"Intel Software Defined Silicon" Coming To Linux For Activating Extra Licensed Hardware Features

[bridgman]

AMD is a saint despite setting exorbitant prices for Ryzen 5000 CPUs and RDNA 2.0 GPUs.

A 128 bit bus, very small die RX 6600 XT for 600 euros? Any time of the week.

Just curious, what conversion mechanism do you use to get 600 euros from a 379 USD MSRP ?

The best guess i could come up with was the old "times two and add thirty" factor, but I don't know why you would apply it here other than for dramatic effect.

I hope I don't need to explain again that we aren't the ones pocketing the extra money when the supply chain marks up products.

[Spacefish]

AMD could easily implement such a thing on their PSP today..
You can even "vendor lock" a EPYC CPU today, and it´s commonly done.
There is a way to instruct the PSP to only load any firmware (bios image) if the signature is correct, by burning some efuses in the CPU via special commands.
As a mainboard vendor you can specify which signatures are correct / your bios image can burn those fuses on the first power-on.

Once you put a EPYC CPU into an HPE Board for example, it´s vendor locked to HPE forever and can´t be used on any other board.

The idea behind this feature: Block firmware based malware / persistent malware which modified firmware to stay persistent.

But basically it leads to vendor lock-in today / making the CPU unusable on any other brands board.

Probably at the factory speed grades / core disabling is done via a very similar mechanism, probably backed by EFUSES which are read / checked by the BootROM on the CPU.

[yump]

The people who are angry about this are ignorant of economics and of the economics of microchips, especially.

It costs like $100 million to make the first CPU, and like $100 to make the second.

Making 10,000 of CPU A and 10,000 of CPU B costs almost 2x as much as making 20,000 of CPU A. So what you do is that you make the smallest number of designs you can get away with, and then mix, match, and gimp to create products people are willing to pay for at a large variety of price points. If I'm counting correctly, AMD's entire current generation CPU lineup is 4 designs. 5 if you count fake 5000-series mobile parts as current generation, and maybe 6 if non-pro Threadripper has a bespoke I/O die.

And price discrimination is not bad for value-focused buyers (which includes me, with my 8 year old i5 and $150 phone). Instead of paying something like ( $100 million / num_CPUs + $100 ), which would be your share of the production cost if every chip was fully enabled, you can pay something like $100.

Intel has been operating this way for decades, and there's nothing objectionable about it. The only thing they want to do different is let people change their minds without physically replacing the processor.

[AmericanLocomotive]

The people who are angry about this are ignorant of economics and of the economics of microchips, especially.

It costs like $100 million to make the first CPU, and like $100 to make the second.

Making 10,000 of CPU A and 10,000 of CPU B costs almost 2x as much as making 20,000 of CPU A. So what you do is that you make the smallest number of designs you can get away with, and then mix, match, and gimp to create products people are willing to pay for at a large variety of price points. If I'm counting correctly, AMD's entire current generation CPU lineup is 4 designs. 5 if you count fake 5000-series mobile parts as current generation, and maybe 6 if non-pro Threadripper has a bespoke I/O die.

And price discrimination is not bad for value-focused buyers (which includes me, with my 8 year old i5 and $150 phone). Instead of paying something like ( $100 million / num_CPUs + $100 ), which would be your share of the production cost if every chip was fully enabled, you can pay something like $100.

Intel has been operating this way for decades, and there's nothing objectionable about it. The only thing they want to do different is let people change their minds without physically replacing the processor.

The issue is that a manufacturer is never going to sell fully functional dies at a loss. This "Software Defined Silicon" scheme requires your chip to have fully functional hardware to work. So every component on that die needs to work, at the maximum possible binned speed. This is opposed to the traditional semi-conductor model where your "cut down" parts are usually made with semi-defective dies. These dies would otherwise need to be thrown away, so selling them at all means you save a ton of money as a manufacturer. You can disable defective cores & cache, bin them at lower speeds and higher TDP, etc.... So it doesn't matter if these cheap, low-end chips are super popular, because you're basically reclaiming what would otherwise be lost product. Once yield improves and you stop getting so many defective chips, these low end salvaged chips become less lucrative and at that point the manufacturer will usually spin-up a proper cut-down die to save on manufacturing costs.

This Intel scheme inherently requires fully functional hardware to work. There is no way Intel is going to give fully functional Xeon dies away for cheap, losing their shirt on every chip sold. No. They're going to increase their "base" prices to protect their margins, then charge even more than they did before if you want to fully enable your chip.

The whole thing is a farce, just like other examples in the technology industry. A notable example in semi-recent times were the Flir E-series of thermal imaging cameras. They had the E4, E6 and E8 - the differentiating feature being the resolution. The E4 being 80x60 and the E8 being 320x240. The E8 cost $3,000 while the E4 cost $1,000. Turns out both cameras had the exact same hardware, and the only thing that was different was a config file.

[drakonas777]

The amount of defective dies is far too low to satisfy the demand for lower tier SKUs, especially when lithography is matured enough. The majority of these dies could easily be fully unlockable.

[oiaohm]

The issue is that a manufacturer is never going to sell fully functional dies at a loss. This "Software Defined Silicon" scheme requires your chip to have fully functional hardware to work. So every component on that die needs to work, at the maximum possible binned speed. This is opposed to the traditional semi-conductor model where your "cut down" parts are usually made with semi-defective dies. These dies would otherwise need to be thrown away, so selling them at all means you save a ton of money as a manufacturer. You can disable defective cores & cache, bin them at lower speeds and higher TDP, etc.... So it doesn't matter if these cheap, low-end chips are super popular, because you're basically reclaiming what would otherwise be lost product.

