Despite Initial Claims, AMD Confirms Ryzen 8000G APUs Don't Support ECC RAM (tomshardware.com) 64
Slashdot reader ffkom shared this report from Tom's Hardware:
When AMD formally introduced its Ryzen 8000G-series accelerated processing units for desktops in early January, the company mentioned that they supported ECC memory capability. Since then, the company has quietly removed mention of the technology from its website, as noted by Reddit users.
We asked AMD to clarify the situation and were told that the company has indeed removed mentions of ECC technology from the specifications of its Ryzen 3 8300G, Ryzen 5 8500G, Ryzen 5 8600G, and Ryzen 5 8700G. The technology also cannot be enabled on motherboards, so it looks like these processors indeed do not support ECC technology at all.
While it would be nice to have ECC support on AMD's latest consumer Ryzen 8000G APUs, this is a technology typically reserved for AMD's Ryzen Pro processors.
We asked AMD to clarify the situation and were told that the company has indeed removed mentions of ECC technology from the specifications of its Ryzen 3 8300G, Ryzen 5 8500G, Ryzen 5 8600G, and Ryzen 5 8700G. The technology also cannot be enabled on motherboards, so it looks like these processors indeed do not support ECC technology at all.
While it would be nice to have ECC support on AMD's latest consumer Ryzen 8000G APUs, this is a technology typically reserved for AMD's Ryzen Pro processors.
I can't make much of it. (Score:2)
I can't make much of it excepted that maybe they want to force customers to buy the pro version. ECC memory isn't rocket science after all.
Re: (Score:2)
I can't make much of it excepted that maybe they want to force customers to buy the pro version. ECC memory isn't rocket science after all.
Screwing consumers for more money. Paying more and getting less. What part of the re-definition of “progress” is unclear to you?
Old people used to say they don’t make ‘em like they used to. Now everyone says it. Because it’s true.
Re: (Score:1)
BS.
I would argue they would force people to pay more and get less if they actually introduced ECC support in those CPUs.
Most people in the target group wouldn't need it, therefore they would pay more for a feature that's useless to them. And if some of the others wanted it, they would have had to also get ECC RAM, at which point they would be better off getting a Pro CPU.
Re:I can't make much of it. (Score:5, Insightful)
BS. I would argue they would force people to pay more and get less if they actually introduced ECC support in those CPUs. Most people in the target group wouldn't need it, therefore they would pay more for a feature that's useless to them. And if some of the others wanted it, they would have had to also get ECC RAM, at which point they would be better off getting a Pro CPU.
I would argue that everyone needs ECC RAM. Sure, a random bit flip causing a crash or hard drive corruption every so often might not be the end of the world, but it still sucks. ECC should be the default, not the exception. The tiny extra cost in the overall price of computers caused by eliminating non-ECC RAM across the board would be well worth it in terms of overall stability.
Re:I can't make much of it. (Score:5, Insightful)
And just to add to that, I would note that nearly every time I've helped someone with a computer that was crashing constantly, bad RAM was at fault. ECC would have made it possible for the OS to detect those problems and alert the user that their RAM was defective. Instead, some of them had suffered for years, thinking that their computer's behavior was normal.
That's why ECC should really be an "everyone" thing, not a "server" thing.
Re: (Score:2)
I am curious about what kind of daemon would notify the user of errors reported by the ECC code. I had to look at the system logs to see some ECC reported errors that the layer managed to fix. Sometimes even filling up the logs. There must be some kind of daemon for that right?
Re:I can't make much of it. (Score:4, Interesting)
On some Dell tower PC's from around 2008, the notification was that at the next boot, the BIOS firmware showed a message saying something like "A single bit memory error was detected and corrected, press F1 to continue."
I have some sticks of RAM (in a box of junk somewhere) that caused the message on the hottest days of the year.
Re: (Score:3)
I think you might be confusing boot time or off-line write/read memory check which require no ECC at all and run time memory check which can only be performed realistically with ECC. Of course, you could write ECC data to disk but this would slow down the computer so much that it isn't a practical solution.
