Hacker News Reader

Comment by ChuckMcM | Hacker News Reader

This illustrates one of the truism's in semi-conductors which is that you need to hold an advantage for 18 months before you have enough momentum to actually move the needle. If I were AMD's marketing team I would start hammering on ECC memory in the desktop and on laptops. Anything with 16GB or more of main memory. Intel has held that as a Xeon differentiator for a long time and it adds something like 20% margin to the part. That AMD can push PCIe 4.0 and ECC at 'desktop' prices is pretty neat.

View on HN

Outside of our bubble, people don't care about ECC memory on their laptop or desktop. It also adds ~12% (iirc) more memory cells, meaning higher memory cost, which is going to be a hard sell given that it's a solution to a problem that most customers don't even know about, nor would most care much if it was explained to them. From a user perspective, memory is fine as it is, and if memory corruption does occur, it's probably blamed on buggy software. That would make any such push look like a desperate "we need a differentiator to tell people why they should buy our CPUs and upsell them in the process" move. Competing on price, speed, and (for laptops) power consumption would appear much more promising.

It needs to be marketed as a reliability improvement it is. Lots of people simply don't know about it at all, but would love better stability.

This is mass market retail we’re talking about. Unless it’s easily demonstrable, people are going to buy the prettier box on the shelf or the one that’s cheaper.

I’d bet money that when company X adds ECC and calls it “super stable memory”, company Y adds a button in the settings menu that calls fsck and calls that “hyper stable memory” and undercuts company X by 12%.

Actually, you can have ECC without an extra chip, and 4 gen DDRs actually have quote low single bit error rates because they have error correcting PHYs

People don't care about a lot of things until they're told it matters. I mean, look at the "gaming" and "X with RGB" market and tell me people name economically reasonable choices when buying computer hardware. I'm sure marketing could sell quite a few ECC sticks.

I agree with your view. Just want to add that from recent intel and AMD server purchases, the price difference in processors though considerable, is shadowed by the cost of memory sticks... for a fully loaded server, the cost savings from going to AMD appears like a small number next to the cost of mem sticks.

PCIe 4.0, more mem channels, and faster memory (3200) does make things very attractive.

Iirc, all AMD CPUs support ECC- but motherboard support is flaky.

That is correct, it requires that the BIOS set up the appropriate machine check vectors and initialize RAM during POST. My current desktop with 32GB gets 2 - 3 single bit errors a month. (it has a Xeon, will be replacing it with a Ryzen 3000 series relatively soon with 128GB of RAM)

ECC UDIMM RAM is way too expensive comparing to LRDIMM and you can barely get 32GB ECC UDIMM modules. Not going to happen...

You under estimate my willingness to overspend on things :-).

Sounds like Epyc to me.

7302P (16c/32T), 7402P (24c/48t), 7502p (32c/64t), or 7702P (64c/128T).

You also get 4 times the memory bandwidth and motherboards guaranteed to work with ECC.

Of you could go half way, threadripper comes out tomorrow or so and has twice the memory bandwidth of Ryzen/half of Epyc in the TRX40 parts. There's also a TRX80 which is using the same socket/memory bandwidth as Epyc.

Yeah, I was trying to figure out how to cool a Supermicro dual socket board in my desktop case. Its a challenge, they really want the big datacenter type rack fans blowing through the system and that is a bit too noisy for me. On the plus side it could have way more RAM. :-)

Indeed, for noise and cost reasons I'd just got with a single socket. Any decent case with 2x140mm in front and 1x140m in back should easily handle a single socket quietly.

Dual 7702P with 4TB ECC RAM and 4x RTX 8000 would be a dream ;-) At the price of Corvette...

And there it is; ASUS Prime TRX40 Pro, 128GB ECC memory, and a 3950x CPU.

I've lost track of all the different kinds of RAM going these days. Does anyone have a good link to a primer of what's current?

That's because there is no demand. If AMD were to move towards ECC the memory demand for those ECC DIMMs would increase and price would go down as more are produced.

Sadly 128GB is easy without ECC, but hard with ECC. There are 16GB ubuffered ECC dimms around for $100 or so, but no 32GB ECC I could find.

So 64GB ECC is easy, 128GB ECC not ... unless you upgrade to one of the "P" Epycs like the 7302P or 7402P. That gets you 16 cores/32 threads (ryzen 3900 x = 12 cores/24 threads) but also 4 times the memory bandwidth.

Threadripper is an option there also. The main issue I've run into is that I can't find anything faster than 2133 in ecc

M391A4G43MB1-CTD

Yes. Most (all?) Asrocks support it. Some ASUSs do. Most (all?) Gigabytes don't.

EOY and beginning of year is when most purchases for servers come in too. I suspect we will see AMD will spike in the next six months.

I know I’m also waiting to buy a laptop until the 7nm mobile chips are out. That’ll be a massive battery life increase (A couple hours), to the point that people will flock to AMD.

Chances are, whatever laptop gets made with that new chip will just have a thinner battery and the same battery life unfortunately :(

I think AMD's marketing team is more based on partnerships ( and custom solutions), than on big ads.

They should push more for partnerships with OEM's though.

> ECC memory in the desktop and on laptops

Wow and here I though RGB lighting was pinnacle of pointlessness.

Non-ECC RAM bit flips are pretty common at modern densities. They can cause everything from a stray pixel on screen, to crashes, to corrupted files and file systems.

At $dayjob we only have like 200 laptops but our help desk staff find 1-3 bad non-ECC DIMMs a year causing crashes or borked files.

Well, you said it... 3 crashes a year vs cost of 200 ECC RAM. (say refresh of pc every 3 years). So in lifetime of those machines you will experience 9 crashes. Is it really cost effective? Also file versioning, automatic backups of files should be a standard in any serious business.

So in reality no work would be lost to crash, only bit of time.

The problem is that backups don't work if the file is already corrupted when it's being written, most filesystems aren't very robust to on-disk corruption (and even those that are like ZFS have quite a few fun bugs if there is memory corruption present). Even versioning will not save you if you version already corrupted data.

I am programing over 10 years, that never happened to me and I never heard of anyone who had experienced something like that. Your words/excels save a copy of files upon opening (to handle crashes or corruption).

Is it worth entertaining this scenario? Unless you are running on a space probe or nuclear facility pc that's non-issue worth any time.

Oh, I've experienced such things, it's a matter of scale. If you have 1000 machines running, the probability of one eating importing project data due to cosmic bitflips starts to matter.

Processors use ECC in their on-board caches for a reason.

Transient memory errors are not uncommon when you have many terabytes of collective RAM in use.

Google’s study reports 8% of DIMMS see a transient error per year. https://ai.google/research/pubs/pub35162

> find 1-3 bad non-ECC DIMMs

The problem of faulty memory is separate and can be addressed by memory testing (which can even be run while computers are online).

My understanding is that ECC's main purpose is to guard against corruption of non-faulty memory, due to ionising radiation for instance. It's rare for this to be a problem in normal scenarios (not in space or high-altitude), and use cases.

Reply on news.ycombinator.com