Memory Bandwidth

[TheRyuu]

I was just running through a few test's to see just how fast my computer is.

I ran the memory bandwidth test in SiSoft Sandra and it turns out my DDR500 RAM (2gb's of it) at 3-3-2-7 timings has a memory bandwidth of ~6,800mb/s. Seems like I'm maxing out the theoritical maximum bandwidth of DDR memory.

Now the weird thing is, thats higher than both the memory bandwidth of the 965EE with DDR2-800 ram from Intel, and the new AM2 running DDR2-800 ram from AMD.

WTF? I thought that DDR2 was suppose to give more memory bandwidth?

I also ran Everest Read and Write tests.
Read: 7616mb/s. (once again faster than the DDR2-800 ram used in the AM2 benchmarks)
Write: 7752mb/s. (what a surprise, faster than even DDR2-800)

Memory Copy: 6573mb/s (nothing to compare it to)
Memory Latency: 48.5ns (well, fastest thing on the chart)

 

[RichUK]

The theoretical max bandwidth for DDR500 in a dual channel config, would be 8,000MB/s, So your not really maxing it out at 6.8GB/s

 

[DrMrLordX]

Not really surprising. DDR2 just isn't very good performance-wise. It's cheaper to manufacture and uses less power. That's why it's being used.

 

[F1shF4t]

Well DDR 2 800 and DDR 400 are clocked the same speed except DDR2 has its I/O signals at twise the speed. DDR 2 has more bandwidth at the expence of much more longer latency.

 

[stevty2889]

Originally posted by: Dark Cupcake
Well DDR 2 800 and DDR 400 are clocked the same speed except DDR2 has its I/O signals at twise the speed. DDR 2 has more bandwidth at the expence of much more longer latency.

NO, DDR2 is still double data rate, it's DDR2 800 runs at 400mhz, DDR 400 runs at 200mhz..

 

[Brunnis]

Originally posted by: stevty2889
NO, DDR2 is still double data rate, it's DDR2 800 runs at 400mhz, DDR 400 runs at 200mhz..

No, the memory chips of DDR2-800 memory, for example, run at 200MHz. The buffers run at 400MHz and in DDR mode, thus motivating the "800MHz" designation.

 

[BitByBit]

Originally posted by: stevty2889

NO, DDR2 is still double data rate, it's DDR2 800 runs at 400mhz, DDR 400 runs at 200mhz..

DDR2 800 and DDR 400 both run on an internal clock of 200MHz.
DDR2 simply uses a 4-bit prefetch, whereas DDR uses 2 bits, allowing it to provide the theoretical bandwidth of a DDR module running at twice the internal clock.

The reason DDR2 has higher latencies in terms of external clock cycles is because its internal clock is slower.

Consider this:

We have a DDR 400 Module running with an internal clock of 200MHz and a DDR2 400 Module running with an internal clock of 100MHz.
Both can supply the theoretical bandwidth of a 400MHz SDR module while running at 1/2 and 1/4 the internal frequencies respectively. However, because they are not actually running at 400MHz, their latency in terms of 'effective cycles' is increased.
So if a 200MHz DDR400 module has a latency of 1 internal clock, that translates to a latency of 2 effective clocks, or CAS 2.
Now if a 100MHz DDR2 400 Module has a latency of 1 internal clock, that translates into a latency of 4 effective clocks, or CAS 4.

When comparing the latency in ns of DDR and DDR2, DDR2 should be treated as if it runs at the same internal clock as DDR, since the way their cycle latencies are calculated account for the differences.
The latency of CAS 2 DDR400 memory is calculated as follows:
Latency = Cycles / Frequency
= 2 / 200MHz
= 10ns.
The latency of CAS 4 DDR2 400 memory is calulated as above:
Latency = 4 / 200MHz
= 20ns.

It should be noted that it is this latency that the processor sees, and not the number of CAS cycles.

DDR3, as far as I know, will use an even larger prefetch and lower internal clocks to achieve the same or greater bandwdith. So a DDR3 800 module would run at an internal clock of 100MHz (compared with 200MHz for DDR2 800) but could have a latency of 8 cycles!

In my opinion, it is latency that should be addressed in future memory, not bandwidth.

 

[stevty2889]

Originally posted by: Brunnis

Originally posted by: stevty2889
NO, DDR2 is still double data rate, it's DDR2 800 runs at 400mhz, DDR 400 runs at 200mhz..

No, the memory chips of DDR2-800 memory, for example, run at 200MHz. The buffers run at 400MHz and in DDR mode, thus motivating the "800MHz" designation.

I stand corrected, you are correct.