Re: (Score:2)
Instead, some of them had suffered for years, thinking that their computer's behavior was normal.
People thinking a computer crashing is normal would have suffered through the OS throwing up errors they didn't understand thinking it was normal as well. You're not building a case for ECC, you're building a case for basic computer educations.
If a computer crashes once, shit happens. If it crashes twice it something to investigate, and it's insanely trivial to test memory.
Re: (Score:2)
True - if you don't expect your test to identify 100% of the problems - esp. those that may be temperature sensitive.
And it's definitely not "insanely fast" even to do a "kind of intense test" which actually belies the "insanely trivial" claim. I've had memory failures that would show up on, perhaps, once every 20 hours of running MemTest+ - and go through many passes which showed no errors.
Re: (Score:2)
I would argue that everyone shouldn't need to pay extra for suffering one extra crash per maybe thousand people per lifetime of their PCs.
Bonus points for not having to suffer a couple of extra deaths per those maybe thousand gamers for hardcore gamers because ECC memory latency caused a hitch at just the right moment.
Re: (Score:2)
If it was only that, I could go for it but the worst is data corruption, your computer keeps running most of the time and data on disk becomes corrupted slowly but surely.
Re: (Score:2)
This isn't about the disk.
Re: (Score:2)
I would argue that everyone shouldn't need to pay extra for suffering one extra crash per maybe thousand people per lifetime of their PCs.
The probability of one-bit errors is proportional to the amount of RAM. When computers had 64k of RAM, that might have been a realistic number. With the average computer having five orders of magnitude more RAM, the probability is five orders of magnitude greater.
Some studies have shown as much as 1 bit error per gigabyte of RAM per 1.8 hours. That means over a person's lifetime, a computer with 16 GB of RAM, the average person would experience over 6 *million* one-bit errors.
Re: (Score:2)
Excellent. How many of those flips are relevant to anything useful being done?
How many of these are catastrophic?
How many of these are related to processes that are critical in some way?
You're essentially making the oncology argument. That we get constant mutations across cells in our body, and therefore we're obviously in danger of horrible cancers!
Whereas in real world, almost none of the mutations are relevant. Of those that are relevant, almost none cause a terminal failure in the multiplication limitat
Re: (Score:2)
Excellent. How many of those flips are relevant to anything useful being done?
Somewhere between 0% and 100%? Every single-bit error has the potential to cause harm. The only question is whether the page of code or data will be used before it gets purged out of RAM.
How many of these are catastrophic?
Define catastrophic? Causes a crash? Causes some file that the user is working on to become corrupt? Causes filesystem corruption that results in the user losing a large percentage of files? All of those outcomes could be catastrophic or not, depending entirely on luck, and on whether the user is actually using the co
Re: (Score:2)
>Somewhere between 0% and 100%
Same as your chance of dying in the next second after reading this, yes.
The level of dishonesty of direction of this argument is staggering.
Re: (Score:2)
>Somewhere between 0% and 100%
Same as your chance of dying in the next second after reading this, yes.
That's a "who the heck knows" answer. Ultimately, the answer is that it doesn't matter how often it is going to cause data corruption, because when you have billions of devices, even relatively rare events happen frequently to someone. It's the law of large numbers.
And machines aren't affected equally [toronto.edu], so if you're in the lucky one-third who gets a machine that has a higher rate of problems, you're going to have a much higher risk of losing data. If every machine had ECC, this wouldn't matter, but becaus
Re: (Score:2)
>That's a "who the heck knows" answer.
Unless you're the world's fastest typist, you got the answer before typing that. It was a trick question aimed at trying to understand if you get the basics of how statistical mathematics works. You failed. The chance is zero.
Re: (Score:2)
Re: (Score:2)
ECC should be the default, not the exception.
No thanks. ECC comes with a massive performance limitation. I'll happily trade off the insanely rare chance that one of the bits in my memory is going to be flipped and may lead to a crash (odds are it won't) for the added performance benefit of not being limited to slow RAM speeds. And that's before we start discussing the general availability of DDR5 ECC in the first place.
