DDR vs. SDR or Why more CPU bandwith will not help future macs

MacBandit · Aug 27, 2002

If anyone out there hasn't read this they should. It's a comparison of DDR L3 Cache to SDR L3 Cache.

http://www.powerlogix.com/downloads/SDRDDR.pdf

This exemplifies what i've thought all along. What I mean is it doesn't matter if the System Bus in the new Macs is Double Data Rate or not it will not help it because DDR for all normal purposes is not any faster. Where DDR will help is with Direct Memory Access to from Memory from the ATA, PCI, and AGP busses. We already have that so what we need now is a faster CPU.

Everyone be happy with the new Macs there will be noticeable speed differences in real world use in graphics processing etc..

MacBandit · Aug 28, 2002

I'm so dissapointed that I haven't received any replies. I thought someone would surely disspute this and maybe I could learn something from it.

alex_ant · Aug 28, 2002

This exemplifies what i've thought all along. What I mean is it doesn't matter if the System Bus in the new Macs is Double Data Rate or not it will not help it because DDR for all normal purposes is not any faster. Where DDR will help is with Direct Memory Access to from Memory from the ATA, PCI, and AGP busses. We already have that so what we need now is a faster CPU.

Everyone be happy with the new Macs there will be noticeable speed differences in real world use in graphics processing etc..

I was under the impression that the reason the new Power Macs are barely any faster than the old ones is that their DDR implementation is crippled. Although DMA performance has been improved, the MPC7455 CPU is still not capable of DDR, which means it's still starved for bandwidth. So no, DDR is for all normal purposes not any faster in the Power Macs, but it would be if the PMs had a CPU that took advantage of it.

Alex

Sun Baked · Aug 28, 2002

L3 CACHE TEST FAILURE

The old reliable BareFeats testing methods used to poke holes in DDR vs SDR theories rise up again to prove the same thing with L3 cache memory.

Unless you preload the cache with a 1MB file how is your huge Barefeats test going to prove anything but FSB speed limitations?

---

The sort of start out talking about the latency of the cache and the throughput, then immediately switch over to AND IN REAL WORLD TESTS THE DDR L3 CACHE IS USELESS and SDR is a good cheap alternative.

Seems this is solely a prep for putting a SDR cache on board their next set of CPU upgrades for L3 cache.

---

On the 1GHz 7455 Motorola is claiming throughput of L1/16GBs L2/10.67GBs DDR L3/4GBs with L3 being 1/4 clock and L2 being full clock but read every 3rd clock (something like that.)

Motorola's main comment on the L3 latency issue seems to be "This is another latency versus throughput confusion. The SRAM access time is the latency but once data starts coming back, the cache designer tells us they have shown in the past that they can sustain four (8 byte) data beats on every two clocks or or 4GB/sec at L3CLK = 250MHz."

nuckinfutz · Aug 28, 2002

That paper mentions Cache differences

You're talking a whole seperate issue when it comes to processor benefits.

Caching is determined by the OS or application so it's dependent on the app itself.

Processors are always looking to get information stuffed in their pipelines..the faster the better. Stalls for bubbles are to be prevented at all times.

Sun Baked · Aug 28, 2002

Anybody else have a clues as to what Powerlogix is doing with this set of DDR vs SDR CACHE tests?

Or what the motives are?

MacBandit · Aug 29, 2002

I realize that they're specifically showing cache tests. They are also showing DDR throughput potential which is only better then SDR when it's dealing with a very large file. In the real world most data we work with is made up of a bunch of little files thus reducing the advantage of DDR because it's access latency is not faster then SDR and in some instances slower.

MacBandit · Sep 6, 2002

Please explain then why when AMD went to full DDR with a DDR capable bus and everything they saw a less then 10% increase in overall speed. Not double the speed as you are implying by doubling the throughput of the RAM.

This is because most files on the hard disc are very small and to access a bunch of them requires time for each thus the latency. When they are finally transferred the file size is so small that the speed at which they are transferred hardly makes up for the time the CPU had to wait for the RAM to access. This is because the transfer time is nearly instantaneous compared to the access time. This is why SDR and DDR in most real world cases shows no real speed increase.

I'm not trying to fight about this I'm just looking for a good wholesome discussion and someone that can explain to me why what I'm saying is false. Though at this moment I believe I am correct.

amnesiac1984 · Sep 6, 2002

so macbandit, what your saying is that there is no advantage to the processor of having DDR FSB but the advantage is seen by the AGP, PCI and ATA busses which are directly connected to the system controller and thus running at full DDR with the RAM. thats the impression i got, correct me if I'm wrong?