 

[Dribble]

Originally posted by: wizboy11
I also ran Everest Read and Write tests.
Read: 7616mb/s. (once again faster than the DDR2-800 ram used in the AM2 benchmarks)
Write: 7752mb/s. (what a surprise, faster than even DDR2-800)

That says more about how good the memory controller is on 939 and how rubbish the memory controller is on AM2 then anything.

 

[Fox5]

Originally posted by: wizboy11
I was just running through a few test's to see just how fast my computer is.

I ran the memory bandwidth test in SiSoft Sandra and it turns out my DDR500 RAM (2gb's of it) at 3-3-2-7 timings has a memory bandwidth of ~6,800mb/s. Seems like I'm maxing out the theoritical maximum bandwidth of DDR memory.

Now the weird thing is, thats higher than both the memory bandwidth of the 965EE with DDR2-800 ram from Intel, and the new AM2 running DDR2-800 ram from AMD.

WTF? I thought that DDR2 was suppose to give more memory bandwidth?

I also ran Everest Read and Write tests.
Read: 7616mb/s. (once again faster than the DDR2-800 ram used in the AM2 benchmarks)
Write: 7752mb/s. (what a surprise, faster than even DDR2-800)

Memory Copy: 6573mb/s (nothing to compare it to)
Memory Latency: 48.5ns (well, fastest thing on the chart)

Hmm, memory latency is higher than I'd expect, I would have thought you'd have gotten down to around 40ns with that ddr.

 

[TheRyuu]

Originally posted by: Dribble

Originally posted by: wizboy11
I also ran Everest Read and Write tests.
Read: 7616mb/s. (once again faster than the DDR2-800 ram used in the AM2 benchmarks)
Write: 7752mb/s. (what a surprise, faster than even DDR2-800)

That says more about how good the memory controller is on 939 and how rubbish the memory controller is on AM2 then anything.

Well, I think it was stated in the AM2 DDR vs. DDR2 article on the anandtech website that AMD has had (i think) 4 core revisions and a few years to perfect the DDR controller in the 939 A64's and Opterons.

The memory controller (if you judge it by the Everest readings) on the newest 90nm A64's and Opterons (at least for me) is operating at ~96% efficiency.

Lets hope by the time AM2 is released that the can make that memory controller a little more efficient and actaully make use of the extra bandwidth (even though the A64 isn't dependant on memory bandwidth)

Originally posted by: Fox5

Originally posted by: wizboy11

Memory Latency: 48.5ns (well, fastest thing on the chart)

Hmm, memory latency is higher than I'd expect, I would have thought you'd have gotten down to around 40ns with that ddr.

Wow, would a hole 8ns. really make that much of a difference?

 

[zephyrprime]

Originally posted by: BitByBitIn my opinion, it is latency that should be addressed in future memory, not bandwidth.

Unfortunately, doing that is hard while merely increasing the speed of the interface is much easier.

 

[BitByBit]

Apparently, XDR's latency is in the order of ten times lower than DDR's, according to various sources.

"The biggest complaint made about DDR2 is their latency. Latency plays a role in the overall bandwidth. DDR has a latency of roughly 10 ns, while DDR2?s latency is closer to 15ns. This is why DDR beats DDR2 at the same speed. XDR's latency will be lower than that of DDR. It will be available in latencies of 1.25/3.0/2.5/3.33 ns. This will provide an additional advantage over DDR2. The clocks should remain on par with DDR2, but when XDR is clocked to higher speeds the gap will grow further."

Source

 

[Fox5]

Originally posted by: wizboy11

Originally posted by: Dribble

Originally posted by: wizboy11
I also ran Everest Read and Write tests.
Read: 7616mb/s. (once again faster than the DDR2-800 ram used in the AM2 benchmarks)
Write: 7752mb/s. (what a surprise, faster than even DDR2-800)

That says more about how good the memory controller is on 939 and how rubbish the memory controller is on AM2 then anything.

Well, I think it was stated in the AM2 DDR vs. DDR2 article on the anandtech website that AMD has had (i think) 4 core revisions and a few years to perfect the DDR controller in the 939 A64's and Opterons.

The memory controller (if you judge it by the Everest readings) on the newest 90nm A64's and Opterons (at least for me) is operating at ~96% efficiency.

Lets hope by the time AM2 is released that the can make that memory controller a little more efficient and actaully make use of the extra bandwidth (even though the A64 isn't dependant on memory bandwidth)

Originally posted by: Fox5

Originally posted by: wizboy11

Memory Latency: 48.5ns (well, fastest thing on the chart)

Hmm, memory latency is higher than I'd expect, I would have thought you'd have gotten down to around 40ns with that ddr.

Wow, would a hole 8ns. really make that much of a difference?

