Intel teases 'software-defined silicon' with Linux kernel contribution – and won't say why
- Reference: 1634622973
- News link: https://www.theregister.co.uk/2021/10/19/intel_sdsi/
- Source link:
SDSi popped up around three weeks ago in a [1]post to the Linux Kernel mailing list, in which an Intel Linux software engineer named David Box described it as "a post-manufacturing mechanism for activating additional silicon features".
"Features are enabled through a license activation process," he wrote. "The SDSi driver provides a per-socket, ioctl interface for applications to perform three main provisioning functions." Those provisioning functions are:
Provision an Authentication Key Certificate (AKC) – a key written to internal NVRAM that is used to authenticate a capability-specific activation payload.
Provision a Capability Activation Payload (CAP) – a token authenticated using the AKC and applied to the CPU configuration to activate a new feature.
Read the SDSi State Certificate – containing the CPU configuration state.
Box's post also pointed to a [2]GitHub page that includes the following explanation:
Intel Xeon family processors with support for Intel Software Defined Silicon (SDSi) allow the configuration of additional CPU features through a license activation process.
Between that GitHub mention and the three functions added to the Linux kernel, it seems plain that Intel could ship Xeons with latent features you could enable by sending it some cash.
Intel's offered precious few other details. The GitHub page includes a [3]document detailing how to use SDSi-equipped silicon to enable dormant features, but with no detail on what new features could be activated with this tech.
[4]
The Register asked Intel to explain its Linux Kernel mailing list post. Chipzilla offered us the following non-committal response:
“We’re not going into a lot of details about Software Defined Silicon at this time. As you know, Intel regularly submits code to the Linux Kernel that could be used in future products. And that’s what happened in this case. If we plan to implement these updates in future products we will provide a deeper explanation of how they are implemented at that time.”
Yeah, right. Intel has gone to all the trouble of cooking up a way to license highly configurable Xeons, but hasn't decided if it will become a product, and tossed the tech into the Linux kernel anyway.
If you believe that, The Register has a bridge we'd like to sell you.
[5]
[6]
So let's ponder what Intel could be up to here – starting with why Intel wants to license CPU features.
Today, Intel sells a CPU and as often as not doesn't see any more cash from its customers until their next purchase – which could be years into the future. Licensing CPU features would potentially give Intel more revenue, more often, perhaps even letting it create the kind of subscription services that investors adore because they boost revenue – and make its arrival more predictable.
[7]
Intel is going to need predictable cashflow to fund its plans to spend [8]tens of billions on new factories.
[9]Alibaba Cloud unveils home-spun 128-core Arm-powered server CPU
[10]Apple arms high-end MacBook Pro notebooks with M1 Pro, M1 Max processors
[11]Arm puts virtual hardware in the cloud so you won't have to wait for the actual chips
Those factories are infamously complex, and Intel works them hard – partly because it makes many variants of its products. If Intel could make fewer variants, and instead pack all its tech into a smaller number of SKUs that could be re-configured in software, production savings could be substantial. Customers would still pay a premium for high-end kit, which would be switched on by software rather than created as discrete products.
We also know that Intel plans to make its products yet more complex, with the [12]"Alder Lake" architecture that mixes and matches cores of different kinds on the same die.
Configurable CPUs could delight customers, by letting them buy a CPU with advanced capabilities like Intel's AVX-512 extensions – a tech aimed at speeding machine learning – but pay to use those extensions only when needed, rather than wearing an up-front cost. Or buyers could acquire servers knowing they have some extra overhead to turn on as their needs increase.
That kind of flexibility is not far-fetched. In fact, Intel already offers something similar in its [13]Speed Select Technology (SST) – an offering that allows users to set CPUs into configurations matched to different workloads. SST also allows definition of virtual CPUs with characteristics that differ from the physical CPU.
[14]
Another current option emphasising composability and flexibility comes from HPE, which offers [15]silicon-on-demand that allows customers of its GreenLake ITaaS environment to vary the numbers of cores that are active on Intel-powered servers and pay only for those cores, rather than having to rent a whole server.
Remember, too, the mid-2010s fad for [16]composable infrastructure – the notion that a collection of connected components could be assembled into servers that meet the requirements of the day. That sort of concept nearly always takes a few years to go from nifty idea to practical adoption.
Intel ended its response to The Register 's questions about SDSi by stating: "We're continuously innovating to ensure we enable flexible solutions to meet the unique demands of our customers and partners and lead the industry in product capabilities and features."