This is not true. https://premiumbuilds.com/knowledge-...at-is-binning/
A component doesn’t even need to be faulty for binning to happen. Sometimes mid-range components are in higher demand than high-end ones. The high-end models might be binned down to mid-range models to meet the new demand.

The reality here with Silicon Lottery is fully functional dies do get sold cut down to meet demand. AmericanLocomotive the problem why company would do this is simple problem of warehouse space. Yes a high end parts have not had enough orders so the allocated warehouse space for high end parts is filled so now all the chips in a batch are going to be binned down no matter if they are fully functional or not.

Remember parts sitting in a warehouse don't sit there for free. So it can in fact be more profitable to down bin from a high end part to min-range part because the cost of warehouse storage is going to cross the difference.

Once yield improves and you stop getting so many defective chips, these low end salvaged chips become less lucrative and at that point the manufacturer will usually spin-up a proper cut-down die to save on manufacturing costs.

True but that depends if you are getting the high end sales in the first place. Yes intel has to make high end silicon to have something to self to benchmark against amd epyc and other vendors even if they are not selling it. This software defined binning by intel does not block defective parts of the silicon being disabled per chip.

https://en.wikipedia.org/wiki/Intel_Upgrade_Service this also came about when AMD was doing well. Yes AMD doing well historically results in Intel high end silicon not selling well. This result in high end silicon backing up in warehouses while the mid and low end versions of that silicon sell out in cost sensitive builds.

Everything is old is new again. Yes I do suspect the unlocked chips again will end up cheaper than the Intel Software Defined Silicon fully unlocked. Remember its the same problem as 2010.

1) Intel production is not yielding enough failed chips to truly down grade.
2) Intel has a backlog of high end chips as in 2010 that people are not going to pay the price on.(this is what going to lead to the costing problem again)
3) AMD is highly competitive so releasing chips with features fully cut off may also result in less sales.
4) Intel is stuck currently in the cost sensitive market.

Reality here Intel here is attempting to have it cake and eat it too. Yes down bin a high end chip to mid/low end to fill their orders for mid to low and after that chip is installed in system sell it direct to user at the price difference between high and low. This is tempting when your high end chips are not moving anyhow based on the idea if you can get them in the door you can possible make a sale. Think about it person has bought the motherboard they have the cpu installed if they can upgraded without having to change anything they just might do that instead of going to the competition.

Of course just like the old Intel Upgrade Service we can expect this Software Defined Silicon to split the low/mid end in form of shattered mess. As in the chips that were binned down because they were defective so have no Software defined upgrade options and chips with software defined upgrade options.

Also there are reasons why you may want to change the binning of a CPU on a fly. Take the EPYC chips that only have 8 core enabled out of a possible 64. This is done allow higher clock-speed with the same cooling. You do have the case where particular workload perform better when the CPU is binned a particular way.

So time will tell how Intel uses this.

[tildearrow]

And fairies, right. Only it doesn't explain the exorbitant MSRP AMD has set for the RX 6600 XT ($380), die size 237 mm², and a soon to be released XT-less version. This is basically a 2021 alternative to RX 480, die size 232 mm², which was released for $230. Only 65% more expensive.

Must be scalpers again who made AMD set this MSRP. Right.

Again you've just proven that for AMD fans the company is a second coming of Christ no matter what they do or how much they charge. And I've seen more than enough bug reports and issues for their open source drivers, actually a ton more than I've ever had with NVIDIA proprietary "crap".

How desperate are you to seek conflict? I was referring to the $600 price for a 6600 XT you found, and then you suddenly switch topics/subjects.
I have not called AMD a "second coming of Christ". I even am aware of its flaws, so don't start.

Even Bridgman (an AMD employee!) has joined the debate and told you the truth, but you cannot understand.

[TemplarGR]

The 6600 XT does appear to list for around 600$ in stores.... That is what i read.

[mdedetrich]

And fairies, right. Only it doesn't explain the exorbitant MSRP AMD has set for the RX 6600 XT ($380), die size 237 mm², and a soon to be released XT-less version. This is basically a 2021 alternative to RX 480, die size 232 mm², which was released for $230. Only 65% more expensive.

Must be scalpers again who made AMD set this MSRP. Right.

Again you've just proven that for AMD fans the company is a second coming of Christ no matter what they do or how much they charge. And I've seen more than enough bug reports and issues for their open source drivers, actually a ton more than I've ever had with NVIDIA proprietary "crap".

Uh you can't blame AMD for the current supply shortage, they don't even manufacture their own silicon. AMD is constrained by TSMC and wafer amounts are agreed upon years in advance when they are delivered.

This is ontop of the fact that TSMC is also the top sillicon provider for other companies (Apple, Playstation, Microsoft, Intel with their new discrete GPU) etc etc so even if AMD could get more chips from TSMC, they can't because everyone else is fighting over it.

There has already been a lot of leaks and apart from demand (which AMD cannot do anything about) almost all of the markup over MSRP is actually due to price gouging by distributors.

[ssokolow]

If Intel instead spun this as "Here's a free and open-source tool to check *what* was defective in your low-end chip, and we'll sell you a key to let you unlock whatever was pref'd off purely to match the bin it went into, but, like with overclocking, we make no guarantees about what the inspector tool will tell you about any specific unit", then I'd be less against this.

...of course, then it wouldn't have the marketing angle they want out of it.