[Peder Rice]

http://www.kingston.com/hyperx/

For once, it looks as though varying speeds of DDR memory will actually matter (small instances aside).

At the bottom of the page are the new 500MHz DDR modules, which would fit oh-so nicely with the dual-2.0GHz PowerMac G5, and would probably offer a nice performance boost.

[kkant]

Originally posted by Peder Rice:
For once, it looks as though varying speeds of DDR memory will actually matter (small instances aside).

Yes, getting CAS2 DDR400 (as opposed to CAS 2.5 or 3) will help the speed some. See:

http://www.hardcorecooling.us/produc...=201&3=462

Originally posted by Peder Rice:
At the bottom of the page are the new 500MHz DDR modules, which would fit oh-so nicely with the dual-2.0GHz PowerMac G5, and would probably offer a nice performance boost.

Unfortunately, it would appear that faster-clocked RAM will not be utilized in the G5's. The RAM will be clocked at 400 (really 200 doubled-up), even if it is DDR500. I have asked about this many times, and no one has been able to say for certain that the motherboard will detect the clock-speed of the RAM chips and automatically uprate the memory bus timing. So I think that it won't happen.

[Commodus]

Apple isn't going to use DDR500 yet, simply because it's too new - it was announced weeks after WWDC!

However, don't be surprised if your early-2004 2.5 GHz PowerMac G5 has DDR500 on it. I can easily see Apple wanting to run that memory.

[Peder Rice]

What I'm curious about, and kkant touched on this, is will the 500MHz DDR memory provide a speed boost over the current 400MHz DDR memory in the PowerMacs (latency factored in)?

For example, if I throw in two 512MB 500MHz DDR chips into a dual-2GHz PowerMac, I should see a performance boost. Here, I'll provide some numbers:

Single 1.6GHz PowerMac
- 1600MHz CPU
- 800MHz FSB
- 333MHz DDR Memory

333 x 2 = 667 != 800
(667 / 800 = 83.3%)

Single 1.8GHz PowerMac
- 1800MHz CPU
- 900MHz FSB
- 400MHz DDR Memory

400 x 2 = 800 != 900
(800 / 900 = 88.8%)

Dual 2GHz PowerMac
- 2000MHz (x2) CPU
- 1000MHz FSB
- 400MHz DDR Memory

400 x 2 = 800 != 1000
(800 / 1000 = 80%)


Now, as far as I can see, those percentages show me that the Single 1.8GHz PowerMac makes the best use of its available bandwidth. The top-of-the-line Dual-2GHz PowerMac can't fill 20% of its bandwidth. Either way, each model should support higher frequency memory to fill this gap.

Single 1.6GHz PowerMac
Implements - 333MHz (PC2700) DDR Memory
Should Support - 400MHz (PC3200) DDR Memory

Single 1.8GHz PowerMac
Implements - 400MHz (PC3200) DDR Memory
Should Support - 466MHz (PC3700) DDR Memory

Dual 2GHz PowerMac
Implements - 400MHz (PC3200) DDR Memory
Should Support - 500MHz (PC4000) DDR Memory


EDIT: I'm talking of course, not about Apple using this high-speed memory, but users throwing it in their own systems for performance boosts.

[kkant]

Originally posted by Peder Rice:
What I'm curious about, and kkant touched on this, is will the 500MHz DDR memory provide a speed boost over the current 400MHz DDR memory in the PowerMacs (latency factored in)?

This is a good question, but I think the answer is "no". Here's why:

1) The motherboard has a certain memory bus clock, that you can't change (well OK, maybe you can change if you overclock, but I don't know if you can independently change the clocks of the FSB and the RAM). Therefore, the mem will run at 400 (or 333) no matter what RAM you install.

2) DDR500 has lower latency than DDR400 in an absolute sense, but not when measured by clock cycles. That is to say, both CAS2 DDR500 and CAS2 DDR400 require the same number of clock cycles to read or write. Of course if the motherboard is running at 500, then the total amount of time taken by 2 clock cycles is less than the time taken by 2 400mhz clock cycles. However, from point (1) we see that the motherboard is running at 400 no matter what, unless you overclock.