Which means absolutely nothing. Yet going to all the trouble of inserting SDSi in the Linux kernel surely signals something substantial is brewing. And if that something is configurable and/or composable CPUs, Intel may have something rather more substantial to say. ®
Get our [17]Tech Resources
[1] http://lkml.iu.edu/hypermail/linux/kernel/2109.3/01165.html
[2] https://github.com/intel/intel-sdsi
[3] https://github.com/intel/intel-sdsi/blob/master/os-interface.rst
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/front&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2YW6XT52sx0mhTLQV0xOyhQAAAAQ&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[5] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/front&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44YW6XT52sx0mhTLQV0xOyhQAAAAQ&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[6] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/front&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33YW6XT52sx0mhTLQV0xOyhQAAAAQ&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[7] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/front&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44YW6XT52sx0mhTLQV0xOyhQAAAAQ&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[8] https://www.theregister.com/2021/10/07/intels_80bn_european_chip_plant/
[9] https://www.theregister.com/2021/10/19/alibaba_yitian_710_arm_server_cpu/
[10] https://www.theregister.com/2021/10/18/apple_macbook_pro/
[11] https://www.theregister.com/2021/10/18/arms_virtual_hardware/
[12] https://www.theregister.com/2021/08/19/intel_architecture_day/
[13] https://www.intel.com/content/www/us/en/architecture-and-technology/speed-select-technology-article.html
[14] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/front&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33YW6XT52sx0mhTLQV0xOyhQAAAAQ&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[15] https://www.theregister.com/2021/06/22/hpe_discover_silicon_as_a_service/
[16] https://www.theregister.com/2015/10/26/composable_infrastructure_servers_that_give_software_more_to_define/
[17] https://whitepapers.theregister.com/
If they're unwilling to document the features that can be enabled
one might reasonably draw adverse conclusions.
my guess is (last line if you want to skip the background)
The way chips were made, I have no idea if it is still the same or not, was that you would aim to have every chip be the best possible but through imperfections during the manufacturing process some parts of the chip may not work as designed. They would test them all to bin the dies by maximum stable clock rate and functioning cores. And traditionally they used to burn silicon fuses to permanently disable parts of the chip that just did not work, and also permanently disable access to perfectly functioning parts that just did not allow the chip to fit into any of the current product ranges sold. Say a 4 core 8 thread silicon die that only had three corectly working cores (6 threads) and one partially work core (1 thread) could never be sold as a 4 core CPU, 3 core CPU is not in any product range, so they would sell it is a dual core (4 thread) CPU and traditionally use blow fuses to permanently remove access to the extra working and partially working cores.
At a guess the silicon fuses have been replaced by the Intel Management Engine, which even reports back what the chip is when queried. And this new mailbox talks to the IME to "re-fuse" disabled fully functioning cores/threads.
At the end of the day they can sell the exact same silicon inside the same package that fits into the exact same socket and then get you to cough up extra cash to enable cores that are there but you did not pay for when you initially bought the chip.
The start of licensed instructions
Fast forward 2030 when buying a CPU:
You will get one SUB instruction and one JMP on zero instruction for free (i.e. the CPU is Turing complete). All other instructions require additional licenses for specific instruction set groups. There are no perpetual licenses. You can only acquire a license for one month at a time with an option for a 12 month discount license.
The software defined CPU is perfectly normal. It is an innovative solution to cover the development costs of our highly effective processor products. The license system will ensure a more stable revenue stream for better innovation in the CPU space.
Please, don't be alarmed.
Re: The start of licensed instructions
On the other hand, you would imagine that the future trend would likely be for an increasing proportion of "serious" processor sales to go to cloud providers where you really need the flexibility to move workloads around between identical systems.
Unless that's actually the target: Amazon or Google or whoever can get cheaper chips for no-frills workloads but run more-demanding workloads on the same hardware when required by temporarily enabling the different features, reducing the different hardware platforms they might otherwise need.
This is where competition is a good thing. One more reason to use an alternative rather than get caught.
ARM increasingly is the future if recent announcements and marketing stack up to reality.
Clearly this means...
The boffins at Intel have found a way to emulate a Xeon running on a 286 chip. We have a solution for the silicon chip crisis!
Now, where did I put that old IBM PC?
There goes the brand
Crippleware chips. A bold move.
Re: There goes the brand
This is done in the real world already. I know that some manufacturers of mainframe style systems do that. You buy a license for a certain number of cores that are enabled, and can license additional cores should workload increase in the future.
Yes, it is a license to print money. No, I don't really like it. And there are other license models (like paying for the CPU time actually used) by the same vendors. In the end, market (and your company's bean counters) will decide.
Yield?
I'd always understood the variety of SKUs to be how Intel gets a good yield - if function A doesn't work on this chip, it gets packaged up and sold as the SKU that doesn't have that functionality.
If it won't clock up to the max, it's sold as the version with a lower maximum clock speed that it can do.
If Intel start selling licences to use specific pieces of silicon, then they'll also have to make sure all those bits actually work as advertised, even if the customer never buys a licence as a customer who rents a licence only to find it doesn't work is going to be mightily and publicly annoyed.
Re: Licenses
Licensing CPU features would potentially give Intel more revenue
Larry, is that you?
Won’t bode well
This will work well if Intel can get great quality from their fabs and purposely SD downgrade capable chips for customers to then pay for enhancements later.
But everyone in the chain wants their money from delivering from end to end. Intel are cutting out a large chunk of that chain, including their own FAB’s by not constantly shipping new.