If you're doing something important than ECC that shit. Most of the world's computers are not and it shouldn't be the default.
Re: (Score:2)
Some time ago, I read that more and more with RAM density increases, the RAM sticks had ECC internally, and a quick search seems to confirm [synopsys.com] it:
For every 128 bits of data, DDR5 DRAMs has 8 additional bits for ECC storage.
Which means that it doesn't even cost significantly less to make non-ECC RAM, it's just crippled by lacking the interface on the bus.
Re: (Score:2)
ECC is nice to have, but not a critical feature in the vast majority of consumer-oriented systems.
Re: I can't make much of it. (Score:2)
Once memory sizes get big the statistics don't look great for those running without ECC. If apps simply crashed on a bit error it wouldn't be a big deal. But introducing subtle corruption that an app later writes to disk can lead to a tech support nightmare for users. And costs software vendors millions in needless support calls.
Re: I can't make much of it. (Score:4, Informative)
There is a significant amount of errors introduced on the data lines between CPU and memory module. For starters the signal voltage is quite low on DDR5 (1.1V) and impedance is around 50 Ohms. It doesn't take much external noise to occasionally insert bit errors in a transmissoin. And that's assuming your DDR5's on-die termination is working correctly, without ODT I doubt DDR5 would even be viable as a memory standard. Regardless of error correction.
Moving forward, I think that so-called "consumer grade memory" should include CPU side ECC. And "server grade memory" should include multi-processor ECC and rDMA ECC. Anything less is a pre-2020 way of thinking about computer architecture.
Re: (Score:2)
Single bit RAM error rate is measured as some number of bits total per some amount of time. More time is more errors. But also more RAM is more errors. Do you think consumers should have more RAM or not? I think we should probably plan on them having more, but that's just me being funny basing assumptions on the past exponential growth of installed memory in consumer PCs.
The value that ECC adds is fewer tech support calls. Which saves businesses money, but it doesn't necessarily save Intel or AMD any money.
Re: (Score:2)
I'm convinced that we have tickets in our system where customers are having unexplained problems that are likely due to bad ram or bit flips. It's a waste of time and effort and does nothing but infuriate customers. But hey, someone saved a few pennies selling "consumer grade" RAM and made a ton creating an artificial tier of "enterprise grade" RAM.
CRC is, as you say, there to protect the channel. It doesn't actually prevent the propagation of errors. Which was kind of the whole point of my argument that co
Re: (Score:2)
BS. I would argue they would force people to pay more and get less if they actually introduced ECC support in those CPUs. Most people in the target group wouldn't need it
Until we start seeing Backblaze-level crash detail statistics for non ECC memory as a whole, I would argue we have no damn idea who the “target” audience is. All we do know is how many people rely on memory chips daily. And all memory has one main feature and service that is required beyond even capacity; reliability. Non-ECC memory is a bullshit downgrade that shouldn’t exist.
Re: (Score:2)
What about solar flares? /s
Re: (Score:3)
Non-ECC memory is a bullshit downgrade that shouldn’t exist.
100% AGREE
I've run Linux as my main desktop for nearly 3 decades. I got very used to not rebooting. I started getting more and more crashes (still pretty rare though... like 8 months), so I looked into building a system with ECC memory for my next system. I quickly learned that Intel didn't support ECC on the consumer line of processors, but AMD did! AMD supported ECC memory on most of their processors since a looong time ago, though it could be tricky to find the right motherboard that also supported it, a
Lack of demand (Score:4, Informative)
Re:I can't make much of it. (Score:4, Interesting)
ECC RAM has pros and cons. Pros are marginal extra reliability, mostly for things where flipped bit once a year isn't going to have any meaningful impact, nor ever get noticed.
Cons are higher cost and higher latency at similar binning levels per module.
(In before someone tosses in that infamous article that ran a few synthetic benchmarks and found no differences between memory with different timings to say that latency in memory doesn't matter).
