Who is the customer that's simultaneously willing to spend $2,000 on a CPU and so desperate to save a few bucks that they'd settle for non-ECC RAM? I want to meet that person.
Suppose I'm doing scientific computing, is ECC solving a super rare problem, or something that might happen weekly or monthly?
Thus ECC would be super helpful to those that expect that most of the 64GB will be filled with bits that shouldn't be changed.
This article by a system integrator is a pretty decent data point: https://www.pugetsystems.com/labs/articles/Advantages-of-ECC...
But this would be if you were using 64GB nonstop for a month, which is quite a bit rarer than just having 64GB. And certain filesystems (ZFS, Btrfs) and most databases have software checksumming functionality to detect bitrot.
Not to say ECC isn't useful or multiple layers of protection aren't prudent, just that the problem it solves is something you're unlikely to encounter below datacenter scale.
Google did a big study in 2009 [0] and found error rates between 2.4E-11 and 6.7E-11 errors/bit-hour, or, in more familiar units: about 1 error per 2-GByte of memory per hour.
Obviously most errors go unnoticed, but for anything where random bit flips are unacceptable, ECC is necessary.
Note too that error rates go up with memory density, not down, so things might be worse with more recent chips.
[0] http://www.cs.toronto.edu/~bianca/papers/sigmetrics09.pdf
Wow! So that means a 16GB laptop will on average experience 8 memory errors per hour under moderate load? I had no idea it was that high.
This likely explains why my Mac crashes randomly but incredibly rarely. It's like once every few months. I guess a bit flip happens at the wrong place about that often?
Per lane (unidirectional) bandwidth by PCIe revision [0]:
PCIe 1.0: 250MB/s
PCIe 2.0: 500MB/s
PCIe 3.0: 985MB/s
PCIe 4.0: 1.969GB/s
PCIe 5.0: 3.9GB/s
> A 10gb card does not need 8 PCIe lanes.
It depends on the PCIe revision(s) the card supports. If the card needs to support 10GBit on a PCIe 1.0 link, then it will need an x8. The bandwidth of PCIe 1.0 is ~250MB/s per x1 link, but there is protocol overhead and 4 lanes doesn't provide enough bandwidth, so you'd have to use an x8 (since x6 keyed cards are not a thing).
The most popular cards with enthusiasts are refurb Mellanox cards, which are PCIe 2.0 but support PCIe 1.1, so they need to utilize PCIe x8 to provide full speed 10GBit:
PCI Express 2.0 x8 (1.1 compatible) [1]
[0] https://en.wikipedia.org/wiki/PCI-Express
[1] http://www.mellanox.com/page/products_dyn?product_family=20&...
No, it's not. Not when the cards people are using are targeted at PCIe 2.0/1.1.
Since the context was 10GBit cards, most people are using the Mellanox card I linked to, which is PCIe 2.0.
Whine about the age of the standard all you want, what's on the market today and in use by a majority of home users is PCIe x8 for 10GBit.
The original post was saying there is a shortage of PCI lanes because of situations like a 10gb network card using 8 lanes. The truth of the matter is that if someone is running out of lanes in that situation, it is because of exceptionally inefficient use of the lanes, not because 40 lanes is as limiting as what is being implied.
No, it's because the customer is installing the cards they want to use, and those cards are consuming PCIe lanes.
A system with more PCIe lanes lets the customer install more of whatever they want. The PCIe revision doesn't matter, and neither does the "efficiency" of the cards they're installing. More lanes means the customer can install more of whatever they want/need to install.
Sure, you could spend more on a 10GBit card that supports a newer PCIe revision and doesn't use 8 lanes, or just go and buy an AMD system which costs less than Intel and provides 64 PCIe lanes anyway.
You can recommend whatever you want, but at the end of the day the client is going to use what they want to use. I find it's best just to accept that and not argue with the client about how they're doing things wrong. You can of course make a recommendation on how to improve things, but if the client doesn't accept it, move on. You're usually hired for your professional opinion, not to stand around whining about how no one implements your great ideas.
If the client decides not to take your professional opinion, then either:
1) accept that they're paying you for your opinion which they're not following (which is their choice, it's their money)
2) quit and find another client who treats you like the demigod you so clearly are
My point about AMD providing more PCIe lanes than Intel at a lower price point still stands. The original point of OP was that Intel is not providing more than 40 PCIe lanes on anything but Xeon SKUs:
> That and refusing to up the number of directly connected PCI lanes. The latest chips support 40 lanes but some of them are outboard of the CPU and connected via a controller.
I wish I'd known about AMDs chips at the time though. I suspect my next machine will be an AMD, which hasn't been true in a decade.
Those who care are supposed to spend even more on Xeons.
These are odd moves to make during the long tail of Moore's law. Not sure if protecting the server market from a small amount of overly cheap buyers putting workstations into server use is worth hobbling everyone else and forcing them to pay the Xeon and Naples premium.
No idea who is spending $2k on this either, especially when the 1800X is $449 right now. Whale gamers I suppose.
How important is ECC RAM if I want 32 GB or more RAM? What are the consequences of not having ECC RAM?
I'll probably go for AMD this time anyway, but was wondering in general how important that is. Are non-ECC machines known to crash? Wouldn't that be a factory design flaw that would need to be fixed, at least according to EU consumer protection laws?
Most memory is non-program data, however, and most non-program data won't cause obvious errors if corrupted. A picture that renders incorrectly, a broken model in a game, a single burst of static in a video...
It's insidious precisely because it's hard to notice, and corruption never gets better over time.