Jump to content

3600MHz DDR4 ECC-UDIMMs?

After being stuck for years now with dual-core Haswell/Broadwell laptops, I finally want to upgrade to a decent Workstation (Software Development and occasionally some CPU+RAM intensive image processing tasks).
I've laid my eyes on Zen3, specifically a Ryzen 5900X.

For reliability and peace-of-mind reasons, I want to spend the extra money for ECC UDIMMs.
Just a small problem: that Processor would very much prefer 3600 MHz RAM. I failed searching for sold-as-3600-MHz ECC UDIMMs.

I hope to find some suggestions that are better than "take cheap ECC UDIMM sticks, transplant memory chips from binned 3600 MHz gaming non-ECC UDIMMs". I theoretically would be able to handle/organize the necessary reballing, but I'd much prefer to just OC some known-suitable factory-fresh ECC UDIMMs, or buy binned ones if there is a decent source.

I was hoping to get a pair of 32 GiB sticks at first, and buy a second pair in a few months if RAM ever got tight. The rest of the system should get watercooling, and I'm not picky for asthetics. I'd consider coating everything but the black top of the dies and the edge contacts that go into the socket with conformal coating, and just flow water around the sticks (with direct contact to where the heat spreader of gaming RAM would contact the memory chips).

Does a 5900X actually handle 64 or even 128 GiB RAM at 3600 MHz? I was hoping/wishing for CL16, but CL18 would be an option if the price difference signficantly surpasses the performance difference.

I know overclocking and maximum reliability don't really combine, but that just means I won't push the infinity fabric too hard and keep an eye out for ECC errors (in case I need to relax timings/frequency or increase voltage).
If 3600 MHz CL16 is too much for a 5900X if one cares about reliability, please let me know.

Link to comment
Share on other sites

Link to post
Share on other sites

There is no Ecc udimms of that speed, as any ddr4 stick currently over 3200 is overclocking, and ecc chips don't support things like xmp to overclock them.

 

You can try to overclock the chips and see if you can get high speeds, but there are no garantees. The nice thing about overclocking ecc ram is you know if there are errors.

 

 

THe fastest jdec speed for ddr4 now is 3200 mts cl20. But for your uses, ecc probably won't make a difference.

 

 

 

Link to comment
Share on other sites

Link to post
Share on other sites

As far as I know, no 3600 MHz ECC memory exists. It's not a JEDEC bin, so it wouldn't be considered stable for most enterprise applications. Even 3200 CL22 is just barely starting to hit the server space.

 

Honestly, if you're not running clients' webservers off your personal machine, it's probably not worth getting ECC. You're very very unlikely to see ECC-correctable errors interrupting your workflow on even a high-end workstation, and the combination of the cost increase and major speed & latency downgrade make non-ECC look a lot better.

 

Honestly, if you want to improve your reliability, ditch the watercooling. That will prevent more problems than ECC would ever solve.

Link to comment
Share on other sites

Link to post
Share on other sites

actully ryzens a better choice for cost but the i9 better for prefomance definately in a workstation a xeon may even do good depending on what your doin

 

Link to comment
Share on other sites

Link to post
Share on other sites

7 minutes ago, Meatball the man said:

i9 11900k with 64gb ram maybe a gtx 980 ti for good workstation graphics you dont need any high end gpu right

What...

No.

The 11700K and the 11900K are basically the same thing, except the 11900K costs more, can boost 100 MHz higher, and has some fancy boosting tech.

 

Also, the 5900X is 12 cores. The 11900K is 8 cores. In the rendering world, more cores = more better most of the time.

elephants

Link to comment
Share on other sites

Link to post
Share on other sites

3 minutes ago, FakeKGB said:

What...

No.

The 11700K and the 11900K are basically the same thing, except the 11900K costs more, can boost 100 MHz higher, and has some fancy boosting tech.

 

Also, the 5900X is 12 cores. The 11900K is 8 cores. In the rendering world, more cores = more better most of the time.

in the case they nedd more cores wouldent a server chip like xeon  or amd server chip be better  then again you gotta point

Link to comment
Share on other sites

Link to post
Share on other sites

1 hour ago, Electronics Wizardy said:

There is no Ecc udimms of that speed, as any ddr4 stick currently over 3200 is overclocking, and ecc chips don't support things like xmp to overclock them.

 

You can try to overclock the chips and see if you can get high speeds, but there are no garantees. The nice thing about overclocking ecc ram is you know if there are errors.

 

 

THe fastest jdec speed for ddr4 now is 3200 mts cl20. But for your uses, ecc probably won't make a difference.

 

 

 

Well, you don't need XMP to go beyond JEDEC. The reason for ECC is due to the scheduled reboot interval target, which I set at 6 months. Unless Kernel security fixes need to be applied, I'm not gonna reboot earlier than that.

And yes, seeing that you're pushing the RAM a little bit beyond comfort by way of getting much more than 1 single bit error per hour of memtest86+ is a very nice benefit of ECC.

The thing is: the only two differences between ECC UDIMMs and Gaming RAM are that the former has an extra chip for every 8 chips, and the latter has a heat spreader from the factory.
We're not talking registered RAM here. These are literally the same chips, just 12.5% more (which also means 12.5% higher load on the command and clock lines, but 3600 MHz shouldn't be a problem for Zen3, from what I understand).

Link to comment
Share on other sites

Link to post
Share on other sites

1 hour ago, FakeKGB said:

What...

No.

The 11700K and the 11900K are basically the same thing, except the 11900K costs more, can boost 100 MHz higher, and has some fancy boosting tech.

