Red Hat signals Intel's software-defined silicon will debut in Linux 5.18 Intel has revealed it plans to have its mysterious plans for software-defined silicon delivered in the next version of the Linux kernel. The existence of software-defined silicon (SDSI) emerged in October 2021 when Intel staffers posted to the Linux Kernel mailing list with hints about new functionality that would allow users …
I'm fairly sure that stuff with disabled features is not new. The same silicon did more if you pay more. I'm scratching my head a bit but disabled cores on CPUs seems to ring a bell. IIRC you could DIY by enabling a connection with silver paint.
With modern line widths etc, software is probably the only feasible method. But if you are giving it away with the GPL why not just enable it in the first place?
Intel has been selling CPUs with disabled features for ages. The only difference is instead of permanently fusing off features in the factory and selling them with different SKUs depending on what's disabled, they can have fewer SKUs and allow CPUs sold to be field upgraded for a fee.
I remember getting an old (pre-AMD) ATI not quite top of the line graphics card, when they were first released, they had too much of the "good" silicon so they were selling the cheaper version with the same chipset as the absolute top of the line card.
All you had to do was pop the heatsink and fan off, connect a trace that had been cut on the side of the chip using a pencil or some solder, and you had the full card for the £150 less.
Found it - early Radeon X800 Pro could be turned in to Radeon X800 XT Platinum Edition, or more accurately "Graphite Edition"
You're probably thinking of the three core Athlons.
The original plan was to take four core dies with a duff core and repackage them as a three core.
It turned out that picking out single core failures and feeding them back into production was more expensive than just running one, four core process line and fiddling it at the end.
I recall unlocking cores in various AMD processors. Same CPU, different SKU. Often lower SKU had some cores fail in testing and it was hence sold with less cores. Sometimes you got lucky and could unlock cores and they worked fine.
I believe that at one time Intel sold 486s in which the maths co-processor had failed testing as 486SXs to the status hungry executive marked, then reverse two pins (power and ground?) on perfectly standard 486s and sold them as 487 co-processors. Both 486SX and 487 cost more than a 486, netch.
"But if you are giving it away with the GPL why not just enable it in the first place?"
The key to unlock it is not GPL-ed, and that is what you pay for.
But will that key be unique and blown-in? Will there be a master key for test purposes or for resetting the enabling key? How long before keys start appearing on the black market?
"new functionality that would allow users to purchase licenses that turned on capabilities physically present in processors, but which are not available to use out of the box."
The difference here is that this opens the door to dynamic enablement and disablement. Once the licence becomes subscription based, you can be locked out of essential functionality at the drop of a hat.
So, if the Internet isn't working at all, isn't working properly, or their license verification servers blortch out, your CPU won't run (or analogously to those wimpy spare automobile tires, your CPU will run only 1 core with no hyperthreading)?
This is a mal-prescription for your system reliability.
"(Intel didn't said)...but maintained silence on its full capabilities, said nothing about what processors it would work with, and even suggested it might not be related to a real product."
So hell the how can the Linux devs review this to include it? Will the source code have all the details in the Git headers/comments?
Re:
> But if you are giving it away with the GPL why not just enable it in the first place?
and
> So hell the how can the Linux devs review this to include it? Will the source code have all the details in the Git headers/comments?
I strongly suspect the kernel code is nothing more than a channel from user-space to the hardware, which will be used to send encrypted keys that will have as-yet undefined purposes (but likely, as suggested, enabling features in return for extra £, DRM etc.).
Given the prime movers of this (Intel, IBM), I also strongly suspect it is intended to be of more benefit to corporate users than us proles.
I don't know how Intel is going to battle the consumer's brain when it comes to marketing these. Whilst entirely the same as the hardware based restrictions in CPUs, consumers are just going to to think of this as a nickel and dime attempt. We have a tough time selling newer routers and other pieces of equipment that do the same, to the man on the Clapham Omnibus it feels like a scam and that paying more should mean more hardware.
I will however look forward to it, if they price the additional features correctly and if you can pick and choose what you need.
Since Intel has bought Altera years ago I was waiting for an integration of a FPGA with a modern Intel CPU. My speculation is that Intel has combined a modern variant of its Aria FPGAs with a Xeon CPU using a cache coherent interface and that will allow all sorts of software defined network accelarators to be defined during runtime of the SOC.
While switchable features have been a feature of computing from the earliest days, this represents a wholly new development moving from purchase of hardware to purchase of a license. The ability to purchase core usage seems like a good idea (especially with Oracle licensing) that you could upgrade or downgrade as needs change, but there are regressive aspects.
Intel will further be able to experiment in production with features like transactional memory that does not necessarily work.. pressuring vendors to add support for features that could be enabled to break AMD or emulation.
Licensing could be used to force customers to upgrade, by expiring licenses or changing to a subscription model that borks old kit after a period of time.
A 8-core CPU discounted as a 4-core seems like a good way to reduce end-user cost, but could be used to manipulate the market by field-upgrading Intel processors whenever AMD launches a new product.
Doesn’t seem like a consumer friendly or open initiative – Linux should have nothing to do with it.
I'm used to having to buy "server grade" CPUs for certain features.
It's something we've always had to do, but it usually only meant paying more for the unit.
But subscription based CPU features? So we pay for a server grade CPU (already more expensive) and furthermore must pay a recurring subscription?
