Richard Cownie (tich@pobox.com) on 10/2/09 wrote:
---------------------------
>David Kanter (dkanter@realworldtech.com) on 10/2/09 wrote:
>---------------------------
>
>>The first issue is why do you need your own core? There are reasons I can imagine,
>>but in most cases it's probably not necessary. Especially for ARM which has a lot of cores floating around there.
>
>Historically, it seems to have been a question of building
>a faster core than any of the available ones (and I
>suppose by building it yourself you also ensure that
>ARM plc doesn't then sell it to your competitors).
>Thus we got StrongARM/XScale, and the 1.2GHz Marvell
>Sheeva.
>
>But I wouldn't rule out the "fun" explanation: every
>writer wants to write a novel; every software developer
>wants to write a compiler; every digital hardware
>designer wants to design a cpu core ... those factors
>of personal motivation can swing a close decision.

There is another possible reason: in the early days when ARM was small, having an architectural license would hedge against the possibility of ARM not developing suitable cores for their partners, or worse, being bought by a competitor and wound up. For some companies having a safety net might have been important.

>>I suspect the cost is closer to 20-30M. The real question is how much money is
>>the arch license compared to taping out the bloody chip.
>>
>>I think the rumor is that the high-end ARM designs are costing around $100-300M.
>>That's a lot of money...against which $10-30M may not be a big factor.
>
>Indeed. The design and verification effort for a
>superscalar OoO core is huge, and $30M for a license
>isn't big in relation to that, and helps to reduce some
>of the risks.

Developing a new architecture is as expensive as designing a CPU, and even large companies will not have enough expertise. The software and documentation cost is incredible, especially if you count everything involved, including having to redo things several times for research and evaluation prototypes and then the final version. Designing and implementing the ARM assembly syntax took at least a man year, and that was just the dot on the i.

>Also I don't know how the architecture license affects
>the per-chip royalty fees on the final product. If you're
>planning to make say 20M chips, and having your own core
>saves you $0.50 in royalties on each chip, then that's
>$10M saved. All those figures are entirely speculative.

I don't know whether an architectural license has any royalties at all, but for core licenses the royalties vary depending on the core, the volume and the type of license. You can get a cheap license with high royalties, or a more expensive license with lower royalties.

Wilco