[stevefan1999]

I don't. In fact we shouldn't have. ECC RAM is supposed to be serving services that requires extremely high RAS (Reliability, Availability and Serviceability). Any normal DDR4 RAM out here is already having a good enough RAS for a relatively long time.

As an anecdotal reference, my home "NAS" few weeks ago is running only with 16GB of cheap ADATA DDR4 2666MHz RAM ran for half a year non-stop, and I didn't even see a single RAS event in dmesg. It could be survivorship bias though.

---

For a typical home/personal workstations we normally run stuffs that is burst period (i.e. used during 9 to 6 and some high usage in 2-3 hours). It is also okay to shrug a few bitrots or even tolerate a system crash like if the butterfly effect hits.

So another reason I don't use ECC RAM on workstations: Remember that ECC RAM is notoriously long to do memory initialization and memory training.

Not sure if its an AMD thing but my friend booted his EPYC server with 256GB of DDR4 ECC RAM for half an hour.

You are using a "work station" so getting the work done as quickly as possible is more important than having a system too safe and reliable. Speed and work throughput is of utmost importance so ECC I'd say is okay to be sacrificed.

[theevilsharpie]

> So another reason I don't use ECC RAM on workstations: Remember that ECC RAM is notoriously long to do memory initialization and memory training. > > Not sure if its an AMD thing but my friend booted his EPYC server with 256GB of DDR4 ECC RAM for half an hour.

This server likely had a setting enabled to perform a full test of the memory on boot.

I'm not sure why server vendors do this (I've never once in my career seen that test catch anything), but the boot times have nothing to do with ECC. My Ryzen 9 3900X-powered PC with 64 GB of ECC boots in a few seconds.

> ECC RAM is supposed to be serving services that requires extremely high RAS (Reliability, Availability and Serviceability). Any normal DDR4 RAM out here is already having a good enough RAS for a relatively long time.

Without the ability to detect errors, how do you know it's good enough?

[bcrl]

Server BIOS is just very poorly written. It never gets the boot time optimizations that desktops get.

At a former employer, every time we got a new generation of Intel dual socket server boards, the bugs that we had to get fixed in the previous generation would always reappear. Specifically, there were issues with the board properly detecting and configuring PCIe gen 1 boards. One of the custom boards used an older FPGA and only used gen 1 speeds. Intel would eventually get the BIOS fixed about 6 months after the first test systems would arrive. I saw this happen in no fewer than 3 distinct generations of server boards. It's bonkers how broken their processes are. My best guess is that the code for the next generation of CPU always branches off before the bug fixes are reapplied.

Sometimes I hate computers.

[thulle]

> Not sure if its an AMD thing but my friend booted his EPYC server with 256GB of DDR4 ECC RAM for half an hour.

That could be a memory training issue. Using 256GB listed as compatible with the Supermicro motherboard my EPYC server boots in a minute or so. If i configure my workstation memory with weird timings it can take a while for the motherboard to figure out the remaining Auto-timings, possibly adding 4-5 minutes to the boot procedure.

> Speed and work throughput is of utmost importance so ECC I'd say is okay to be sacrificed.

But you don't have to, see my other post in the thread regarding my workstation memory.