Understanding DDR5

[Avaviel]

  

On 2/13/2024 at 11:25 PM, RONOTHAN## said:

This screenshot was on a 13700K with DDR5 7200 CL34-44-40-40 with custom sub-timings, so about as fast as you can hope for. 

 

A lot of this will come down to specific software that you're using and what has the least amount of overhead. This was the first one I found (ImDisk), so take that for what you will. For reference, with the 2nd piece of software I found (SoftPerfect RAM Disk) performs significantly differently on the exact same setup:

This was over in software, but it's a specific hardware question... so posting here instead of a reply to the old thread.

I done goofed.

Bought DDR5. Thought I had a good one... It's DDR5 6400. Like large size, but slow. (Compared to other DDR5 sticks.) Would be great if I needed lots of RAM.
 

Budget (including currency): $300ish USD

Country: USA

Games, programs or workloads that it will be used for: Work, trying to get really fast RAM disk.

Other details:

GIGABYTE X670E AORUS PRO X
 

AMD Ryzen 9 7900X

Corsair MP700 Gen5 1TB


I did this upgrade today, and fell victim to the standard mistake: Not looking at details and going for size over speed. I bought this: https://www.newegg.com/g-skill-96gb/p/N82E16820374499?Item=N82E16820374499 DDR5 6400. I can return it still.

I am trying to get very fast DDR5 to use as a RAM disk. This is much faster than what I had, but it isn't the speeds I was hoping for. I'm sure there's some config to help, like XMP (I think that's it) is enabled.

Would this be a good choice? https://www.newegg.com/team-48gb/p/N82E16820985110?Item=N82E16820985110

It's less total RAM, but much faster.

Also, This: https://www.bestbuy.com/site/corsair-vengeance-32gb-2x16gb-7200mhz-pc5-57600-c34-intel-xmp-udimm-desktop-memory-with-rgb-lighting-multi/6554930.p?skuId=6554930

Cheaper, and I could go for 64GB but be stuck. (And isn't 4 sticks slower than 2? I remember something like that.)

 

 

 

 

 

 

 

 

 

 

[Electronics Wizardy]

I'm pretty sure CPU overhead and the design of the RAM disk is much more of a limit than ram speed. DDR5 6400 is pretty fast for most uses. I have DDR5 4800 on a 12900k and got like 60GB/s reads with fio on linux.

 

What CPU + Mobo is this for?

[RONOTHAN##]

3 minutes ago, Avaviel said:

Would this be a good choice? https://www.newegg.com/team-48gb/p/N82E16820985110?Item=N82E16820985110

I'd be shocked if the XMP worked on it. 

 

3 minutes ago, Avaviel said:

Also, This: https://www.bestbuy.com/site/corsair-vengeance-32gb-2x16gb-7200mhz-pc5-57600-c34-intel-xmp-udimm-desktop-memory-with-rgb-lighting-multi/6554930.p?skuId=6554930

With a single kit, the XMP should work most of the time (not on every CPU, but more than it doesn't), but if you're doing 4x16GB there's no shot of it working. 

 

 

With DDR5, the speeds you want do greatly depend on the platform you're trying to run it on. With AMD, pretty much your only option is to buy DDR5 6000 or slow down a kit you buy to DDR5 6000. The AMD memory controller has two modes, 1:1 and 2:1 mode, where 1:1 gets significantly better latency (from the last post, this is what you were primarily wanting) and 2:1 mode gets significantly higher memory speeds. Unfortunately, that higher clock speed doesn't translate into much higher bandwidth, and by the time you're clocked high enough to overcome the latency penalty of 2:1 mode, stability becomes hit or miss on pretty much anything but the X670E Gene. Speeds above 6000MT/s default to 2:1 mode, so unless you want to manually go to 1:1 mode and test for stability it's not worth the effort. 

 

On Intel 13th/14th gen, there's also memory controller issues, but it's more so in that it's incredibly inconsistent rather than performance penalties . On any given motherboard, the range of settings that work can be over 1000MT/s, so some chips will max out at 6800MT/s while others will top out at over 8400MT/s. If you want it to just work and not to have to think about stability testing it (since this is a workstation, you will want it to be fully stable), you want to stick to 6400MT/s max as that's the maximum speed I'd expect to just work on 99.9% of CPUs on most motherboards. 