It makes sense that for most people and manufacturers, ECC RAM is not worth it for general consumer PC market. Because adding it means wasting silicon and software work time to support it. And this is a strict negative for those who need low latency memory, as ECC memory adds memory latency. Spreading this among everyone, rather than a small group of users that actually do need it would significantly lower costs for this small group of users at the cost of everyone else having a minor increase in costs.
Re: (Score:2)
It's not just latency. Outright clockrate limiting as well. The highest performance parts are not ECC, period. I just looked at all my normal suppliers. They all stock ECC RAM, none of them 8000MHz which they all have available in non-ECC DDR5 variants. Heck 6000MHz is hard enough to find.
Re: (Score:1)
You can always disable or avoid ECC if you're extremely concerned with latency. Gamers like you who don't care about persistent data or consequential decisions based on data in RAM can do that. But I can't imagine most sysadmins in hospitals or manufacturers or millions of other companies purposely disabling ECC if it was already present in the millions of non-Xeon PCs they run their businesses on.
The situation we find ourselves in is created by Intel artificially segmenting the market into Xeon workstation
Re: (Score:3)
Mod parent up!
Lucky talks like the majority of people need the bleeding edge 8000MHz memory and don't care if their game crashes once in a while.
The converse is true. Most PC's aren't running high end games most of the time, but they would certainly benefit from more reliability. Most PCs would benefit from more memory, even if it was slower.
And bringing up overclocking??? Cause that's something EVERYONE needs, right? That's a corner that can safely be cut without impacting a single average user. "Because a
Re: (Score:2)
Like most people with highly specialized needs, you experience significant added cost. You have deep vested interest in making everyone bearing a large part of this cost by universalizing your highly specific needs.
Those who don't much care about latency still care about added cost, that comes with no benefit.
Re:I can't make much of it. (Score:4, Informative)
ECC memory is required if you want say, 16GB or more of RAM. After that, the incidence of bit flips starts coming down from once a month to daily or weekly. Now, for a desktop CPU, no big deal since you're going to reboot far more frequently than that for it to matter.
But if you're talking about 128GB of RAM, which isn't getting unusual for workstations (64GB is par for the course for higher end systems) then it starts getting to the point where you want ECC because you're going to get events on a weekly, if not daily, basis.
In fact, lack of ECC RAM is a limiting factor for large arrays of memory. It's so important even DDR5 has "soft ECC" logic built into it. It's not strictly ECC RAM, but it's implemented in the RAM hardware itself because the array is that vulnerable. (It's also what makes buying DDR5 memory exceedingly annoying because if you need ECC RAM, the fact DDR5 has ECC built in makes it hard to search for DDR with explicit ECC support).
Re: (Score:3)
Then don't use it. The problem at hand is that chipsets don't support ECC RAM, not that people would be forced to use it.
Also, I believe the real additional cost is negligible. ECC costs a lot more due to economy of scale and pricing politics, not parts cost.
Re: (Score:1)
This is not a problem, but a good thing. Because adding on crap that you don't need means paying for things you don't need.
It is a "problem" only for the small minority that actually needs ECC memory that wants others to subsidize it for them though much larger production numbers and economies of scale. For everyone else, lack of inclusion of useless features you will not use is an issue of reduced cost of goods. Which is a feature.
Re:I can't make much of it. (Score:4, Interesting)
Exactly this, and hardware companies are very skilled at creating artificial tiers.
Remember back in the day it was possible to patch nvidia GPU ROMs to change device ids? It would magically enable features that were not supposed to be supported on your device. You paid for it, they programmed it, but uses a paid switch, sorry upgraded model to enable those. (And of course they started encrypting that firmware).
For a very long time, for example, the "homelab" community had to play excessive dances to have their GPUs "passthough" to the virtual machines. Why? It was an enterprise feature only enabled in workstation or higher edition of the same exact chips (minimum 2x price). They finally relented, and enabled it without hacks, but the entire thing was a sh--storm.
Same with Intel + AMD and ECC. And of course the PCIe lanes. They have been relaxing a bit, since modern systems use 2 or even 3 storage devices over PCIe (nvme), hence we now have more lanes coming out of the CPU. But still not enough that it will run two full sized x16 GPUs on the same system. Again, market segmentation that is entirely artificial.
