I'm not as sketched out about this as if it were single socket workstation ryzen/threadripper CPUs. In the market from $1000 to $6000 workstation desktops where enthusiasts and people with specific requirements (or just 10, 15, 20 years of experience building x86-64 PCs themselves) would want to build their own desktop from individual components ordered off Newegg.
I doubt more than a single digit percentage of 'serious' dual socket (64-128+ core) rackmount server customers are going to be buying their own barebones motherboards and CPUs and assembling it themselves. They're going to buy it from a Dell, HP or a Supermicro integrator or similar. If you're buying a $12,000+ server with 128 cores and 512GB to 4TB+ of RAM and some fast NVME storage it's highly unlikely you're putting it together yourself.
Any massive hosting/cloud scale operations that want to DIY their own EPYC systems from pieces will be doing it through a Taiwanese integrator, such as those that supply the ecosystem components for open compute platform server motherboards. And as such they'll also not encounter any technical issues or procurement issues with this. At the point where you have two $3000 CPUs on a motherboard that costs $1200, the full firmware/motherboard/CPU integration and qualification process is very different than putting a $399 ryzen into a $300 board.
Not yet. We will in 5 years time when their resale value meets our budgets.
I work in a rather budget-constrained lab environment. “Beg, borrow, steal” is the order of the day. Just today I was pricing out pre-loved Gen8 HPs. In 5 years time I could be exactly the hypothetical the article outlines.
This isn’t today’s problem - it’s a problem we’re creating today. We’ll hit it when your examples start retiring them and my example are eager to recycle them.
Yes - and no, I know lots of people including myself who have things like older dual socket Dell R710 as home test hypervisor servers. Also a very tiny percentage of people will bother to ever upgrade the CPUs on them.
For home lab stuff... When people buy a $200 used Dell R610 off ebay with two 8-core CPUs they most likely expect to use it in the exact same CPU configuration. Maybe add RAM. And probably use their own choice of SATA 6Gbps SSD in the drive trays instead of whatever old, possibly unreliable used spinning drives might come with it.
I have a 4U, quad socket Dell R910 with 32 total cores and 256GB of RAM that I got for $350. I'm absolutely not going to go messing around with replacing the CPUs on it with something I've purchased from ebay. When it's too old or slow, or I'm tired of having a 500W electrical load in my garage, I'll replace it with another thing that's come off a 3-4 year lease cycle.
My go-to vendor’s business model is to do that for me. So I say, I want a DL360. I pick processors, ram, disk controllers from their stock. I can even tell them how many caddies I want (looking at you, Dell). And they ship me a build-to-order server from second-hand parts.
So in the future I’ll likely have a smaller bin of CPUs to choose from. If firmware keys get more specific than per-vendor, it could be potentially a very small bin. And small bins typically mean higher costs. The cheapest cpu is typically the biggest bin, not the highest specced.
I have never purchased a used server and aruck with the included CPU’s, they’re either power hungry beasts or bottom-rung SKU’s. All of my 12th gen PowerEdge servers at home run E5-2450L’s (they’re all -EN platforms), for example. The one exception is the R210 II I use as a firewall/router.
By percentage the number of 1U/2U servers sold with ultra power efficient CPUs is fairly low. When people buy those new they will absolutely be going for CPUs that are 85W to 130W TDP per socket, times two sockets.
As a person that's formerly worked for a server manufacturer for a number of years I would say that the mid to upper performance range of the CPU market is 80%+ of the servers by volume. The other 10% is either the very low power models, and the top 10% of the units sold by volume are the very most expensive CPUs available at the time.
If you buy a used 1U Dell R610 with two six-core CPUs and 64GB of RAM, nobody should be surprised that a 120VAC watt meter at the wall shows it idling at 150W power consumption, with cpu load at 0.00... [surprisedpikachu.gif]
I mean, they don’t have to be the ultra-efficient ones - but for my home lab use I want < 100W idle usage and even my R520 can handle that with the 2450L’s (Ivy Bridge-EN could do this without the L suffixed SKU’s, but HCC chips in that family were more expensive when I was buying).
I haven't yet purchased used server gear, but I do a lot of window shopping. Most of the servers come with CPUs, but not all of them, and there's always a lot of loose CPUs for sale. Very occasionally, I've seen new-old-stock server motherboards for sale for not too much.
> I work in a rather budget-constrained lab environment. “Beg, borrow, steal” is the order of the day. Just today I was pricing out pre-loved Gen8 HPs. In 5 years time I could be exactly the hypothetical the article outlines.
Likewise for me. I AM building a $20,000 HPC because simply put, no one will sell us one for anything close to what we can actually afford and when it affects the speed and capability of my research and publications personally, it feels like a waste to leave extra performance on the table.
If I may ask, using what motherboard? The problem of EPYCs being locked to a certain vendor platform should not be a problem for you if you're buying a factory new Supermicro or Tyan or competing board. And a new set of individually purchased EPYC CPUs.
As a University student, I had dual Athlon XPs on a Tyan board. It was a lot of money for me back then, and the power supply melted to the board (not the power supply's fault either. I soldered on a new ATX connector and the 2nd power supply also melted to the board). If I couldn't use those processors in any other machines except a Tyan, I would be insanely pissed off.
I doubt more than a single digit percentage of 'serious' dual socket (64-128+ core) rackmount server customers are going to be buying their own barebones motherboards and CPUs and assembling it themselves. They're going to buy it from a Dell, HP or a Supermicro integrator or similar. If you're buying a $12,000+ server with 128 cores and 512GB to 4TB+ of RAM and some fast NVME storage it's highly unlikely you're putting it together yourself.
Any massive hosting/cloud scale operations that want to DIY their own EPYC systems from pieces will be doing it through a Taiwanese integrator, such as those that supply the ecosystem components for open compute platform server motherboards. And as such they'll also not encounter any technical issues or procurement issues with this. At the point where you have two $3000 CPUs on a motherboard that costs $1200, the full firmware/motherboard/CPU integration and qualification process is very different than putting a $399 ryzen into a $300 board.