apple had no choice, but I suspect this is something that may well change next year.


Also, lack of hardware RAID suggests that the Mac Pro is not designed for server environments. For that you'll need to attach an external XServe RAID.

No, Apple machines in the past haven't required ECC ram, but most workstation class machines have, in my experience.

It's a workstation. I store my data on a server with a raid-5 setup. My workstations don't have raid-5. The workstation processes the data in RAM and ECC is the only way to protect that data. The servers of course also have ECC/Registered ram.

If I were to use my Mac Pro to store important data, I'd insist on a hardware based raid-5 setup. Fortunately, this is just a workstation for me and it does its job as a workstation just fine.

As you suggested, you would require an external XServe RAID to store your data.

There are many people who do not want this level of protection and would rather have raw speed. This is why I do think they will have to address this market, as I stated. I have no doubt they will introduce some sort of Conroe/Core 2 Duo system. Just as they suprised us with the cube back in the day, they will suprise us again.

 

[ksz]

The quad G5 supports both ECC and non-ECC memory. Most workstations do the same.

I suspect that a successor to the 5000X may support non-ECC DDR2 memory at the expense of some DIMM slots, or simply realize the full potential of FB-DIMM. I like the idea of serialized access to memory -- much like serialized access to disk (SATA) -- because it is theoretically faster, more reliable (simpler and much narrower bus), and allows for more memory slots. But the technology is new and apparently hasn't been perfected. Even with the overhead of ECC, it should be faster than it is.

According to AnandTech, the peak theoretical bandwidth in quad-channel mode is 21.3 GB/sec, but the actual is 4.292 GB/sec, leading to a low efficiency rating of 20%. Compare this to 60% efficiency for DDR2.

 

[MacsAttack]

From the reviwe conclusions...

The memory performance of the Mac Pro is noticeably better than the PowerMac G5 and competitive with other products in the Mac lineup

Simple fact - the machine has been designed and built as a Workstation (and can also serve as a small server - which is why there is an option to have OS X server installed as part of the CTO). Both applications can call for more memory than you would ever install on a desktop games system. FBD memory is the best way to get all those memory slots into a system and hit the kind of price points that Apple is shooting for.

Apple have now completed their transition. During that period they have not introduced any "new" systems - the MacBook design was by far the most radical redesign. They may (or many not) produce some kind of mid-range system at some point in the future. That however would be a new from the ground up design aimed at a new market segment. Problem is, that segment looks like it is a small vocal group of hobbyists concentrated on Mac forums. For 99% of people the current lineup probably does very well indeed

Anyway. Apple have now completed their transition. Now they are in a position to start doing "new stuff". Whatever it may be. Probably in the new year. After Leopard ships.

 

[sirnh]

I believe that the memory benchmark used may not tell the whole picture. Most memory benchmarks use a single thread. If you look at the memory architecture for the Woodcrest w/ 5000X Chipset, you see that each processor has a Front Side Bus to its own bank of memory.

When a thread is locked on physical CPU 0, it has a higher latency when retrieving data from the 3rd and 4th channels that are assigned to CPU 1.

I believe a proper test might be to run an instance of the bench mark loop on each CPU, then look at the total memory throughput. It may not give an accurate amount of the real bandwidth of the memory bus, since there will be core contention for the FSB, but I will bet that you will find an increase in performance when channels 3 and 4 can be directly accessed by the cores assigned to them.

This chipset and processor wasn't designed to run a single high-load thread. You can't realize the full potential of the system until you take advantage of ALL of the available resources.

BUT I could be wrong.

 

[Fadl]

The numbers at anandtech from the Everest Memory Benchmark seems a bit too low for me. Look at the difference from dual to quad channel memory access. There is noch much difference. But when I look at other Woodcrest reviews there have to be much more bandwidth available.
To all Mac Pro users who can run windows: Please do a Sandra Sisoft 2007 and Everest test and post your memory results.
Maybe there is something wrong with the Apple mainboard or anandtech has been wrong.
Also please tell us if you have user dual or quad channel.

 

[Lollypop]

Im glad they made a push for a mid range mac in the conclusion, yes FBD memory is a issue, and its more expensive and its not as popular, for that you get reliability and more expansion... most of us dont need that, im not worried about it, all I want is to have a fast, quiet, mac that has a bit more expansion than a imac, but doesnt cost as much as the mac pro.

Very glad to see that the Mac Pro's price compared so wel againts the competition!

 

[DavidC1]

The numbers at anandtech from the Everest Memory Benchmark seems a bit too low for me. Look at the difference from dual to quad channel memory access. There is noch much difference. But when I look at other Woodcrest reviews there have to be much more bandwidth available.
To all Mac Pro users who can run windows: Please do a Sandra Sisoft 2007 and Everest test and post your memory results.
Maybe there is something wrong with the Apple mainboard or anandtech has been wrong.
Also please tell us if you have user dual or quad channel.

