[GTALUG] Linus blaming Intel for lack of ECC in consumer systems

[Lennart Sorensen]

On Sun, Jan 17, 2021 at 10:26:07AM -0500, D. Hugh Redelmeier via talk wrote:
> If the vendor chooses to segment their market by leaving out zero-cost 
> features for the lower priced product, that's a demonstration of monopoly 
> power in the market.
> 
> Intel has done that a lot.

Oh boy have they ever. :)

> There is a plausible excuse in some cases.  If all the cost is in 
> engineering and not in manufacturing, it may make sense to only recoup the 
> engineering costs from those who need it.
> 
> AMD, when it has been the underdog, has often thrown a lot of features in 
> without extra charge.  This has sometimes appeared to drive Intel to move 
> broadly provide the feature (i.e. when Intel appeared to be threatened 
> by competition with AMD).
> 
> An example that irked me was virtualization hardware.  For the first few 
> years, Intel only provided the feature on an odd subset of chips that were 
> expensive.  Whenever you were considering a system, you had to consult the 
> spec sheet for the CPU.  AMD provided it in all new processors after a 
> they had engineered it.

I didn't help how many companies left the feature disabled in the
BIOS and in some cases didn't even provide an option to enable it.
Supposedly because it might cause someo compatibility issues, although
I don't believe it ever has for anyone.

> Other examples that come to mind:
> 
> - Instructions to accelerate cryptography (eg. AES).
> 
> - PCI lanes, as Nicholas mentioned.

Hmm, I can't recall cases of this, but I may just have missed it.

> - physical address bits leaving the CPU bus
> 
> - 64-bit mode on x86

I have seen some intel CPUs that the official specs say don't have 64
bit support, but every actual chip I have ever seen had 64 bit support.
So is the chip manufactured wrong because they didn't turn it off or
was the spec writen wrong?  Can you use it and trust the chip you buy
next month will still have it or not?

> Interestingly, there are examples of the opposite behaviour that also
> indicate to me monopoly power.
> 
> - Intel engineered a fairly powerful and intricate MMU for the 80286.
>   They essentially forced all customers to pay for it, even though
>   almost nobody used it during the service life of the processors
>   bought in the first years.

It's design flaws hurt it a lot.  Not being able to leave proteceted
mode once you entered was apparently a big problem for operating systems.

>   Motorola spent very roughly the same amount of silicon to implement
>   a much more elegant 32-bit processor (the mc68000).

Well they had the benefit of not being backwards compatible with
anything else.  Starting from scratch can be a big advantage.

> - Intel, a generation later, made customers buy a 32-bit processor
>   (the 80386) when all that they'd use was the 16-bit subset.

At least the 386 had a much better protected mode design, although ram
was probably too expensive still for most people to have much use for
the 386's features.  And it was able to help out dos in 16 bit mode
using the MMU to allow TSRs and such to be loaded outside 640k and was
able to pretend extra ram was EMS or XMS which was likely much cheaper
than the special EMS memory boards people had been using before.

> In both those cases, the evidence is muddied because it was
> Microsoft's monopoly behaviour that left hardware capability unused.
> 
> Famously, it was the latent power of the 80386 which incentivized
> Linus to produce Linux.
> 
> In general, UNIX could and did exploit both the MMU of the 286 and the
> 32-bit architecture of the 386 long before Microsoft software.

Yes unix on a 286 didn't have to care about going back to real mode so
it was able to use it.  Microsoft Xenix 286 came out in 1984 which is
only 2 years after intel introduced the 286 to the market.  That's pretty
quick support from Microsoft.

> ECC is generally provided by one extra bit per byte of memory.  So the 
> cost premium should be no more than a factor of 9/8.  It should be even 
> less since the cost of the board should not increase by that much, only 
> the RAM.

Yeah 10% extra at most would seem reasonable.

> I imagine that they are incentivized into this behaviour by their market 
> position.
> 
> There is a pattern of underdogs changing if they become dominant.
> 
> It's not driven by morality, it's driven by game theory.
> 
> That's why anti-trust enforcement is so important for healthy free 
> markets.  Something that has faded away in the last fifty years.

-- 
Len Sorensen