All nodes lead to Rome: Epyc leak spills deets on second-gen Zen 32-core AMD server chippery Benchmarks of engineering samples of AMD's second-generation Zen-based server processor silicon, codenamed Rome, have once again found their way online. Back in March, we saw stats detailing the performance of the chip designer's upcoming 64-core, 128-thread flagship second-gen Epyc part. The info was revealed to the world by …
It’s great to see not only competition and AMD beating Intel at their own game. Unfortunately an upgrade other than for very specific workloads is still largely pointless even versus a 5yr old cpu. Lower power consumption is great, however if comparing cost of ownership plus power consumption, you probably need to retain your older Xeon for a few years before swapping to optimise total cost of operation. AMDs research and manufacturing success is still going to be a hard sell in the boardroom. That said, when you do hit that optimum, no question AMD are now king of the hill.
I doubt mainstream businesses upgrade CPUs very often. If a company has a rack of dual CPU servers supporting their business and they are working acceptably, they don't risk disrupting their business by tearing apart that rack. Instead, they'll just keep those servers going until they are depreciated, then replace them completely with newer ones.
The only large scale upgrades I've ever read about were supercomputers, where they bring the system offline and replace (e.g.) 8096 processors at once. They can do that because time on that system is scheduled, and it's not used for daily purposes like processing credit card transactions.
We usually do expansion and upgrade. First buy new rackfull of new servers to replace that rack (or two racks), then disconnect old rack(s) and pop new CPUs. Then use old rack(s) as dev for new deployment, and when those are moved to prod, rack(s) are moved to production again. But that does not happen every month or even every year.
Yes, agreed - I still use a few very long lived UltraSPARC systems in corporate environment. Been dependent on eBay for spares for quite some time. I'm more thinking of high demand cases, VM nodes and the like where there is a reasonable demand for CPU and clock count. Cost benefit for trading Xeons in that role is very hard to make stack up.
"I doubt mainstream businesses upgrade CPUs very often"
It depends on scale.
Smaller companies likely have a 1-3 node VM cluster and moving from one CPU vendor to another will require a replacement of the existing environment, although the migration itself will likely be a case of shutting down VM's and moving them to new storage connected to the new servers, adjusting resource allocations and powering up. i.e. it is probably part of a X year refresh where X>>5
In larger environments where you have multiple VM clusters or replace racks at a time, rolling upgrades are relatively common - spin up a new cluster/rack, test, arrange downtime, shutdown the guest and move it to the new host and repeat until the old VM cluster is empty and you can repurpose it or decommission it. The challenge is if the new CPU's provide any real advantage as more cores can affect licensing and the different CPU architectures prevent you merging VM clusters, although the larger package (more IO, more cache, probably lower power/heat, and less concerns around speculative execution and future performance) may alter that decision. Naturally, physical servers are harder.
At cloud scale, build/test your racks off site, deliver X racks of the new hardware to required locations, add the new instance type and pricing and let the customers at it... If customers like it (particularly given it being ~10% cheaper than Intel at present on AWS), profit :)
I am biased as have way too much invested in AMD shares.
Maybe Cloud Datacenter expansion will be mainly using AMD servers. Capacity on the existing (degrading) Intel servers can be managed by using pricing to persuade users to move from Intel instances to the AMD instances. The 10%+ discount is a relatively painless way for a CIO to meet budget.
No doubt Intel have some long term contracts with the cloud vendors, although these may be void due to the security issues. I would not be too surprised if Intels market share drops faster than anticipated in the second half of the year.
Don't underestimate the lure of vMotion / live motion / whatever else it's called.
If you get the basics of your VMware cluster right, you can migrate to newer Intel CPUs as and when they become available. Simply add the new machines to the cluster, vMotion across, and the users are none the wiser. Once everything is across you can remove the old servers from the cluster and upgrade the CPU baseline. Seamless.
If you choose to switch to AMD instead, you have to set up a different cluster. Then you have to shut down each VM, migrate it to the other cluster, and then start it up again. It doesn't take long but it's disruptive. Server people don't like "disruptive".
Where I work we're now down to about 2 hours per day without people scheduled to be logged in, so taking down a core system to move it in that way could be a chore. Not saying it can't be done, not at all. All I'm saying is that's going to cause significant resistance against switching vendors.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