3) The motherboard sets the latency timings based on clock cycles, not absolute time. The SPD ROMs on the memory chips give CAS latency, like 2-3-3-6, which means 2 cycles to start a read or somesuch, etc etc. The motherboard reads the SPD ROMs at boot-time, and times memory accesses based on those CAS settings--regardless of whether the RAM is DDR400 or DDR500.

4) From a system efficiency perspective, it is "OK" for the RAM bus to be slightly behind the FSB--especially in a dual system. This is because the FSB is used by the processor(s) to communicate with everything else, not just memory. In partucular, the processors use the FSB to communicate with each other. So if the FSB is faster than the RAM bus and needs to wait for the ram, it can make productive use of that time by passing data from the other processor, or the AGP slot, or whatever.


Please feel free to argue or correct. I consider my points to be less statements of fact than bases for discussion.


EDIT: One corollary to all this is that CAS3 DDR500 ram would actually be slower than CAS2 DDR400 RAM, in a non-overclocked system.

[Peder Rice]

Originally posted by kkant:
4) From a system efficiency perspective, it is "OK" for the RAM bus to be slightly behind the FSB--especially in a dual system. This is because the FSB is used by the processor(s) to communicate with everything else, not just memory. In partucular, the processors use the FSB to communicate with each other. So if the FSB is faster than the RAM bus and needs to wait for the ram, it can make productive use of that time by passing data from the other processor, or the AGP slot, or whatever.

Your other points are completely valid. The CAS latency makes all the difference in the world, and that's why, ideally, a comparison would have to be made between memory of equal latency.

However, I cannot see how slower memory than what the FSB will support is a good thing. In AthlonXP systems that I've built and run, the general rule: use the fastest memory possible, just for that little improvement.

With the AthlonXP (not the new, higher-FSB ones), the jump to 400MHz DDR memory makes very little difference, except in allowing the other systems (hard drive, etc.) just a fraction of bandwidth, freeing up more bandwidth for the CPU and video card.

With the PowerMac, we have the case of everything waiting for the memory. Now maybe Apple noticed no significant change between the memory types and threw in the lesser modules for price, but for performance freaks, it'll make a difference.

Either way, it's fast enough for me, I just don't understand their position on this.

[Simon]

Originally posted by Peder Rice:
Now maybe Apple noticed no significant change between the memory types and threw in the lesser modules for price, but for performance freaks, it'll make a difference.

Since DDR500 came out after WWDC and therefore much after the G5 boards were designed your argument is only valid for the 1.6GHz model (the others had the fastest RAM around at the time).

In case of the 1.6GHz model I guess it's just cutting corners to keep the price down. Like the PCI slots and the missing 4 RAM slots.

[kkant]

Peder: I agree, it would've been nice if we could have DDR500 or at least DDR466 so that we get that last bit of performance in situations where memory is the limiting factor. Actually, what would be nice is if Apple provided "advanced config tools" which let you set the mem clock, FSB clock, etc. But I suppose that is a little too PC-BIOSish.

Simon: My question is, why not DDR466? This was available at good prices (to consumers) well before the WWDC announcement. Perhaps the issue was that quantities of even PC3700 were not sufficient; but I find it hard to understand why Apple could not contract with some supplier or another to make this happen.

[Svenmagnus]

Have you guys heard about this new technology called mram?
mram apparently does away with reloading data from the hardrive after startup. This will pretty much do away with startup times.

IBM said that it might come out in early 2004.

here's the link

[kkant]

Originally posted by Svenmagnus:
Have you guys heard about this new technology called mram?
mram apparently does away with reloading data from the hardrive after startup. This will pretty much do away with startup times.

IBM said that it might come out in early 2004.

here's the link

Sounds pretty cool. I wonder if it would be cheap enough to use in solid state HDs.