[RONOTHAN##]

1 hour ago, Electronics Wizardy said:

I'm pretty sure CPU overhead and the design of the RAM disk is much more of a limit than ram speed.

Forgot about the overhead in my initial post, so I just spent the past 30 minutes checking to see how bad it actually was. 

 

5600 CL46 (JEDEC, complete auto timings)

Spoiler

6400 CL32-39-39-102 XMP

Spoiler

8000 CL36-47-47-128 "XMP" (auto subtimings)

Spoiler

6400 CL32-39-39-40 easy sub-timings (copied from Buildzoid):

Spoiler

7200 CL34-45-40-40 Maxed out sub-timings:

Spoiler

 

 

So yeah, the RAM Disk software is definitely the bottleneck here, since with the exception of sequential reads the performance doesn't change much. Even then, the timings have a bigger affect on the sequential read performance than the frequency does for whatever reason, so going for an 8200 CL38 kit makes no sense in this application. 

[brob]

1 hour ago, Avaviel said:

but it isn't the speeds I was hoping for.

 

What speeds are you hoping for?

 

Is that hope realistic, i.e. any reports of others achieving those speeds?

 

[QuantumSingularity]

I am really intrigued to know what applications is the ram disk software still used for? I mean yes, when all we had was mechanical drives, it used to accelerate things to then unimaginable speeds. And yes, the write speeds on RAM drives are usually WAAAAAY higher than on any modern SSDs, but the reads are twice the speed at best. The last time i ran a game installed on DDR4 memory drive and compared it to even a regular old SATA SSD, the difference was barely noticeable and certainly wasn't present when compared to even a Gen3 NVMe drive. Yes in synthetic tests it's a big difference, but what real life application manages to take full advantage of that speed? 

[Avaviel]

Thanks everyone!

Part of this is that I don't really understand what timings mean... previously the goal was always 'number big' get two sticks of 16GB DDR3 and 1600 and I'm golden. But that was the last time I upgraded parts, and it wasn't for work.

So this is stemming from this thread here, where I'm trying to find some Ram Disk software, and choose to go with it or Gen5 Raid 0. When I saw even the bad DDR5 Ram disk has 2000MB/s random read and writes I thought, "I'll get that!" and proceeded to get the wrong one.

I'm doing 3D CAD stuff for work. There are... odd slow downs that I can only attribute to tiny reads. I've replaced part by part trying to find what's making it slow. Tekla Structures is the software. I linked video in the older thread. I've been modeling with it extensively, and there are times when I go too fast for the software.

Much of that is from it being older software. But some of it is because it reads a database that is the CAD model over all, with tiny bits of info. Here's an example of a report... I'm convinced each time you select something, and I'll have... many items I sometimes need to select.

I just selected everything in the model. I don't often have to do this, and avoid it as much as I can. Still, it's like five minutes of flashing program, black screen to to regular trying to select 51882 objects. I'm used to seeing the "not responding, I'm thinking" screen. This one is frustrating, it's in a loop... I break the loop by left clicking, but also deselect all the things. (This gives me an idea though.)

Spoiler

-------------------------------------------------------------------------------------------------
  Name                Profile               Material         Grids     Part           Assembly
                                                                       position       position
-------------------------------------------------------------------------------------------------
  COLUMN              HSS4X4X1/4            A992              3/L         C106           C106
-------------------------------------------------------------------------------------------------
  Total 15 Parts: 1834.53 lbf, 143'-9"
=================================================================================================
 
