Most Reasonable RAM Combination to Work?

[ScottishTapWater]

Right, so I've got this system currently:

CPU: 9950x3d

RAM: Kingston FURY 32GB (2x16GB) 6000MT/s

GPU: AMD 7900 XTX

Mobo: ASUS PRIME X670E-PRO WIFI

 

I'm now involved pretty heavily in developing products that utilise some pretty complex and large ML models (none of the horrible destroying whole industries kind of work don't worry). So, I don't really have enough memory to run these locally and need to upgrade the amount of RAM I have quite substantially.

 

What's my best bets with regards to how much memory is likely to work and what sticks can people recommend that are still fast enough to not completely bottleneck me when I am gaming?

[strange13930]

9 minutes ago, ScottishTapWater said:

Right, so I've got this system currently:

CPU: 9950x3d

RAM: Kingston FURY 32GB (2x16GB) 6000MT/s

GPU: AMD 7900 XTX

Mobo: ASUS PRIME X670E-PRO WIFI

 

I'm now involved pretty heavily in developing products that utilise some pretty complex and large ML models (none of the horrible destroying whole industries kind of work don't worry). So, I don't really have enough memory to run these locally and need to upgrade the amount of RAM I have quite substantially.

 

What's my best bets with regards to how much memory is likely to work and what sticks can people recommend that are still fast enough to not completely bottleneck me when I am gaming?

two of these kits would max out the ram you can have on your motherboard and they are the same speed as your current ram. If possible you could see if who you work for can fund the upgrade because it is costly. https://pcpartpicker.com/product/xQcgXL/corsair-vengeance-96-gb-2-x-48-gb-ddr5-6000-cl30-memory-cmk96gx5m2b6000c30

[porina]

I'm not familiar with how ML work scales, especially if they're bigger than will fit in the GPU. Does more system ram give you a workable system?

 

For big ram, G.Skill just announced 4x64GB kit at 6000C32. https://videocardz.com/press-release/g-skill-reveals-worlds-first-large-capacity-256gb-64gbx4-ddr5-u-dimm-memory-at-ddr5-6000-cl32

 

Depending on your capacity needs, maybe a different solution is more suitable. Nvidia DGX Spark (formerly DIGITS) is looking most promising with 128GB of unified memory combined with fast AI cores. AMD Strix Halo could be cheaper but slower, and there's also Apple.

[CatTNT]

Everyone knows for ryzen gaming, the best is 6000 MHz and lowest CL possible, usually CL30, and some newer ones with Cl28. To maintain stability, try to get as much RAM as possible in 2 sticks, if that isn't possible, then 4 sticks of DDr5-6000 Cl30 might work. My only experience with this is that my brother has been using 4x16 (using two 2x16 kits with identical specs) DDR5-6000 Cl36 on his 9950X for months with no issues. I don't know if adding more capacity with the same speed would introduce instability, but probably don't increase capacity and decrease your CL at the same time. Don't go above 6000 MHz, no point. 

[ewitte]

16 minutes ago, CatTNT said:

Everyone knows for ryzen gaming, the best is 6000 MHz and lowest CL possible, usually CL30, and some newer ones with Cl28. To maintain stability, try to get as much RAM as possible in 2 sticks, if that isn't possible, then 4 sticks of DDr5-6000 Cl30 might work. My only experience with this is that my brother has been using 4x16 (using two 2x16 kits with identical specs) DDR5-6000 Cl36 on his 9950X for months with no issues. I don't know if adding more capacity with the same speed would introduce instability, but probably don't increase capacity and decrease your CL at the same time. Don't go above 6000 MHz, no point. 

Decent chance of not hitting 6000 with 4 chips.  Although, 9000 series has the best chance of doing so for the AMD side.  I'd stick with 2x48GB if it is enough, or 2x32GB if 64GB is fine.

[ScottishTapWater]

37 minutes ago, porina said:

I'm not familiar with how ML work scales, especially if they're bigger than will fit in the GPU. Does more system ram give you a workable system?

 

For big ram, G.Skill just announced 4x64GB kit at 6000C32. https://videocardz.com/press-release/g-skill-reveals-worlds-first-large-capacity-256gb-64gbx4-ddr5-u-dimm-memory-at-ddr5-6000-cl328x22b

 

Depending on your capacity needs, maybe a different solution is more suitable. Nvidia DGX Spark (formerly DIGITS) is looking most promising with 128GB of unified memory combined with fast AI cores. AMD Strix Halo could be cheaper but slower, and there's also Apple.