1. I don't think the efficiency of the ddr memory controller has increased all that much since the original launch of the opteron.
2. 8/48 is a 16 (2/3)% increase, which may be large enough to have a tangible increase in performance. Also, I believe that's lower latency than an L3 cache, and only about twice what AMD's L2 cache operates at. (not certain about that though)

Apparently, XDR's latency is in the order of ten times lower than DDR's, according to various sources.

"The biggest complaint made about DDR2 is their latency. Latency plays a role in the overall bandwidth. DDR has a latency of roughly 10 ns, while DDR2?s latency is closer to 15ns. This is why DDR beats DDR2 at the same speed. XDR's latency will be lower than that of DDR. It will be available in latencies of 1.25/3.0/2.5/3.33 ns. This will provide an additional advantage over DDR2. The clocks should remain on par with DDR2, but when XDR is clocked to higher speeds the gap will grow further."

XDR is not DDR3 though, is it?

 

[BitByBit]

Did I imply it was?

 

[Furen]

If I had to take a guess I'd say AMD tweaked its AM2 memory controller for lower latency while sacrificing maximum theorethical bandwidth. I may be wrong, of course, but AM2 actually yields lower latency with higher latency modules compared to S939.

 

[imported_Seer]

Rerun your test with your cpu running at the same speed as the cpu in those other tests. Pref. the same core, too. You will see lower results.

 

[TheRyuu]

Originally posted by: Seer
Rerun your test with your cpu running at the same speed as the cpu in those other tests. Pref. the same core, too. You will see lower results.

I know it would be lower. I'm just saying, since I'm running at DDR500 at low latencies I really have no need to upgrade to AM2 since my Bandwidth is about what the DDR2 bandwidth would be.

 

[SparkyJJO]

I knew the reason that DDR2 was in laptops first was because it pulled less power, meaning better battery life, but I thought it would be faster too.

 

[jiffylube1024]

DDR2 is higher latency memory for starters (I believe DDR2 800 MHz runs at 4-4-4 timings). To compound things, Intel doesn't use an onboard memory controller (which gives AMD with DDR1 and advantage), and AMD's AM2 currently is not well optimized for DDR2 yet.

IMO AM2 looks like a waste at this point in time. It will undoubtedly be more expensive than S939 at launch, meaning only those who have to buy a new system and don't upgrade for long periods of time should look into AM2 at its launch. I think waiting for Conroe or adopting A64/X2 (for those who haven't already - shame on you!) are better ideas.

 

[DrMrLordX]

Originally posted by: jiffylube1024

IMO AM2 looks like a waste at this point in time. It will undoubtedly be more expensive than S939 at launch, meaning only those who have to buy a new system and don't upgrade for long periods of time should look into AM2 at its launch. I think waiting for Conroe or adopting A64/X2 (for those who haven't already - shame on you!) are better ideas.

Why do you think people are speculating that AMD will raise s939 prices after AM2 launches? to make AM2 look better, that's why.

Conroe should bring AM2 and s939 prices crashing down. Or so it is hoped.

 

[SparkyJJO]

I'm just hoping AM2 ends up being better than it appears right now. If intel gets on top in the CPU world, well, someone get me a rope....

 

[Zebo]

Originally posted by: SparkyJJO
I'm just hoping AM2 ends up being better than it appears right now. If intel gets on top in the CPU world, well, someone get me a rope....

Don;t worry it won't last long. Intel has only lead for 8 months in last 6 years so I would'nt count AMD out yet. Remember Hammer is almost 3 years old, ancient in CPU world - They have K8L in wings which is said to double integer and FPU performance and a bunch more tweaks like PCIe controller and XDR? Not much is known but don't think they are goign to let intel dominate them.

As far as DDR2 it's crap for Hammer. Hammer is sensitive to latency, an increase in latency decreases performance significantly for Hammer and DDR2 has higher latencies which it's increaded bandwidth can not overcome even at DDR2 -800/1000. It's too bad anand did'nt do the testing right (ddr400 2-2-2) or y'all would have seen what a dog highly clocked DDR2800 was. Peoples best choice is get a 939 setup with some LL ram now or wait for conroes.

 

[DrMrLordX]

Originally posted by: Zebo
Peoples best choice is get a 939 setup with some LL ram now or wait for conroes.

Damn straight.

 

[JPH1121]

Hrm...an integrated PCI-e controller would make for some damn nice graphics cards but wouldn't do a whole lot unless the graphics card is designed around a high frequency PCI-e bus (I'm guessing)

regardless of what intel does with Conroe, I'm thinking that 65nm will easily make up for the performance and power gaps.

Better/quicker transistors, better power consumption, and significantly higher clock speeds will give the K8 architecture a new lease on life...