The reason Sandra Sisoft 2007 scores higher is because most uses BUFFERED portion of the benchmark. Unbuffered portion of the benchmark is considered to be more realistic portion of the benchmark and scores close to Everest: http://www.anandtech.com/mb/showdoc.aspx?i=2813&p=4("As we have been saying for years, however, the Buffered benchmark usually does not correlate well with real performance in applications on the same computer. For that reason, our memory bandwidth tests have always included an Unbuffered Sandra memory score. The Unbuffered result turns off the buffering schemes, and we have found the results correlate well with real-world performance as we will see shortly.")

http://www.tomshardware.com/2006/06...10.html#dual_vs_quad_channel_memory_bandwidth

Quad channel: 5.7GB/sec
Dual channel: 3.8GB/sec

Have a take a look at these two first:
http://www.anandtech.com/storage/showdoc.aspx?i=1799&p=10

The target market of SCSI drives towards workstation market allows it to be faster on appropriate applications, but slower in PC apps.

http://www.anandtech.com/showdoc.aspx?i=2307&p=6

3DLabs cards maintain competitiveness until the PC benchmarks, gaming.

So what happens should be same with Xeon. It would be suited to workstation apps where its supposed to, but fair poorly on PC apps, and Core 2 would be better.

I suspect that a successor to the 5000X may support non-ECC DDR2 memory at the expense of some DIMM slots, or simply realize the full potential of FB-DIMM.

The one of the reasons for FB-DIMM is that the requirements for ECC-DRAM disappears. There's no ECC/non-ECC FB-DIMM. The data integrity features are in all FB-DIMM. Having a non-ECC would mean a regular DDR2 memory. The memory controller may have to change significantly to support that. But the target market Xeons are in means for them, FB-DIMMs are gonna remain.

We may see 4x4-like one in the future, but they won't be suited to the target market Xeons are.

 

[DavidC1]

Im glad they made a push for a mid range mac in the conclusion, yes FBD memory is a issue, and its more expensive and its not as popular, for that you get reliability and more expansion... most of us dont need that, im not worried about it, all I want is to have a fast, quiet, mac that has a bit more expansion than a imac, but doesnt cost as much as the mac pro.

PCs may never need FBD memory, as we are fine with two slots, and capacity of memory per slot increases with advances in technology. Prices are higher for ECC DDR2, but we don't care, as we don't use ECC. Same is for FB-DIMM. I bet for target apps, FB-DIMM is faster than regular DDR2.

 

[tobyg]

I believe that the memory benchmark used may not tell the whole picture. Most memory benchmarks use a single thread. If you look at the memory architecture for the Woodcrest w/ 5000X Chipset, you see that each processor has a Front Side Bus to its own bank of memory.

When a thread is locked on physical CPU 0, it has a higher latency when retrieving data from the 3rd and 4th channels that are assigned to CPU 1.

I believe a proper test might be to run an instance of the bench mark loop on each CPU, then look at the total memory throughput. It may not give an accurate amount of the real bandwidth of the memory bus, since there will be core contention for the FSB, but I will bet that you will find an increase in performance when channels 3 and 4 can be directly accessed by the cores assigned to them.

This chipset and processor wasn't designed to run a single high-load thread. You can't realize the full potential of the system until you take advantage of ALL of the available resources.

BUT I could be wrong.

Another thing I read somewhere is FB-DIMM's can read and write memory at the same time, although not at the exact same speeds. I forgot where I read it, but I don't think any existing benchmarks will test for that sort of thing. I know it had something to do with the fact that the memory is buffered so it can actually read and write at the same time.

Yes, I ran the Everest tests and Sisoft Sandra test and the memory bandwidth is not the greatest. Like I said though, I don't mind having ECC ram at this point... I've had enough memory problems on PC's and I have always liked ECC memory if and when I can get it.

 

[ksz]

The one of the reasons for FB-DIMM is that the requirements for ECC-DRAM disappears. There's no ECC/non-ECC FB-DIMM. The data integrity features are in all FB-DIMM. Having a non-ECC would mean a regular DDR2 memory. The memory controller may have to change significantly to support that. But the target market Xeons are in means for them, FB-DIMMs are gonna remain.

That's a good clarification. FBD is not ECC. There is no error correction code in FBD. Instead, it is a form of registered memory, but with a large register array called AMB or Advanced Memory Buffer. This, coupled with the serial path to the FSB, ensures signal integrity and allows the motherboard to contain more DIMM slots without the perils of signal degradation over long circuit paths.

Benchmarks run by Tom's Hardware show a substantial 50% improvement in quad-channel mode over dual-channel. AnandTech should probably rerun their benchmarks with Sandra Unbuffered. I wonder why they didn't do this in the first place...

 

[Mr. Mister]

CAN I run non-ECC RAM in my Mac Pro?