39 comments

[ 3.9 ms ] story [ 86.2 ms ] thread
I hope to god the big three cloud providers pick up EPYC in good quantities. High core counts, lots of PCIe lanes, competitive perf per watt, what’s not to like? We badly need some kind of viable competition for Intel.
Given the pasting meltdown had on Intel performance and cloud providers pretty much have to run the patches, if Epyc doesn't sell well then Intel is pretty much unassailable.

I'm not convinced they are Opteron showed they can be vulnerable and Intel's handling of meltdown and PR was terrible, I imagine lots of pissed off CTO's.

It might be an opportunity for a cloud startup to create a whole business around AMD, it may be too hard for the established players to switch at this point.
The problem for folks below that tier is that the ecosystem for AMD is terrible.

On the software side, basic tools like numatop don't work on epyc. That's because the the equivalent of the Intel uncore counters that measure infinity fabric (AMD's new Hyper Transport, like Intel's QPI) are not available. Perhaps the big three have the time to do these tools themselves, and the NDAs in place to get access to the counters. Since even their single socket offering has up to 4 numa domains, getting access to these counters is pretty important.

There is also a limited selection of motherboard offerings from just a few vendors. This won't impact some of the cloud providers (who do their own board design), but for the rest of us, it is a pain. We've had an epyc we've been testing. I'm its biggest fan, but the motherboard is terrible in an enterprise setting. Just this week I updated the BIOS on it to get the new AGESA. Unlike our Intel / Supermicro boards, the BIOS update wiped all the settings. Earlier BIOS updates have required a physical (at the wall) power cycle, rather than a remote ipmi chassis power cycle.

Why AMD is not publishing these counter docs/specs? Is the Intel equivalent public? (I assume, since numatop is open source, and uses perf counters from the standard Linux kernel - no binary blobs, right?)
numatop was not updated since 2015 AFAIK...
It was actually updated in late 2017 to add POWER support.
Educate a noob: What is the significance of overclocking at this insane frequency when the CPU is not stable? From a digital design perspective I would imagine that a lot of timing constraints are not met, and this makes the CPU unstable. What prevents other CPU's unable from reaching this speed?
I would guess that the remarkable fact is exactly that this CPU is still able to work somewhat properly at these extremely out of spec conditions, thus proving a good design/production process/idk
Yeah, beside the "why not" factor, it's always reassuring that your device has so much room left.
AFAIK, other than "because we can", "stable enough to get into an OS, but not to be reliably used" is pretty close to "stable", so assuming tech iteration doesn't result in a quality drop, you can think of it as a "Coming Soon" for what a generation or two later's high-end chips can be pushed to.
I am not sure if they got into the OS at that speed. From the video I got the impression that they adjusted the core frequency and voltages using software, once inside the OS.
Getting a CPU Z validation is a sport, nothing more. But they have to develop tricks and techniques to further their sport, which may have other benefits in terms of understanding the hardware.
What is preventing someone from doing the same with the latest Intel chip? Do they fry immediately or do they just emergently shut down?
Different chips have different limits. I used to do some of this competitively back in the day. Different architectures can't get as high max frequencies, some respond differently under the extreme cold. These records are typically done using liquid nitrogen or phase change cooling and a lot of chips will freeze up below a certain temperature. Then its also a toss up between different individual cpus of the same model. It's a bit of a crapshoot.
Hell I even remember back in the day hunting down RAM produced from specific batches
That's nothing, the old P4s could do 8GHz on liquid nitrogen! I guess it's impressive that it's a multi-core AMD processor rather than an architecture that clocked well but had little else going for it.
Some one managed to put a 486 to 1GHz using liquid nitrogen (Probably I read it wrong)
POWER6 had stock clocks that went up to 5 GHz. I wonder what the most someone has overclocked one of those to is?
(comment deleted)
I'm pretty happy about Ryzen + architecture. They have at least proven that they can iterate on their existing architecture and work on a 12nm die. I was an early adopter of Ryzen and was happy to get my 1800x running stable with all cores at 4GHz. The price to performance ratio of Ryzen is great for almost all workloads.

While this article doesn't mean too much for real world applications. It does show that AMD is really competing with Intel. I agree that if Epyc can penetrate the server market (which I think it can), then it's really going to start an innovation war between Intel and AMD which is going to be good for the consumer.

With no onboard graphics you are required to buy a graphics card. That can raise your cost a significant amount if you weren't planning on buying one. If it ever dies, and every build I've owned has had the GPU die at some point, your computer is unusable until you buy another. Plus you can't do any Android development unless you boot into Linux. I'm sure these issues don't affect most people, and probably saves some gamers a little money that they can put toward their GPU, but I had two similar builds for the same price, picked AMD Ryzen over Intel to support the underdog and I regret it. I'm planning to give my Ryzen build to my folks and buy an Intel.
you regret buying a ryzen because you had to buy a gpu, and why can't you do any android development? can you not use genymotion or a real device for development?
No, but I am saying that should factor into the cost. If I buy a CPU with similar performance to Intel for $20 less but have to buy a graphics card for $50, it's not really cheaper. The regret is just the difference between being able to do everything and having to work around it because my CPU isn't supported. I doubt there's anything you can run on an AMD chip that you can't run on Intel but the opposite isn't true.
Wow. I did not know about the Linux/Android development thing. Just looked it up. That is so weird.

I'm about to build a Ryzen system. I run Linux and I don't do any Android development, so neither of those affect me. Still, that is a weird limitation.

Getting a CPU with GPU is actually something I try to avoid. I don't want or need it, as I use 3d software primarily. It's a complete waste for me. So, as you can see this argument depends on the needs of the user. I have avoided buying a CPU with built in GPU so far, glad AMD is this way, I am converted.
There are APUs for the Ryzen Platform and IIRC Ryzens latest have GPUs.
Thats cool, I can spin my legs at 200rpm on a child's bicycle for 1 minute before it breaks.
And shortly after, sustainable fusion was achieved.
While impressive, it's not really a barrier is it.
Since TechRadar redirects me to some boilerplate BS because I block 3rd-party scripts by default, I needed to use a mirror. Here it is for anyone else who needs it:

http://archive.is/yyBpa