G.Skill has a 256GB memory kit for Threadripper 3990X users

[HEXUS]

G.Skill Trident Z Neo DDR4-3600 CL16-20-20 (32GBx8) kit will become available from Q2.

Read more.

[Gentle Viking]

And here i thought i was a bad ass rocking 64GB

[LSG501]

Just what every 'workstation' user needs..... led lighting on their memory to interfere with colour accuracy etc....

I like the 32GB sticks, I like the 3600mhz, I like the CL16 but why is there no dual and quad packs according to their website.. there are going to be people out there who only need 2 or 4 of these sticks. Oh and I'm not a fan of the led lighting as I see no need for it on a workstation, let alone if the case doesn't have a glass side panel, which seems to be a rarity these days sadly.

[DanceswithUnix]

I see no mention of ECC ram, one of the basic needs for a workstation???

[LSG501]

I see no mention of ECC ram, one of the basic needs for a workstation???

It isn't necessary for every type of workstation, I'd see no real gain by using ecc for my 3D work for example.

[DanceswithUnix]

It isn't necessary for every type of workstation, I'd see no real gain by using ecc for my 3D work for example.

But if you care about the correctness of your work, then you should be using ECC ram.

It would seem odd to invest all that money in a machine to do that video render really fast, then risk haviing to do it again because there was a glitch in the middle.

I think modern ram is actually better than the old DDR and DDR2 ram and I see surprisingly few errors, but having worked on corporate server design with that sort of memory quantity I know they happen. Who the heck wants 256GB but doesn't care about the integrity? Same person that wants RGB I presume.

This PC has 16GB of standard DDR4 and it has honestly bugged the hell out of me since I built it. I don't *know* of any problems, it has had zero crashes under Linux where I care about such things, but I don't *know* for sure. As production releases are all compiled on a 2600X based server that does have ECC ram it won't ever impact customers, else I couldn't have taken the punt.

The kids even have ECC ram in their PCs, being hand-me-down workstations, so at least their homework is safe

Usually ECC ram is a fairly small uptick in cost vs non ECC, but sadly when I built this there seemed to be a real scarcity and I couldn't justify the cost. At some point I will get some proper grown up ram, and this junk can go in a games machine.


Edit: I notice the Threadripper max supported memory speed is 3200, so these sticks are an overclock. I don't think the official specs for ECC ram go above a 2666 rating, so the best you can get for a proper workstation is something like this:

https://www.scan.co.uk/products/32gb...-dual-rank-12v

which at £167 per stick is going to cost you a bargain £1336 to fill your slots. OFC you could try and overclock your ECC ram, and unlike the cheap stuff you can look in your /var/log/messages file to see if the kernel is reporting memory errors so you actually know if the overclock is working. It surprises me that overclockers don't buy ECC.

[LSG501]

But if you care about the correctness of your work, then you should be using ECC ram.

It would seem odd to invest all that money in a machine to do that video render really fast, then risk haviing to do it again because there was a glitch in the middle.

There have been a LOT of tests done on this, at least within the 3D community, and the consensus is pretty much that the benefit of using ecc memory is so slim it barely even registers in the tests, if at all, so like I said there's no real benefit to me using ecc. IIRC I'm pretty sure I'd actually take a performance hit by using existing ECC ram because the ram is 2666mhz too.

Now if you're doing something like scientific tests where the formulas etc can take days to calculate or doing something where an error could cost someone their life then that is an area is can be of benefit.

I think modern ram is actually better than the old DDR and DDR2 ram and I see surprisingly few errors, but having worked on corporate server design with that sort of memory quantity I know they happen. Who the heck wants 256GB but doesn't care about the integrity? Same person that wants RGB I presume.

Servers aren't really the same as a workstation though, I doubt many would say ecc doesn't have a place in the right environment, servers being one of them.

[DanceswithUnix]

Servers aren't really the same as a workstation though, I doubt many would say ecc doesn't have a place in the right environment, servers being one of them.

I'm not saying they are the same, I'm saying that if you have quarter of a terabyte of ram you had better be expecting the odd bit flip in that ram. If your work is somehow resistant to those errors then that's super, but most workstation tasks just aren't. I expect two things from a workstation: heavy compute ability (including decent single thread performance) and basic fault tolerance. Otherwise, it's just a quick PC.

[LSG501]

I'm not saying they are the same, I'm saying that if you have quarter of a terabyte of ram you had better be expecting the odd bit flip in that ram. If your work is somehow resistant to those errors then that's super, but most workstation tasks just aren't. I expect two things from a workstation: heavy compute ability (including decent single thread performance) and basic fault tolerance. Otherwise, it's just a quick PC.

This blog post, while a little old, basically summarises things for 3D CAD as opinions haven't really changed.... basically to sum it up, you're more likely for the program to crash due to poor coding than have a memory issue cause you an issue. And in all honesty, I'd say that is pretty accurate lol.

[DanceswithUnix]

This blog post, while a little old, basically summarises things for 3D CAD as opinions haven't really changed.... basically to sum it up, you're more likely for the program to crash due to poor coding than have a memory issue cause you an issue. And in all honesty, I'd say that is pretty accurate lol.

That's a very dated article. It was even dated when it was written as rowhammer was already announced by then which used an improved understanding of dram problems as an attack vector (with more research since to demonstrate that a skilled attacker can arrange a three bit flip to get past ECC protection). I would hope (but don't know) that newer chips are designed to be immune to these problems, just like many years ago they stopped using slightly radioactive substrates for making memory chips when they realised some chips were their own source of errors.

But the big issue I have with that article is the assumption that memory errors will cause a crash. That is far from the worst case scenario, the problem is if it silently corrupts your design. Now what if you save it out, flaw included, and don't notice for weeks until it bites. The premium to not have to worry about that is for an average 16GB machine something like £40 and a speed penalty that I can measure but not feel.

The issue with not noticing memory errors in the machine's system logs is a Microsoft issue, really they need to get beaten up over that. That is probably a chicken and egg problem though, if more people used ECC then MS might be inclined to fix it.

[LSG501]

That's a very dated article. It was even dated when it was written as rowhammer was already announced by then which used an improved understanding of dram problems as an attack vector (with more research since to demonstrate that a skilled attacker can arrange a three bit flip to get past ECC protection). I would hope (but don't know) that newer chips are designed to be immune to these problems, just like many years ago they stopped using slightly radioactive substrates for making memory chips when they realised some chips were their own source of errors.

But the big issue I have with that article is the assumption that memory errors will cause a crash. That is far from the worst case scenario, the problem is if it silently corrupts your design. Now what if you save it out, flaw included, and don't notice for weeks until it bites. The premium to not have to worry about that is for an average 16GB machine something like £40 and a speed penalty that I can measure but not feel.

The issue with not noticing memory errors in the machine's system logs is a Microsoft issue, really they need to get beaten up over that. That is probably a chicken and egg problem though, if more people used ECC then MS might be inclined to fix it.

I did say it was dated, but there have been numerous tests done since then by others and they all basically come out the same. Also I don't remember it saying the memory errors would cause the program to crash, it does say it's somewhere between nothing and a full on crash which imo is a fair comment (I've actually had ecc hard error and crash, but funnily enough non ecc hasn't...).

Honestly having worked with both ecc and non ecc machines I honestly don't see any benefit from ecc in my 3D CAD work.

[devBunny]

@Mark: If you are going to push the boast out to purchase a 128-thread AMD Ryzen Threadripper 3990X

Ha ha. Nice typo, or even better if it was wordsmithed.