[Svenmagnus]

I found another article on mram, which says to not expect to see this in your personal computer anytime soon.

link

[tie]

Originally posted by Svenmagnus:
Have you guys heard about this new technology called mram?
mram apparently does away with reloading data from the hardrive after startup. This will pretty much do away with startup times.

How is this better than copying the RAM to the hard disk at shutdown, which theoretically computers already do? (I say theoretically, because for me it always crashes.) I guess computers will use less energy sleeping.

[Svenmagnus]

This is better because you dont have to wait for the computer to load the operating system and various applications into ram when you start up your computer. Its already loaded into mram. It uses less electricity and is faster then dram.

[reader50]

Originally posted by Peder Rice:
What I'm curious about, and kkant touched on this, is will the 500MHz DDR memory provide a speed boost over the current 400MHz DDR memory in the PowerMacs (latency factored in)?

For example, if I throw in two 512MB 500MHz DDR chips into a dual-2GHz PowerMac, I should see a performance boost. Here, I'll provide some numbers:

Single 1.6GHz PowerMac
- 1600MHz CPU
- 800MHz FSB
- 333MHz DDR Memory

333 x 2 = 667 != 800
(667 / 800 = 83.3%)

Single 1.8GHz PowerMac
- 1800MHz CPU
- 900MHz FSB
- 400MHz DDR Memory

400 x 2 = 800 != 900
(800 / 900 = 88.8%)

Dual 2GHz PowerMac
- 2000MHz (x2) CPU
- 1000MHz FSB
- 400MHz DDR Memory

400 x 2 = 800 != 1000
(800 / 1000 = 80%)


Now, as far as I can see, those percentages show me that the Single 1.8GHz PowerMac makes the best use of its available bandwidth. The top-of-the-line Dual-2GHz PowerMac can't fill 20% of its bandwidth. Either way, each model should support higher frequency memory to fill this gap.

Single 1.6GHz PowerMac
Implements - 333MHz (PC2700) DDR Memory
Should Support - 400MHz (PC3200) DDR Memory

Single 1.8GHz PowerMac
Implements - 400MHz (PC3200) DDR Memory
Should Support - 466MHz (PC3700) DDR Memory

Dual 2GHz PowerMac
Implements - 400MHz (PC3200) DDR Memory
Should Support - 500MHz (PC4000) DDR Memory


EDIT: I'm talking of course, not about Apple using this high-speed memory, but users throwing it in their own systems for performance boosts.

The calculations are flawed. You are taking MHz into account, but not bus width. MHz alone might work for serial connections, but the bus connections in question are parallel ones.

The CPU bus (processor bus in Apple literature) to the northbridge chip (U2 in Apple literature) is 2 channels that are each 32-bit wide, and is a DDR bus clocked at 1/4 the CPU clock. Because of the DDR part, this becomes effectively 1/2 the CPU clock in practice. For the 1.8GHz system, the CPU <-> northbridge bus would be 2 channels x 32bits x 900 MHz = 7.2 GB/second. Since control signals are also sent over the same bus, the effective useful bandwidth drops down to 6.4 GB/second.

Using this ratio for useful data and recalculating for the different CPU speeds and architectures, each system comes up like this on the CPU bus end:
Processor bus effective bandwidth
1 x 1.6 = 5.7 GB/s
1 x 1.8 = 6.4 GB/s
2 x 2.0 = 14.2 GB/s (each CPU has it's own processor bus to the northbridge)

On the memory end, dimm modules are 64-bit x the effective clock speed per second. PC100 used in older systems comes up as 100 MHz x 64 bits = 800 MB/s continuous bandwidth.

