Intel's Pay-As-You-Go CPU Feature Gets Launch Window

(Image credit: Intel)

Intel's mysterious Software Defined Silicon (SDSi) mechanism for adding features to Xeon CPUs will be officially supported in Linux 5.18, the next major release of the operating system. SDSi allows users to add features to their CPU after they've already purchased it. Formal SDSi support means that the technology is coming to Intel's Xeon processors that will be released rather shortly, implying Sapphire Rapids will be the first CPUs with SDSi.

Intel started to roll out Linux patches to enable its SDSi functionality in the OS last September. By now, several sets of patches have been released and it looks like they will be added to Linux 5.18, which is due this Spring. Hans de Goede, a long-time Linux developer who works at Red Hat on a wide array of hardware enablement related projects, claims that SDSi will land in Linux 5.18 if no problems emerge, reports Phoronix

"Assuming no major issues are found, the plan definitely is to get this in before the 5.18 merge window," said de Goede. 

Intel Software Defined Silicon (SDSi) is a mechanism for activating additional silicon features in already produced and deployed server CPUs using the software. While formal support for the functionality is coming to Linux 5.18 and is set to be available this spring, Intel hasn't disclosed what exactly it plans to enable using its pay-as-you-go CPU upgrade model. We don't know how it works and what it enables, but we can make some educated guesses. 

Every generation of Intel Xeon CPUs adds multiple capabilities to make Intel's server platform more versatile. For example, in addition to microarchitectural improvements and new instructions, Intel's Xeon Scalable CPUs (of various generations) added support for up to 4.5TB of memory per socket, network function virtualization, Speed Select technology, and large SGX enclave size, just to name a few. In addition, there are optimized models for search, virtual machine density, infrastructure as a service (IaaS), software as a service (SaaS), liquid cooling, media processing, and so on. With its 4th Generation Xeon Scalable 'Sapphire Rapids' CPUs, Intel plans to add even more features specialized for particular use cases. You can see an example of the SKU stack above, and it includes all types of different Xeon models:

  • L- Large DDR Memory Support (up to 4.5TB)
  • M- Medium DDR Memory Support (up to 2TB)
  • N- Networking/Network Function Virtualization
  • S- Search
  • T- Thermal
  • V- VM Density Value
  • Y- Intel Speed Select Technology

But virtually none of Intel's customers need all the supported features, which is why Intel has to offer specialized models. There are 57 SKUs in the Xeon Scalable 3rd-Gen lineup, for example. But from a silicon point of view, all of Intel's Xeon Scalable CPUs are essentially the same in terms of the number of cores and clocks/TDP, with various functionalities merely disabled to create different models. 

Intel certainly earns premium by offering workload optimized SKUs, but disabling certain features from certain models, then marking them appropriately and shipping them separately from other SKUs (shipped to the same client) is expensive — it can be tens of millions of dollars per year (or even more) of added logistical costs, not to mention the confusion added to the expansive product stack. 

But what if Intel only offers base models of its Xeon Scalable CPUs and then allows customers to buy the extra features they need and enable them by using a software update? This is what SDSi enables Intel to do. Other use cases include literal upgrades of certain features as they become needed and/or repurposing existing machines. For example, if a data center needs to reconfigure CPUs in terms of clocks and TDPs, it would be able to buy that capability without changing servers or CPUs. 

Intel yet has to disclose all the peculiarities of SDSi and its exact plans about the mechanism, but at this point, we are pretty certain that the technology will show up soon.

Anton Shilov
Freelance News Writer

Anton Shilov is a Freelance News Writer at Tom’s Hardware US. Over the past couple of decades, he has covered everything from CPUs and GPUs to supercomputers and from modern process technologies and latest fab tools to high-tech industry trends.

  • salgado18
    Intel copying the EA business model: make a full product, then strip its features to sell separately. Instead of selling one full processor for $1000, sell one crippled processor for $1000, then demand $100 for re-enabling each feature. I know that's capitalism and market rules and all, but it sounds like dishonest money to me (EA customers agree on that). Hopefully that gives more sales to AMD.
  • InvalidError
    The other way to reduce the logistics costs of marketing a billion variants of each CPU with razor-thin arbitrary differentiations would be get rid of the stupidly super-fine arbitrary market segmentation.

    Making fully functional full-size CPUs only to disable most of their capabilities and sell/lease them back as hardware DLCs sounds stupid.
  • bigdragon
    I recall that Intel tried this in the past -- about a decade ago -- and it didn't work out. The software-defined upgrades/unlocks caused confusion and didn't always work properly. Hopefully this latest effort crashes and burns too. Intel really needs to simplify their overly segmented CPU lineup.
  • InvalidError
    bigdragon said:
    Intel really needs to simplify their overly segmented CPU lineup.
    Yup. Make 2-3 dies to cost-effectively service the major server tiers and maybe have 2-3 bins each dictated by where defects are most likely to land.

    Selling DLC CPUs would require that every DLC-powered CPU sold to cover a given market segment range be a golden sample or at least capable of the highest specs for the market segment. Doesn't sound cost-effective unless Intel's yields excluding fatal defects and edge losses are close enough to 100% that it can just write off sub-par dies, which would imply that the entire i5-i9 range is a manufactured farce.
  • spongiemaster
    Intel just doesn't learn. They are really missing an opportunity here. You can't just let customers pay for the features they want to add. You have to turn it into a loot box type experience where customers pay for a random feature and hope to get what they want. If not, they keep buying more until they do get what they want. Then Intel can setup an online shop where users can resell the features they don't want to other Intel users. That way you get people endlessly buying CPU upgrades in an attempt to win highly sought after features that they can sell for more at the online shop.
  • kal326
    So if the features are already in place then they are past binning for things not to work or work per SKU. So basically Intel is saying we could have been give you all these features in an entry level chip all along. Instead to save us hassle and maintain margins when charging more for arbitrary features we instead invested time and resources in to the Intel Chip Marketplace.
    Now comes the crash and burn of this when it’s linked to security exploits and they have to start disabling offerings or someone just cracks it to enable all features.
  • sabrix
    EA : That's my boy !! 🥰
  • salgado18
    I though of it even worse now: customer, who is a server owner, has 1000 Xeons and 200 EPYCs, for example. Now he needs this one feature enabled, because of a new service he'll provide. He has to contact Intel, pay 1000 new feature licenses, upgrade 500 machines (dual-socket, maybe), and pass validation tests to see everything is working. He also needs to do absolutely nothing on the EPYCs, since they already have said feature enabled.

    Do they really think that's the way to earn more money? In the short term, sure, but this customer's next purchases will probably be EPYCs, which cost less (they do), are more energy efficient (they are), and now cost less in the long run. I'd say it could be cheaper still to change older systems to new EPYCs instead of spending money on feature upgrades.
  • InvalidError
    salgado18 said:
    He also needs to do absolutely nothing on the EPYCs, since they already have said feature enabled.
    If he CPU DLC you got is a core count upgrade, the only way you are getting that out of EPYC-based systems is a physical CPU swap which is far more labor-intensive and will require even more extensive re-validation since you have a whole different chip instead of the same chip with fewer things still pay-walled.
  • hotaru251
    Tesla showed this business model (paying for a thign with features you have to pay more for to use) and it will be future sadly...