 

Also, the 5900X is 12 cores. The 11900K is 8 cores. In the rendering world, more cores = more better most of the time.

I'm not rendering, I'm compiling and running very nicely scaling applications for testing and measuring which parts of the program look like they'd benefit the most from some optimizations.
My workloads generally prefer cores, except some interactive parts like auto-complete in the code editor.

Link to comment
Share on other sites

Link to post
Share on other sites

1 hour ago, Meatball the man said:

in the case they nedd more cores wouldent a server chip like xeon  or amd server chip be better  then again you gotta point

I don't have the money to spend 2000 EUR on just motherboard and CPU.
Also there are no Zen3 EPYCs, yet.

Also server chips tend to be really bad in single-thread performance, and I do care about that as far as it affects basically loading times "when stepping through a door", to use the gaming equivalent to my situation.

I still have a GTX 970 around from my old i7-2600K system (broken motherboard, chipset problem with SATA ports; also fairly outdated and loud), as well as an SSD that can do at first. That leaves a budget of up to 1500 EUR for case+PSU+CPU+Motherboard+RAM.

Link to comment
Share on other sites

Link to post
Share on other sites

40 minutes ago, namibj said:

Well, you don't need XMP to go beyond JEDEC. The reason for ECC is due to the scheduled reboot interval target, which I set at 6 months. Unless Kernel security fixes need to be applied, I'm not gonna reboot earlier than that.

I have many systems without ecc that have gone much longer without reboots. Moderns systems should go years without a reboot just fine.

 

41 minutes ago, namibj said:

We're not talking registered RAM here. These are literally the same chips, just 12.5% more (which also means 12.5% higher load on the command and clock lines, but 3600 MHz shouldn't be a problem for Zen3, from what I understand).

Yes it should be possible to make faster udimms, but until ryzen, there was basically no reason/way to oc udimms. I hope someone brings this product to market, but its not out yet.

 

 

Link to comment
Share on other sites

Link to post
Share on other sites

1 minute ago, Electronics Wizardy said:

I have many systems without ecc that have gone much longer without reboots. Moderns systems should go years without a reboot just fine.

Yes it should be possible to make faster udimms, but until ryzen, there was basically no reason/way to oc udimms. I hope someone brings this product to market, but its not out yet.

 

 

Oh, sure, they do run longer than that without ECC. But the risk of bit rot in kernel data structures increases, and thus the risk of hidden/silent issues and corruption. And the 6 months is due to kernel updates, not due to stability.

Well, maybe some 3rd party is binning promising ECC UDIMMs? Remember, I'm trying to find a better way than to get 5 G.Skill sticks and 4 cheapo ECC UDIMMs, replacing the chips on the ECC UDIMMs with ones from the G.Skill sticks.
Needing to go that route would waste some memory (I can turn 9 Gaming DIMMs into 8 ECC DIMMs, but I have at most 4 slots so I'd waste half a stick), as well as cost me an expected 50-100 EUR extra as I don't have the reballing tools on hand. And I'd need to buy some cheap ECC UDIMMs to get the proper ECC-layout PCBs.

Link to comment
Share on other sites

Link to post
Share on other sites

7 minutes ago, namibj said:

Well, maybe some 3rd party is binning promising ECC UDIMMs? Remember, I'm trying to find a better way than to get 5 G.Skill sticks and 4 cheapo ECC UDIMMs, replacing the chips on the ECC UDIMMs with ones from the G.Skill sticks.
Needing to go that route would waste some memory (I can turn 9 Gaming DIMMs into 8 ECC DIMMs, but I have at most 4 slots so I'd waste half a stick), as well as cost me an expected 50-100 EUR extra as I don't have the reballing tools on hand. And I'd need to buy some cheap ECC UDIMMs to get the proper ECC-layout PCBs.

If you got the tools might s well try, but from what ive seen, its not easy or cheap, and  binning udimms will be much cheaper and easier.

 

Also the performance difference isn't too big. I relly wouldn't worry that much. Its almost always less than 5% jump in speed for good 3200 to 3600 on ryzen. So Id just get 3200 udimms here, and see what they can be oced to, and be happy with that.

Link to comment
Share on other sites

Link to post
Share on other sites

2 minutes ago, Electronics Wizardy said:

If you got the tools might s well try, but from what ive seen, its not easy or cheap, and  binning udimms will be much cheaper and easier.

 

Also the performance difference isn't too big. I relly wouldn't worry that much. Its almost always less than 5% jump in speed for good 3200 to 3600 on ryzen. So Id just get 3200 udimms here, and see what they can be oced to, and be happy with that.

In that spirit: anything specific to look for/look out for? Do they even normally handle those clocks with 64 GiB RAM?
And if so: what'd the performance impact be of populating all 4 slots with 32 GiB sticks (planned for later this year, if it turns out to be necessary)?

I'm asking here, because people stopped self-binning/trying non-factory-OC RAM since about Zen+/Zen2, and that's been a while. Are there maybe "undervolting-capable" ECC UDIMMs on the market? I'd assume those'd be binned a bit higher than normal ones.
And considering that I'm willing to run them at higher power (if adequate cooling can keep them alive for 3+ years), I'd expect a slight voltage increase would make even less-optimal chips run fast? As said, I can get the DIMMs watercooled without much fuss, so thermal issues due to higher voltages aren't the primary concern (overvolting is, though).

Link to comment
Share on other sites

Link to post
Share on other sites

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!

Register a new account

Sign in

Already have an account? Sign in here.

Sign In Now

×