(HEDT is pretty much dead at the moment. Threadripper moved to the higher, more lucrative workstation segment, along with Intel's Xeon Gold).
Re: (Score:2)
HEDT is pretty much dead at the moment. Threadripper moved to the higher, more lucrative workstation segment,
Not sure that's the case. I was recently speccing out machines, a 32 core Threadripper (non pro) was about 2x the price of a 16 core Ryzen. So about twice the machine for twice the money, more or less.
Re: (Score:2)
It might have been true when you built your system, I will not dispute that.
However "this gen" ThreadRipper "non-pro" 32 core is... wait for it... $2500:
https://www.amazon.com/AMD-Ryz... [amazon.com]
The "pro" one is at the "reasonable" price of only $3,900:
https://www.amazon.com/AMD-Ryz... [amazon.com]
(You could buy an entire prebuilt brand name workstations for this price in previous generations).
Whereas the 32 core Ryzen is $550:
https://www.amazon.com/AMD-795... [amazon.com]
Re: (Score:2)
You're comparing a 32 core threadripper with a 16 core Ryzen.
Yeah looks like for more than 16 cores there's a fair premium, though. What's the over all work station price though with the motherboards and ram thrown in. Looks like the Ryzens are down a bit price wise, so you probably pay a bit more per core for a finished machine with threadripper vs Ryzen.
I'm not sure what the difference is now. If it's 2x more per core, I'll concede the point. If it's I don't know 1.1x then that's just a straight up line
Re: (Score:2)
You are might. I might have mixed 32 threads with 32 cores.
Nevertheless, 16 core, old gen TR is already 2x of the regular Ryzen:
https://www.amazon.com/AMD-Ryz... [amazon.com] ($1050 vs $550)
I could not find a current gen 16 core, there was a 24 core one:
https://www.amazon.com/AMD-Ryz... [amazon.com]
And again still does not scale linearly.
And to add insult, the 16 core Ryzen is faster than 16 core TR (though old gen):
https://www.cpubenchmark.net/c... [cpubenchmark.net]
Re: (Score:2)
You're comparing the ryzen to the pro threadripper not regular (still more expensive but not so much). Core for core the threadrippers are basically about the same as Ryzens of the same gen since they're more or less the same processor cores, but the pro ones have 8 channel DDR memory, not 2.
Anyhoo
I just plugged specs into my general go to vendor. (cyperpowersystem.co.uk), came out with a good 16 core ryzen workstation at about GBP 3080, all in and a 32 core threadripper (non pro) at GBP 6180, 24 core being
Re: (Score:2)
Not sure about the prices over there, but I gave both pro and non-pro examples.
This gen:
16 core Ryzen 7950X = $500, or $31.25 per core
24 core Threadripper 7960X = $1500, or $62.50 per core
24 core Threadripper Pro 7965WX = $2600, or $108 per core
Yes, TR is 2x per core for non-pro, and more than 3x per core for the pro version. (They changed one thing though, there is no more 16 core TR, fully separating the segments).
Re: (Score:2)
Oh I see yeah we're talking at crossed purposes.
yeah the raw chips are 2x per core for TR (non pro), though you get 2x the memory channels and higher capacity of course. I was looking at the overall system cost, where of course for the Ryzen, it's amortized over fewer cores.
I'm going on some rando reseller's system costs of course.
From that for anything up to 32 core, the cost per core of a complete system is pretty comparable (surprisingly so) between TR and Ryzen.
I wouldn't mind if it's cheaper, though!
Chris Siebenmann's comments from the past few days (Score:5, Informative)
Options for genuine ECC RAM on the desktop in (early) 2024 [utoronto.ca]
Linux kernel boot messages and seeing if your AMD system has ECC [utoronto.ca]
Laptop components (Score:2)
Re: (Score:2)
No they aren't, those are regular AM5 desktop CPUs.
Re: (Score:1)
Market for Lemons (Score:2)
sucks, again (Score:2)
And now, I was ready to go for a Ryzen 8600G or even 8700G... Asus is actually the only motherb