 Part                       GUID: 217801fb-ee4b-41e8-b612-c5991647d37d
 -------------------
 Global coordinates:
  Start point                   :  X= 20'-0"           Y= 115'-0"             Z= 0'
  End point                     :  X= 20'-0"           Y= 115'-0"             Z= 9'-7 5/8"
  Center of gravity             :  X= 19'-10"          Y= 115'-0"             Z= 4'-10 1/8"
  Top level                     : 9'-7"5/8
  Bottom level                  : 5/8
 -------------------
 Local coordinates, UCS:
  Start point                   :  X= 20'-0"           Y= 115'-0"             Z= 0'
  End point                     :  X= 20'-0"           Y= 115'-0"             Z= 9'-7 5/8"
  Top level                     : 9'-7"5/8
  Bottom level                  : 5/8
 -------------------
  Part position                 :  C106
  Assembly position             :  C106
  Net length                    :  9'-7"
  Gross length                  :  9'-7"
  Weight                        :  109.91 lbf
  Weight(Net)                   :  122.30 lbf
  Weight(Gross)                 :  109.91 lbf
  Volume                        :  431 cu.in
  Area                          :  1844.00 sq.inch
  Name                          :  COLUMN
  Material                      :  A992
  Finish                        :  RED OXIDE
  Profile                       :  HSS4X4X1/4
  Flange slope ratio            :  0
  Rounding radius 2 (r2)        :  0
  Rounding radius 1 (r1)        :  1/2"
  Plate thickness (t)           :  0
                                
  Width (b)                     :  4"
  Height (h)                    :  4"
  Class                         :  11
 -------------------

More:
Load bearing                  : No
LEVEL:                        : 1
Type for BOM:                 : HSS
Gage Material:                : HSS

Of course, part of the solution is finding how to treat the program. But, if I can 'select all objects' without freezing... I can EVERYTHING I need to. (And yes there are filters and stuff, so 90% of the time it isn't like that)

Here I have just a bit of an example of what I'm doing.
https://youtu.be/pUdca7e71oA

 

[Avaviel]

On 2/16/2024 at 11:06 PM, RONOTHAN## said:

Forgot about the overhead in my initial post, so I just spent the past 30 minutes checking to see how bad it actually was. 

 

5600 CL46 (JEDEC, complete auto timings)

  Hide contents

6400 CL32-39-39-102 XMP

  Hide contents

8000 CL36-47-47-128 "XMP" (auto subtimings)

  Hide contents

6400 CL32-39-39-40 easy sub-timings (copied from Buildzoid):

  Hide contents

7200 CL34-45-40-40 Maxed out sub-timings:

  Hide contents

 

 

So yeah, the RAM Disk software is definitely the bottleneck here, since with the exception of sequential reads the performance doesn't change much. Even then, the timings have a bigger affect on the sequential read performance than the frequency does for whatever reason, so going for an 8200 CL38 kit makes no sense in this application. 

Oi! I need to learn about memory tweaking then.

I tried it with the paid software, and much better.

I was running it was x32 earlier, and it was slower. But, if I can keep the 95GB of RAM and get these speeds, I can have larger RAM disks.

Edited February 19 by Avaviel

[brob]

1 hour ago, Avaviel said:

Thanks everyone!

Part of this is that I don't really understand what timings mean... previously the goal was always 'number big' get two sticks of 16GB DDR3 and 1600 and I'm golden. But that was the last time I upgraded parts, and it wasn't for work.

So this is stemming from this thread here, where I'm trying to find some Ram Disk software, and choose to go with it or Gen5 Raid 0. When I saw even the bad DDR5 Ram disk has 2000MB/s random read and writes I thought, "I'll get that!" and proceeded to get the wrong one.

I'm doing 3D CAD stuff for work. There are... odd slow downs that I can only attribute to tiny reads. I've replaced part by part trying to find what's making it slow. Tekla Structures is the software. I linked video in the older thread. I've been modeling with it extensively, and there are times when I go too fast for the software.

Much of that is from it being older software. But some of it is because it reads a database that is the CAD model over all, with tiny bits of info. Here's an example of a report... I'm convinced each time you select something, and I'll have... many items I sometimes need to select.

I just selected everything in the model. I don't often have to do this, and avoid it as much as I can. Still, it's like five minutes of flashing program, black screen to to regular trying to select 51882 objects. I'm used to seeing the "not responding, I'm thinking" screen. This one is frustrating, it's in a loop... I break the loop by left clicking, but also deselect all the things. (This gives me an idea though.)