With the extra mb of cache on the DP1.25's these may really rock.

gbojim · Sep 6, 2002

There seems to be some confusion regarding the different buses.

On the MPC7455, the bus that interfaces to L3 cache can read/write DDR or SDR SDRAM. What Powerlogix is doing in their tests is justifying the use of SDR to lower costs without really impeding performance. They are saying that the latency in getting ready to read the data from cache is the largest component of time actually taken to perform the read. The actual time to transfer data, once it is ready, is very short in comparison.

What they are saying is 100% true. What I do not know (and I don't have time to research it right now - anyone else know?) is whether you can stream data from L3 cache. If you could, you would see a more prominent performance increase with DDR L3 cache because the latency would come into play less often as you read larger chunks of data at a time.

But what about the other buses?

There is the system bus or FSB or whatever you want to call it. The MPC7455 only transfers data once per clock cycle and therefore cannot directly take advantage of the DDR technique of 2 transfers per cycle. This may or may not be really important for overall performance because it is very dependent on the rest of the system and the data you process for the reasons MacBandit mentioned. The folks working with FCP, large PS files etc., would see the greatest benefit of having DDR type transfer between the CPU and RAM. Most of us would see very little if any difference.

Then there is the bus connecting the peripherals. Apple has done a really good job with the DMA design to take advantage of DDR to move data around the system efficiently. But again, if you are working with large numbers of small files spread around the hard disk, you will not see the performance gains of someone working with large contiguous files.

One final thing that Poerlogix mentions that is very important is at least having L3 cache. Apple is quite correct when stating that the PC manufacturers do a pretty poor job of keeping the processor fed with data. The Motorola/Apple approach to that problem is L3 cache. You may argue whether that is the best solution or not, however if your OS and application code is efficient, it works very well.

Hope this helps.

MacBandit · Sep 12, 2002

Originally posted by gbojim
There seems to be some confusion regarding the different buses.

On the MPC7455, the bus that interfaces to L3 cache can read/write DDR or SDR SDRAM. What Powerlogix is doing in their tests is justifying the use of SDR to lower costs without really impeding performance. They are saying that the latency in getting ready to read the data from cache is the largest component of time actually taken to perform the read. The actual time to transfer data, once it is ready, is very short in comparison.

What they are saying is 100% true. What I do not know (and I don't have time to research it right now - anyone else know?) is whether you can stream data from L3 cache. If you could, you would see a more prominent performance increase with DDR L3 cache because the latency would come into play less often as you read larger chunks of data at a time.

But what about the other buses?

There is the system bus or FSB or whatever you want to call it. The MPC7455 only transfers data once per clock cycle and therefore cannot directly take advantage of the DDR technique of 2 transfers per cycle. This may or may not be really important for overall performance because it is very dependent on the rest of the system and the data you process for the reasons MacBandit mentioned. The folks working with FCP, large PS files etc., would see the greatest benefit of having DDR type transfer between the CPU and RAM. Most of us would see very little if any difference.

Then there is the bus connecting the peripherals. Apple has done a really good job with the DMA design to take advantage of DDR to move data around the system efficiently. But again, if you are working with large numbers of small files spread around the hard disk, you will not see the performance gains of someone working with large contiguous files.

One final thing that Poerlogix mentions that is very important is at least having L3 cache. Apple is quite correct when stating that the PC manufacturers do a pretty poor job of keeping the processor fed with data. The Motorola/Apple approach to that problem is L3 cache. You may argue whether that is the best solution or not, however if your OS and application code is efficient, it works very well.

Hope this helps.

Thanks, gbojim that's exactly what I have been trying to explain to people. I'm not trying to do this to burst peoples bubbles. I'm just trying to bring people down to earth and make them realise that the new macs as usual are the best macs ever made.

jefhatfield · Sep 12, 2002

Originally posted by gbojim

But again, if you are working with large numbers of small files spread around the hard disk, you will not see the performance gains of someone working with large contiguous files.

well said

many probably think ddr will help them in everything in a noticeable way, and this helps sales

i sure it helps amd's sales who seem to be getting more of their processors on machines again in this up and down high tech field

sometimes i wonder what happened to rambus and i seem confused about it

MacBandit · Sep 12, 2002

Originally posted by jefhatfield

well said

many probably think ddr will help them in everything in a noticeable way, and this helps sales

i sure it helps amd's sales who seem to be getting more of their processors on machines again in this up and down high tech field

sometimes i wonder what happened to rambus and i seem confused about it

What do you mean what happened to RamBUS? I believe there are Intel systems using RamBus though the latest DDR400 shows higher POTENTIAL throughput.