Yeah, unfortunately, I can't really justify dropping that kind of money to get the dedicated AI hardware.  That GSkill kit looks great, but god knows 1) when it'll be available and 2) how much it'll cost

[ScottishTapWater]

41 minutes ago, strange13930 said:

two of these kits would max out the ram you can have on your motherboard and they are the same speed as your current ram. If possible you could see if who you work for can fund the upgrade because it is costly. https://pcpartpicker.com/product/xQcgXL/corsair-vengeance-96-gb-2-x-48-gb-ddr5-6000-cl30-memory-cmk96gx5m2b6000c30

I can probably manage that. Unfortunately, my employer won't pay for parts to go in my own rig but also won't fork out for an entire system so looks like I'm stuck on this one, I can probably manage the cost though thanks!
 

What's the craic with people saying running 4 sticks is pretty unstable though?

[strange13930]

2 minutes ago, ScottishTapWater said:

I can probably manage that. Unfortunately, my employer won't pay for parts to go in my own rig but also won't fork out for an entire system so looks like I'm stuck on this one, I can probably manage the cost though thanks!
 

What's the craic with people saying running 4 sticks is pretty unstable though?

It puts more strain on your cpus memory controller. start by getting only a set of two and if your maxing it out get another to use all four ram slots.

[ScottishTapWater]

Just now, strange13930 said:

It puts more strain on your cpus memory controller. start by getting only a set of two and if your maxing it out get another to use all four ram slots.

You'd have hoped if it's in-spec then the memory controller should be able to cope but sounds like a plan. I can probably just about get away with running the current model I'm working with on 80GB.

[ScottishTapWater]

47 minutes ago, porina said:

I'm not familiar with how ML work scales, especially if they're bigger than will fit in the GPU. Does more system ram give you a workable system?

Also, just to answer this bit. It's slower, but if your VRAM gets filled, you can offload to system RAM and still run the models. A bit slower is fine when I'm developing.

[porina]

14 minutes ago, ScottishTapWater said:

You'd have hoped if it's in-spec then the memory controller should be able to cope but sounds like a plan. I can probably just about get away with running the current model I'm working with on 80GB.

https://www.amd.com/en/products/processors/desktops/ryzen/9000-series/amd-ryzen-9-9950x3d.html

 

This is AMD's spec for the 9950X3D. They guarantee you can run 4 sticks at JEDEC 3600. Anything above that is not guaranteed by AMD. Mobo/ram makers might make claims above AMD, but that's not AMD's problem any more.

 

13 minutes ago, ScottishTapWater said:

Also, just to answer this bit. It's slower, but if your VRAM gets filled, you can offload to system RAM and still run the models. A bit slower is fine when I'm developing.

7900 XTX is connected to rest of world by PCIe 4.0 x16, for about 32 GB/s bandwidth. Compare that to 960 GB/s VRAM. Even dual channel 6000 ram is about 94 GB/s. PCIe speed will limit you before ram speed. It might be more than "a bit slower".

[ScottishTapWater]

1 minute ago, porina said:

https://www.amd.com/en/products/processors/desktops/ryzen/9000-series/amd-ryzen-9-9950x3d.html

 

This is AMD's spec for the 9950X3D. They guarantee you can run 4 sticks at JEDEC 3600. Anything above that is not guaranteed by AMD. Mobo/ram makers might make claims above AMD, but that's not AMD's problem any more.

 

7900 XTX is connected to rest of world by PCIe 4.0 x16, for about 32 GB/s bandwidth. Compare that to 960 GB/s VRAM. Even dual channel 6000 ram is about 94 GB/s. It might be more than "a bit".

Huh, fair. I suppose XMP is taking the piss a bit now you mention it.

> 7900 XTX is connected to rest of world by PCIe 4.0 x16, for about 32 GB/s bandwidth. Compare that to 960 GB/s VRAM. Even dual channel 6000 ram is about 94 GB/s. It might be more than "a bit".

Even 10x slower is probably fine for what I'm testing. It's still a huge improvement over not really being able to debug locally.