The ECC cost with different sockets/platforms

, ,


Just a quick brain fart.

I don’t know if it’s just me, but I’ve noticed a trend (may be):

  • If a CPU is cheap, the Unbuffered ECC RAM it needs is expensive.
  • If a CPU is expensive, the Registered ECC RAM is cheap.

So you don’t really win either way! :slightly_frowning_face:

Do I have it arse about face?

You’re right about that.
Although I think it’s mostly about availability.
RDIMMs are often easy to find really cheap because they get pulled out of old servers all the time, UDIMMs on the other hand are used less and in lower quantities per machine.

1 Like

Ah cheers felix.

It only crossed my mind when I was looking for Reg and they were cheap (16GB less than 40 of my currency), while Unbuffered was 100 or more. :frowning:


It’s really unfortunate.

Buying used Servers, I’ve often noticed that the higher end platforms will often be cheaper on the used market due to sheer availability, e.g. LGA 2011-3 vs 1151.

So often times you have to calculate advantage in power consumption vs disadvantage in cost of RAM.

1 Like

I’ve noticed this for the last decade at least.

yep, assuming you can actually find them! - When I was looking to build a new system last year, I could not find any unbuffered ECC DDR4-3200 DIMMs at all that would be compatible with the Threadripper 3xxx processors/mobos I was looking at. I went for an EPYC Milan with registered PC4-3200 DIMMs instead, which bizarrely was far easier to find suppliers for (even considering the Milan was only just released).


Even if bought new, I imagine the market for U-DIMMs is much smaller, since an advantage of R-DIMMs is the ability of a CPU to handle more of them.

My impression overall has been that R-DIMMs are the standard on servers, and U-DIMM support for ECC is almost a back-ported feature. Even Apple’s Mac Pro tower uses either R-DIMMs or LR-DIMMs.

Admittedly, I have no server hardware of my own and my interest lies particularly in OpenPOWER, where POWER9 only supports R-DIMMs or a separate memory buffer system (Centaur for POWER8/POWER9, OMI for POWER10) where U-DIMM vs R-DIMM is not exposed to the user.

On POWER9, the Centaur/R-DIMM distinction exists between different chip designs, but is this the case for Intel or AMD?
For AMD, since Zen 2 the IO is on an entirely separate die, so the actual CPU-core dies only see an infinity fabric connection, not DDR4; in this case the question would be whether EPYC and Ryzen must use different IO dies to handle the U-DIMM vs R-DIMM difference.


I started writing about R-DIMMs and ECC in a different thread about HEDT; my posts there have a fair amount of overlap with the discussion here I think.

1 Like

I just looked online today (because of this), and found 3200 ECC UDIMMs even at my favorite consumer etailer. That said, the price is… painful to say the least. Roughly two times non-ECC.