jefhatfield · Sep 12, 2002

rambus 800 certainly didn't live up to its name, or maybe it was too expensive as some have suggested...sometimes only, do i see the ram for sale, but i have not seen an intel machine with rambus on the store shelves for some time...too bad...i like lots of choices for pcs and macs

that is my issue with the pentium 4...not bad as a processor, but the athlon xp systems are a better value dollar for dollar, ghz for ghz, but intel has narrowed the price gap some

on the pc side, my gamer friends are happier with the athlon xp and usually the pc magazines seem to favor it over p4

rambus seemed to lose fans because of its price as much as anything

faster ddr ram does sound exciting though

MacBandit · Sep 12, 2002

Originally posted by jefhatfield
rambus 800 certainly didn't live up to its name, or maybe it was too expensive as some have suggested...sometimes only, do i see the ram for sale, but i have not seen an intel machine with rambus on the store shelves for some time...too bad...i like lots of choices for pcs and macs

that is my issue with the pentium 4...not bad as a processor, but the athlon xp systems are a better value dollar for dollar, ghz for ghz, but intel has narrowed the price gap some

on the pc side, my gamer friends are happier with the athlon xp and usually the pc magazines seem to favor it over p4

rambus seemed to lose fans because of its price as much as anything

faster ddr ram does sound exciting though

It seems to me that I heard that RAMBus was basically extorting manufacturors with there licensing scheme.

Also I think that the net Gaming Mags really like the Athlon because from what I've seen is the Athlon though slower in Mhz is still faster overall and we all know that gamers want the fastest system out there.

jefhatfield · Sep 12, 2002

Originally posted by MacBandit

It seems to me that I heard that RAMBus was basically extorting manufacturors with there licensing scheme.

Also I think that the net Gaming Mags really like the Athlon because from what I've seen is the Athlon though slower in Mhz is still faster overall and we all know that gamers want the fastest system out there.

i am way past the hard core gaming age set in my area, and i have been ask to take on teaching a new class this year...computer gaming

previously, i taught computer networking for the home and small business, but now one of the teachers has free games coming in for tryout and promotion so it's a good thing and they thought i could teach that since i am by far the youngest teacher there at 38...and maybe more in touch with kids/gamers

i know the athlon systems are cheaper, somewhat, to build from scratch, and the gamers i know brad about that...but then they turn around and buy a gazillion games they never play much and have xbox, gamecube, and ps2 sitting in their homes

it's a really big part of computer software, no doubt, but since the class is about computer gaming in general, maybe i can entice some of the people to mac gaming

i am not a hard core gamer, but i will have to learn really quick to become one soon, and to me, i find enough challenge on bugdom on my old ibook

ddtlm · Sep 12, 2002

Hate to have to point out the obvious, but the reason why doubling the bandwidth of the memory system (when going form SDR to DDR) does not double performance is because usually bandwidth is not the limiting factor. When it is, then yes you could see a doubling of performance.

For memory, more often then not, latency is the limiting factor. You generally need little tiny bits of memory as quickly as possible, not huge globs.

I maintain that the peformance of current Macs is not so much due to the memory system as it is due to low clock speeds. I am not worried about getting "true" DDR yet.

MacBandit · Sep 12, 2002

Originally posted by ddtlm
Hate to have to point out the obvious, but the reason why doubling the bandwidth of the memory system (when going form SDR to DDR) does not double performance is because usually bandwidth is not the limiting factor. When it is, then yes you could see a doubling of performance.

For memory, more often then not, latency is the limiting factor. You generally need little tiny bits of memory as quickly as possible, not huge globs.

I maintain that the peformance of current Macs is not so much due to the memory system as it is due to low clock speeds. I am not worried about getting "true" DDR yet.

This is what we have been saying. Thanks for the reafirmations though.

gbojim · Sep 13, 2002

Originally posted by MacBandit
What do you mean what happened to RamBUS? I believe there are Intel systems using RamBus though the latest DDR400 shows higher POTENTIAL throughput.

Not sure if you folks missed this or not, but Intel announced early this year (I think it was in February) that they are dropping support for Rambus in favor of DDR.

DDR vs. SDR or Why more CPU bandwith will not help future macs

macrumors 604

macrumors 604

macrumors 68020

macrumors G5

macrumors 603

macrumors G5

macrumors 604

macrumors 604

macrumors 6502a

macrumors 6502

macrumors 604

Retired

macrumors 604

Retired

macrumors 604

Retired

macrumors 65816

macrumors 604

macrumors 6502

Our Staff