  Reveal hidden contents

-------------------------------------------------------------------------------------------------
  Name                Profile               Material         Grids     Part           Assembly
                                                                       position       position
-------------------------------------------------------------------------------------------------
  COLUMN              HSS4X4X1/4            A992              3/L         C106           C106
-------------------------------------------------------------------------------------------------
  Total 15 Parts: 1834.53 lbf, 143'-9"
=================================================================================================
 
 Part                       GUID: 217801fb-ee4b-41e8-b612-c5991647d37d
 -------------------
 Global coordinates:
  Start point                   :  X= 20'-0"           Y= 115'-0"             Z= 0'
  End point                     :  X= 20'-0"           Y= 115'-0"             Z= 9'-7 5/8"
  Center of gravity             :  X= 19'-10"          Y= 115'-0"             Z= 4'-10 1/8"
  Top level                     : 9'-7"5/8
  Bottom level                  : 5/8
 -------------------
 Local coordinates, UCS:
  Start point                   :  X= 20'-0"           Y= 115'-0"             Z= 0'
  End point                     :  X= 20'-0"           Y= 115'-0"             Z= 9'-7 5/8"
  Top level                     : 9'-7"5/8
  Bottom level                  : 5/8
 -------------------
  Part position                 :  C106
  Assembly position             :  C106
  Net length                    :  9'-7"
  Gross length                  :  9'-7"
  Weight                        :  109.91 lbf
  Weight(Net)                   :  122.30 lbf
  Weight(Gross)                 :  109.91 lbf
  Volume                        :  431 cu.in
  Area                          :  1844.00 sq.inch
  Name                          :  COLUMN
  Material                      :  A992
  Finish                        :  RED OXIDE
  Profile                       :  HSS4X4X1/4
  Flange slope ratio            :  0
  Rounding radius 2 (r2)        :  0
  Rounding radius 1 (r1)        :  1/2"
  Plate thickness (t)           :  0
                                
  Width (b)                     :  4"
  Height (h)                    :  4"
  Class                         :  11
 -------------------

More:
Load bearing                  : No
LEVEL:                        : 1
Type for BOM:                 : HSS
Gage Material:                : HSS

Of course, part of the solution is finding how to treat the program. But, if I can 'select all objects' without freezing... I can EVERYTHING I need to. (And yes there are filters and stuff, so 90% of the time it isn't like that)

Here I have just a bit of an example of what I'm doing.
https://youtu.be/pUdca7e71oA

 

 

Are you aware that in a RAID array the slowest unit determines the array speed? Putting a PCIe 5.0 drive in an array with PCIe 3.0 drives is going to mean the array essentially runs at PCIe 3.0 speeds. Ditch the RAID 0 array. You may want a larger PCIe 5.0 drive.

 

Not sure how you went down the RAM disk path, but these days on current hardware it's an anachronism. Your software is 64-bit and can theoretically make use of more memory than you can provide. Have you gone through its settings to see if you can enlarge storage buffers and raise any memory usage limits?

 

I suspect the system drive and page file should be hosted on a PCIe 5.0 NVMe drive. 

 

If selecting ~60k objects takes 5 minutes, the issue is most likely with the software. Have you consulted the software's support group?

[Avaviel]

7 minutes ago, brob said:

 

If selecting ~60k objects takes 5 minutes, the issue is most likely with the software. Have you consulted the software's support group?

Very much! They... aren't helpful.

 

Quote

Are you aware that in a RAID array the slowest unit determines the array speed? Putting a PCIe 5.0 drive in an array with PCIe 3.0 drives is going to mean the array essentially runs at PCIe 3.0 speeds. Ditch the RAID 0 array. You may want a larger PCIe 5.0 drive.

Oh, I have a Gen5 NVME I'll install windows to later. I have two 1TB's.

 

Quote

Not sure how you went down the RAM disk path

Random reads and writes.

I'm trying to nail down every piece until the only thing left is the software.