PC100 = 800 MB/s
PC1600 = 100 MHz doubled = 1600 MB/s or 1.6 GB/s
PC2700 = 166 MHz doubled = 2667 MB/s or 2.7 GB/s
PC3200 = 200 MHz doubled = 3200 MB/s or 3.2 GB/s
PC4000 = 250 MHz doubled = 4000 MB/s or 4.0 GB/s

All the G5 models use dual memory channels, so the bandwidth figures get doubled again. Each system comes up like this:
Memory bus effective bandwidth
1 x 1.6 = PC2700 x 2 = 5.4 GB/s
1 x 1.8 = PC3200 x 2 = 6.4 GB/s
2 x 2.0 = PC3200 x 2 = 6.4 GB/s
- theoretical overclock -
2 x 2.0 @ PC4000 x 2 = 8.0 GB/s

CPU / Memory bandwidth comparison
1 x 1.6: CPU 5.7 GB/s, RAM 5.4 GB/s
1 x 1.8: CPU 6.4 GB/s, RAM 6.4 GB/s
2 x 2.0: CPUs 14.2 GB/s, RAM 6.4 GB/s
- theoretical overclock -
2 x 2.0: CPUs 14.2 GB/s, RAM 8.0 GB/s

The 1.6 and 1.8 are fine, the RAM can keep up with the CPU. The dual 2.0 would benefit from a stable RAM bus overclock to use PC4000, or a different motherboard with quad RAM channels - add RAM in groups of 4. As it is, the dual 2.0's maximum speed will be dependent on the efficiency of the L2 cache controller in each PPC970.

Concerning the speed of RAM, the latency only affects initial reads, or changing the memory location to read from/write to. The specified speed of the RAM works for sequential reads, where you start at a location and just read block after block in. Going from DDR400 (cas2) to DDR500 (cas3) will cause a slightly longer delay in a new memory read (1 ns approx), but will deliver a 25% increase in sequential bandwidth. In the case of the dual 2.0, such an upgrade should be well worth it, especially in memory-intensive tasks such a media work. Such work mostly involves sequential reads anyway, and largely make the cas value irrelevant.

As it stands now, the dual 2.0 has 14.2 GB/s of CPU bandwidth to the northbridge. The northbridge has 6.4 GB/s to the RAM, 3.2 GB/s to the PCI controller, and 2.1 GB/s to the 8x AGP slot. That is 11.7 GB/s to everywhere except the CPUs, so they had best either spend time talking to each other or do some good cacheing. Using DDR500 would give the northbridge 13.3 GB/s to everywhere except the CPUs. You still could not max out the CPU bandwidth to just the RAM, but you could come closer to maxing out the CPU bandwidth to the system as a whole.

If I'm smoking the good stuff and someone in the industry wants to correct my explanation, feel free to chime in. But this is how I would analyze it.

[kkant]

Thanks for putting together this post, reader50. Makes sense.

[Peder Rice]

I'll try to explain myself and make me not look a fool. (I'll probably fail)


I'm looking at things more from the Athlon approach: speed up access for the hard drive, peripherals, etc., and you're bound to free up room for the CPU. Therefore, a jump to the 500MHz DDR memory in the 1.6GHz PowerMac would probably speed up file transfer during an intense render. Normally, all the bandwidth may be needed by the rendering to finish more quickly, but that file transfer that all needs to be finished is taking away from that bandwidth (an unlikely situation, but the same as we saw in the DDR PowerMac G4s). It's not much of a boost, but it helps.

But at the same time, that jump to 500MHz memory does significantly more. When you are forced to resort to dual-channel, and or DDR/QDR, you have to lift the 'theoretical' label and use real situations. Now, my above example won't happen very often, but the following one does.

The move to DDR memory allows data to move to and from the memory simultaneously, but there is always significantly more coming in than out (i.e. five variables going into a function with only one variable returning). Thus, the output bandwidth isn't as full as the input bandwidth. Bandwidth wasted; and the double becomes an inaccurate word. Maybe 1.5; I honestly have no idea how significant a loss it is.

Now we have a little less than our potential 6.4GB/s, and through some calculation that I forgot in mid typing, a jump to faster memory improves performance there as well.

(I'll get back to you on that when it comes to me... way too much code floating in my head)