What? I know the current "go to" money making method for bigger businesses these days is subscriptions.. but this one is just nuts.
Also, if it is going to be like this how will they handle cituations like dedicated server providers? You rent the server, then must pay a fee to enable VT-d on the CPU?
I really hope I'm misunderstanding Intels intentions here and I'm all wrong really!
PS: Sorry for my poor grammar.. I suck at typing on smartphone keyboards...
I reckon the most likely thing is that there’s been a small area of silicon in i9s reserved for FPGA, (Altera IP), and they’ve just not mentioned it to anyone before.
Now they enable the driver in Linux, and they can sell Intel hardware accelerator IP, effectively bundled as “software”.
The other option is to allow their customers to use it *as* a small internal FPGA, without buying separate boards. I’m sure that the investment banking HFT market would *lurve* that (since they accelerate loads of trading within FPGAs) as would a bunch of academia. But…that just cannibalises Intel’s Altera FPGA market, so I don’t see what’s in it for them.
I go for Option A. Intel gets to be a monopoly “software” seller of a very specific type of software.
As noted above, devices shipping with disabled hardware is not new. At least one of the major mainframe manufacturers had a wonderful upgrade that doubled the storage that was performed by an engineer coming out, flipping a hidden switch and turning the 2nd drive head on.
Also, don’t Tesla’s ship with the Autopilot hardware installed, but only enabled when you pay an extra fee?
I'll stick with AMD processors, thank you very much. I've very much soured on Intel after their most recent spate of CPU design flaws (Spectre et. al.), and am certainly not interested in buying processors with "mystery features" baked in instead of useful functionality out of the chip real-estate I'm already paying for.
To me, it sounds more like a lever for shutting off features, and that worries me no little bit. It is bad enough that we have systems vulnerable to a variety of infections and attacks that can corrupt data - but now we're talking about the possibility of attacks completely disabling your physical hardware through software vectors.
But I’m not sure that the consumer-grade ones would have that in? It’s not really in Intels interest to sell you over-provisioned chips that you’re never going to pay to upgrade.
I suspect this is more for server-grade stuff, and provisioned with *software-defined silicon* FPGA-style. Then Intel announce an “Intel Firmware Store”, you buy some firmware IPs with hardware keys, and on Tuesdays and Wednesdays you have some Deep Packet Inspection accelerators running on your pizza boxes when you need it, while on Thursdays and Fridays you have TensorFlow accelerators when you need it.
And yes, these are potential attack-vectors, but no more than any of the *other* microcode firmware updates. Which is both Intel and AMD.
Hello? Developer here. AMD 5900x is just barely enough to do the job, so I'm planning to upgrade in a couple years when the supply crunch tanks and parts flood the market, driving prices down. The next box will be server-grade; I'm out of upgrade options in my current kit regime.
That doesn’t necessarily disagree with my point.
The *assumption* being made on these pages is that people will be sold a software-crippled part, for commercial reasons. This might be right, but there’s alternative option which would be commercially and technically much smarter for Intel.
What if you could buy a CPU that is *customer software-selectable* between being optimised for machine-learning type applications (TensorFlow), brute-force packet categorisation, or science-like number-crunching. Not because the silicon was disabled, but because the silicon was being software reconfigured?
Why did Intel buy an FPGA manufacturer? It seems odd for a large multi billion acquisition, with so far just merging the sales channels, but zero technical strategy. Maybe they are just too dumb to see any technical innovations at all, or maybe they can. Worth waiting a few months to see, I think.
Test equipment is expensive because the market is relatively small.
I am not talking about multimeters or low end oscilloscopes here but rather the higher end; kit that can be used to test complex circuitry such as PCIe systems and really fast DDRx memories (and prove compliance to standards).
I purchased a PCIe analyser for a company some years ago where we did not require anything beyond Gen 1 capability and it was £20k cheaper than paying for the full Gen 2 capability. It could be upgraded as and when that was required.
The equipment I might need may have no requirement (today) for the (for example) DDRx compliance suite or PCIe gen 4 testing. That being the case, I don't have to pay for it even though the equipment may be fully capable of doing the tests in the hardware (Tek front ends have been 63GHz capable for decades).
That can make sense for the test equipment manufacturers and for me as I am not paying for a capability I don't need and the equipment manufacturer does not have to maintain extra physical hardware.
A licence update can unlock that functionality and it has always been a one time install, not a subscription. A lot of test equipment, although networked, does not connect beyond the local network so the 'call home for a server check' is simply not feasible. There are companies working with classified equipment and the network policies prohibit such outside access; the test equipment manufacturers know this.
Another reason for prohibiting / restricting network access is that software updates are highly undesirable in many circumstances.
Other areas that can be unlocked are things like protocol analysers (SPI, I2C, UARTs and the like) although there is no reason (usually) I cannot make my own as these pieces of kit are usually user programmable but it can make sense to simply buy the licence which may be a licence key or a slot in key (sorta like a dongle but much smaller).
There are areas where this sort of thing makes sense.
Every comment appears to be about feature selection.
Intel has WCS - https://www.liquisearch.com/microcode/writable_control_stores
It might be they are coy about this because they are using this feature internally to create their application oriented chips sets, like the blockchain mining processor.
Just saying.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
