Intel is heading down the pay-as-you go route for CPU features on its Sapphire Rapid server CPU's; for now, it's only on server CPU's, but give it time.
Do you think its a smart move by Intel?
https://www.techradar.com/news/intels-pay-as-you-go-cpu-plat…
Intel are getting their traditional stronghold in the corporate x86 server space kicked around so badly by AMD that even Azure is offering cheaper prices on the same services hosted on AMD hardware compared to intel in order to try and get clients to migrate to cheaper to buy and run AMD servers. This is just another case of Intel trying to pull its tradtional trick of trying to extract as much money from the market by locking features and functionality behind certain "pay a premium" scheme like the whole K series CPU/ Z series chipset overclocking enablement fiasco.
This is what happens when you put accountants in charge of tech companies - in less than a decade Intel managed to turn
market dominance and a 4+ year lead on anyone else in the space into:
1. Being well over a year behind their competition's fabs in actually producing something on a given process generation
2. Losing one of their largest consumer product customers due to incompetence (Apple) so badly that it turned it into a fierce competitor
3. AMD going from a joke in the consumer space to leading performance and sales wise in pretty much every x86 category including servers
AMD servers are not cheaper to buy or run but they allow run more customer VMs. 64 cores in Epyc Milan vs 28 in Intel Xeon is a big difference.
Never said the hardware was cheaper, just that it was more cost effective overall. Intel has offered 56 core CPU's for a while now so the core count comparison is not quite as bad as you outlined but they were walked all over by Rome, Milan and soon Genoa generation Epyc products.
When cost and power draw is the same / similar and you can put significantly more customers or workloads on the that hardware - that is the very definition of cheaper as initial hardware purchase price is not even a blip on the radar of TCO for these systems.
to cheaper to buy and run AMD servers.
Never said the hardware was cheaper,
Make up your mind.
@[Deactivated]: Answer was in the first line.
"cheaper to buy and run" is very different to "cheaper to buy" - was pointing out the TCO on these systems is significantly cheaper than intel in a lot of cases where you can put those extra resources to use effectively. Not to mention that in a lot of cases the systems are/were actually cheaper to buy than an similar spec Intel system depending on the chosen vendor.
If you are calling out a 28 core Xeon as a comparison to a 64 core Milan then the 2S/4S end of the market probably isn't something you are keeping an eye on and in that space AMD has a significant pricing / TCO lead to the point where big cloud vendors are passing on some of the savings to customers in order to get them to jump their workloads across so they can retire the older / less efficient gear faster.
was pointing out the TCO on these systems is significantly cheaper than intel in a lot of cases where you can put those extra resources to use effectively
You're talking about TBO. TCO is about the same.
If you are calling out a 28 core Xeon as a comparison to a 64 core Milan then the 2S/4S
You obviously have no idea why cloud vendors don't like to use 4-socket systems.
@[Deactivated]: I have read through that comment a few times and am still at a loss to explain where I claimed that cloud vendors are using 4S systems for any significant portion of their fleet?
When I did my last large scale migration in a DC we reduced the number of compute nodes significantly (around 27%) while almost doubling capacity thanks to more cores, higher base/boost speeds and signifincantly better memory bandwidth available - thus the TCO costs of the solution including climate, power, spares and staffing etc fell significantly. TBO is for the sales and accounting departments to work out, TCO and occupancy / utilisation rates are what the people running the hardware care about.
@baronorder: Sigh.
Total cost of ownership (TCO) is an estimation of the expenses associated with purchasing, deploying, using and retiring a product or piece of equipment. TCO, or actual cost, quantifies the cost of the purchase across the product's entire lifecycle.
Do you see savings from upgrade or higher revenue from being able put more customers into same footprint in that definition? Me neither. That's TBO.
TCO of Intel and AMD servers is approximately the same and mainly depends on power consumption of CPU, everything else is exactly the same.
@[Deactivated]: Sigh.
As you said it refers to the product and in this case it was a DC for a single client and the product was the entire solution that needed to meet set performance and reliability requirements. As you said "TCO, or actual cost, quantifies the cost of the purchase across the product's entire lifecycle".
With less hardware needed due to much higher capacities on offer affordably by AMD hardware at the time - less power use, less cooling required, less spares,less UPS, fewer generators and less staffing / maintenance required by the solution meant that the TCO per item dropped as some of those are not linear costs. TCO is metric designed to include opex costs for that item across the entire solution required to operate it for the lifetime of the item and that being able to cut down the size of the solution signifincantly and opex required impacts the TCO.
TCO is the same if the products on offer are the same spec, cost and opex wise. In this case they were not and the AMD solution was significantly cheaper for the performance required compared to Intel based solutions available at the time thanks to the performance density on offer allowing the solution to be smaller.
@baronorder: TCO does not talk about performance, specs and whatever else you think it talks about.
It's just price plus running costs, no matter whether server is sitting idle or not.
Get a grip with basic terminology and come back.
@[Deactivated]: Smaller solution for identical performance level, fewer higher performance machines at similar price per unit, less opex = lower TCO
Larger solution for identical performance level, more lower performance machines at similar price per unit, more opex = higher TCO
See how performance plays a part in deciding the solution and thus the TCO for equipment in that solution? In addition to this machines are spun down and up dynamically based on usage and requirements in order to reduce running costs so yes idle time and utilisation also plays a part in the calcs for figuring out an accurate TCO in certain circumstances.
However I am sure that my understanding of the concept is wrong and will defer to your expertise on the subject.
@baronorder: You seemingly unable to understand difference between TCO of server, which you started from, and TCO of solution, to which you switched later. Those are related but different things.
@[Deactivated]: You are missing the forest for the trees and doing so wilfully.
I never started on the TCO of the server - you just made that presumption because it suited your argument.
It was always implied (pretty clearly might I add) as part of a solution or system as these servers are not purchased to free-range in a field somewhere doing nothing.
Clients want these server as part of a system / solution and pay the total cost for that system or solution, not just individual line items with complete disregard to the total figure.
They tried this 12 years ago with the Celeron and later some i3's. 'Intel Upgrade Service'. Suffice to say, it was not well recieved and they ended up unlocking the features for free after it folded.
I already pay as I go for CPU with my VPSs. I usually go AMD tho.
People should remember that these features are already on the chips and it cost money to put them there, thus unless Intel is willing to take less of a profit on the sale of these chips, this is nothing more than double dipping.
That's the same as what's happening with cars though. Heated seats paid subscription? Your car obviously has the tech.
How is it different from apps with premium versions or different plans for cloud SaaS?
They have already paid the cost to develop those software features and the marginal cost to make them available is practically $0.
It is completely different! This is not software, It costs money to build these features into each and every chip!
It's more a matter of degree.
Whilst not nothing, the marginal manufacturing cost difference from the additional silicon die area dedicated to those accelerators is minuscule relative to the product customer price (ballparkish ~$10 per $5000 retail chip).
The majority of the cost of those specialised accelerators is actually the hardware R&D, validation, library software development and ongoing bug fixes & features, enabling feature support in third-party open source projects + ISV software, and ongoing specialised customer support.
That's why I think it's not that different to premium app versions, plugins, or SaaS plans. I think it's quite different to the ridiculousness of heated seats subscriptions where there is no IP or ongoing costs to the car manufacturer.
‘You will own nothing, and you will be happy’