Question Help me understand the memory subsystems of Alderlake DDR4, Zen 3 DDR4, Raptorlake DDR5 and Zen 4 DDR5.

[Markfw]

OK, so I thoughtt DDR4 in Zen 3 was memory speed and Bus speed (may not be the correct term), so like 3600 and bus at 1800 was about the best.. And Zen 5, right now I have 6000:3000:2000 but I am not sure what all those mean. Then I know my alderlake will only go to 3000 or 3200 before it does Gear 2. Then I asked about Raptor lake memory, as they are hitting 7400, and @Carfax83 was kind enough the explain his bus at 7400 was only 1850.

So if somebody could help explain all this, and which is faster. It would seem that Zen 4 at 6000 with a bus of 2000 would sort of be the fastest. But I don't get the big picture. And I am sure memory controllers and latency are in there somewhere too. Remember, I started computers when PC100 was it, and it ran at 100.

 

[A///]

I too look forward to this explanation because it's a lot more complex than I remember ram being.

 

[Markfw]

Here is my aida64 extreme(trial version). Kind of worthless.

 

[Carfax83]

Basically, DDR memory runs at a double clock rate technically speaking because it can transfer data twice per clock cycle compared to SDRAM which can only do it once. I think that is where most people get confused.

DDR5 10,000mhz memory isn't really running at 10ghz. It's running at 5ghz, but due to the double data dispatch it's marketed as twice the clock speed that it really is.

So your Zen 4 CPU uses DDR5 6000, which is actually running at 3ghz, the same as your memory controller since AMD engineered Zen 4 to be capable of using DDR5 in a 1:1 ratio with the memory controller. This is different from Alder Lake and Raptor Lake which both use DDR5 in a 2:1 ratio, meaning the effective DDR5 clock speed is twice that of the memory controller. That's why I said that even though I am running DDR5 7400, my memory controller is only operating at 1850. 7400/2 is 3700 and if you divide 3700/2 again, you get 1850.

As for which is better or faster, it's complicated. There are lots of factors involved, and I'm not knowledgeable enough unfortunately to explain it all. But basically, AMD knew what they were doing when they went with a DDR5 memory controller that was capable of running in gear 1 mode, because having a chiplet based architecture means they are more subject to latency. So running in gear 1 mode definitely helps with that. Intel isn't affected by latency as much because they have a monolithic architecture and can afford to run in gear 2, but with higher memory frequencies to make up the difference.

BTW, none of this is accounting for the fact that DDR5 has many latency fighting abilities compared to previous versions which is often why it improves gaming performance the way it does.

 

[VirtualLarry]

I remember the Core2Quad era, the FSB was a quad, memory was a dual, so 200Mhz was the mostly fastest you could go, I think. Or was that 400Mhz? 1600 QDR FSB, and 800 DDR DRAM. They had higher DRAM multipliers, you could go to 1066 DDR2, but I recall reading that there was no observable benefit on the CPU side to going higher than 1:2 ratio.

 

[Markfw]

Basically, DDR memory runs at a double clock rate technically speaking because it can transfer data twice per clock cycle compared to SDRAM which can only do it once. I think that is where most people get confused.

DDR5 10,000mhz memory isn't really running at 10ghz. It's running at 5ghz, but due to the double data dispatch it's marketed as twice the clock speed that it really is.

So your Zen 4 CPU uses DDR5 6000, which is actually running at 3ghz, the same as your memory controller since AMD engineered Zen 4 to be capable of using DDR5 in a 1:1 ratio with the memory controller. This is different from Alder Lake and Raptor Lake which both use DDR5 in a 2:1 ratio, meaning the effective DDR5 clock speed is twice that of the memory controller. That's why I said that even though I am running DDR5 7400, my memory controller is only operating at 1850. 7400/2 is 3700 and if you divide 3700/2 again, you get 1850.

As for which is better or faster, it's complicated. There are lots of factors involved, and I'm not knowledgeable enough unfortunately to explain it all. But basically, AMD knew what they were doing when they went with a DDR5 memory controller that was capable of running in gear 1 mode, because having a chiplet based architecture means they are more subject to latency. So running in gear 1 mode definitely helps with that. Intel isn't affected by latency as much because they have a monolithic architecture and can afford to run in gear 2, but with higher memory frequencies to make up the difference.

BTW, none of this is accounting for the fact that DDR5 has many latency fighting abilities compared to previous versions which is often why it improves gaming performance the way it does.

Excellent ! Exactly what I was looking for. But DDR4, so both AMD and Intel run 1:1 (gear one mode) up to 3200 or so, and AND can go up to 3733 I have seen, but that is not common. So at 3200, 1600 is the ram speed ? And as I said, I know my 12700F goes to gear 2 above 3200.

 

[Shmee]

I remember the Core2Quad era, the FSB was a quad, memory was a dual, so 200Mhz was the mostly fastest you could go, I think. Or was that 400Mhz? 1600 QDR FSB, and 800 DDR DRAM. They had higher DRAM multipliers, you could go to 1066 DDR2, but I recall reading that there was no observable benefit on the CPU side to going higher than 1:2 ratio.

Yeah I had a Q6600, OCed to like 3.4GHz. So my rated FSB would have been about 1511 instead of 1066 I think. I had it with 2x2GB of Gskill DDR2 1100, was a great kit. I think I ended up running it at like DDR2 1000 or something, with tightened timings. Or maybe it was like 1133? So FSB to DRAM would be like 2:3 or 3:4.

 

[Thunder 57]

Excellent ! Exactly what I was looking for. But DDR4, so both AMD and Intel run 1:1 (gear one mode) up to 3200 or so, and AND can go up to 3733 I have seen, but that is not common. So at 3200, 1600 is the ram speed ? And as I said, I know my 12700F goes to gear 2 above 3200.

3200 Is MT/s (Megatransfers/s), and yes runs at 1600MHz.

 

[eek2121]

Here is my aida64 extreme(trial version). Kind of worthless.
View attachment 70680

Wow, that latency...

What are your speeds/timings?

I am down to about 65ns thanks to the latest BIOS update Shamino provided for ASUS users, and I haven't tweaked timings yet.

Note that AM5 boards all have immature AGESA versions and BIOS/UEFI versions. As an example, a single beta BIOS dropped 5-10ns off my memory latency and added more than 10% to my read/write speed. Synthetic benchmarks increased between 3-9% as well. Gaming is also up, but percentage varies by game.

I bring this up because we won't have a real picture of how DDR5 performs with AMD for quite a while. Currently, on many boards, bumping clocks above DDR5-6000 (3000mhz) kicks you into 2T mode, for example. My board doesn't even have options to disable GDM or change the command rate currently.

EDIT: AM4 had similar teething issues. My first AM4 board would not work with the memory I bought for it at XMP. It took AMD almost 2 years to fix memory support so that it would work with XMP timings. They eventually did fix it.

AMD definitely needs to improve in this area, IMO.

 

[Markfw]

Wow, that latency...

What are your speeds/timings?

I am down to about 65ns thanks to the latest BIOS update Shamino provided for ASUS users, and I haven't tweaked timings yet.

Note that AM5 boards all have immature AGESA versions and BIOS/UEFI versions. As an example, a single beta BIOS dropped 5-10ns off my memory latency and added more than 10% to my read/write speed. Synthetic benchmarks increased between 3-9% as well. Gaming is also up, but percentage varies by game.

I bring this up because we won't have a real picture of how DDR5 performs with AMD for quite a while. Currently, on many boards, bumping clocks above DDR5-6000 (3000mhz) kicks you into 2T mode, for example. My board doesn't even have options to disable GDM or change the command rate currently.

EDIT: AM4 had similar teething issues. My first AM4 board would not work with the memory I bought for it at XMP. It took AMD almost 2 years to fix memory support so that it would work with XMP timings. They eventually did fix it.

AMD definitely needs to improve in this area, IMO.

First, the test may be flawed, because I have the CPU at 100% load, including the caches. I can rerun after my contest is over. But regardless, it gave me an error that my cpu and config could not work with this software (or something like that) so. the explanation above is really what I wanted.

BUT I am running 6000 cl30-38-38-96 EXPO memory. with what I am sunning is absolutely killing everything else I have, including (these are 8 core tasks) my 12700F with e-cores disabled. In other words it kills a 12900k and most likely a 13900k at this task an thats 8 cores to 8, but this runs 2 tasks, which the 13900k can not do with p-cores. det0x has the good numbers for my memory, I am just running all out 100% which he is not.

 

[RTX]

OK, so I thoughtt DDR4 in Zen 3 was memory speed and Bus speed (may not be the correct term), so like 3600 and bus at 1800 was about the best.. And Zen 5, right now I have 6000:3000:2000 but I am not sure what all those mean. Then I know my alderlake will only go to 3000 or 3200 before it does Gear 2. Then I asked about Raptor lake memory, as they are hitting 7400, and @Carfax83 was kind enough the explain his bus at 7400 was only 1850.

So if somebody could help explain all this, and which is faster. It would seem that Zen 4 at 6000 with a bus of 2000 would sort of be the fastest. But I don't get the big picture. And I am sure memory controllers and latency are in there somewhere too. Remember, I started computers when PC100 was it, and it ran at 100.

AMD's mem controller runs at 1/3 of the MT/s value ( 2066 x 3 = 6200 MT/s )
On Intel, 2066 on the mem controller results in 8266 MT/s ( 2066 x 4 ) in Gear2
For AMD to run 8266 MT/s, the mem controller needs to be 2755mhz in Gear2

 

[Carfax83]

So at 3200, 1600 is the ram speed ?.

You got it!

 

[Carfax83]

Man, this Hynix A die memory is so easy to overclock. I punched in the same timing values for G.Skill's DDR5 7600 into my ram, and presto chango I now have DDR5 7600.

 

[AdamK47]

First, the test may be flawed, because I have the CPU at 100% load, including the caches. I can rerun after my contest is over. But regardless, it gave me an error that my cpu and config could not work with this software (or something like that) so. the explanation above is really what I wanted.

BUT I am running 6000 cl30-38-38-96 EXPO memory. with what I am sunning is absolutely killing everything else I have, including (these are 8 core tasks) my 12700F with e-cores disabled. In other words it kills a 12900k and most likely a 13900k at this task an thats 8 cores to 8, but this runs 2 tasks, which the 13900k can not do with p-cores. det0x has the good numbers for my memory, I am just running all out 100% which he is not.

Not sure why you would even bother running it at all then.