On Mon, Feb 05, 2024 at 12:30:55PM -0500, D. Hugh Redelmeier via talk wrote:
High-end processor fabrication is now a specialized business. Only Samsung, TSMC, and Intel have the capability as far as I know. There are stories out of China suggesting that they are trying really hard to get there (SMIC, for example).
Most ARM processor designs are done by ARM in-house and are licensed at rates that seem to be good enough for the market. Interestingly, ARM charges considerably more for you to use their architecture without using one of their designs. The few companies with that second license include Apple (who got their license by being a fo-founder of ARM as we know it) and Qualcomm (I think).
Well there is a decent list of them. Ampere making server CPUs, Broadcom, Marvell (including Cavium), Nvidia (Although most of their chips use stock cores), Samsung (Seems they have moved to only stock cores on newer chips though). Certainly not very many really making custom cores anymore.
This second license requires conviction that your processor design team can do better than an off-the-shelf design for your application. A few companies have tried and failed (eg. AMD). Of the shelf ARM designs have not been good enough to break into the datacentre but a couple of non-ARM designs are doing OK (Amazon's Graviton, for example).
Note: users of the second license could just use RISC-V instead, for free.
In my opinion, the software support for RISC-V is mature. Linux supports it as a first class architecture and that's all I need. Android supports it or soon will (I don't remember which). Microsoft does not support it, as far as I know.
Well they certainly are starting to look interesting. Microsoft support doesn't seem to mean much, just look at their level of ARM support so far. And their support for powerpc, alpha, mips and itanium didn't seem to offer anything of use to those architectures.
Yes. ARM has taken a lot longer than I expected.
The NetWinder was a credible machine for desktop or server Linux something like 25 years ago. It failed for several reasons but one was that there was no ARM implementation that got near x86 performance.
Remember: that same fact killed off all the RISC desktops (HP, MIPS, SPARC, Power) and 68k, NS32032, etc.
Alpha was faster than x86 for a while, but DEC internal infighting and pricing strategies certainly didn't help it sell. Can't have alpha hurt vax sales, much better to let everyone else do it instead.
NVidia is hamstrung selling to PRC. Long term, this will likely hurt the US.
NVidia tried to buy ARM. That would have even further aided RISC-V because NVidia competitors would be a bit wary of buying ARM designs.
Right. The ARM China thing was crazy (we don't understand the importance of corporate seals in China). It has been resolved (that's an old article).
RISC-V can be used by both sides of this divide. ARM cannot.
China rightly views the capability of access to fast processor as strategically important. You can be sure that they have more money than venture capitalists. They will make it happen.
I wonder how LoongArch is doing performance wise.
The US is trying to block PRC and they have somewhat clumsy defence in depth. They block AMSL from shipping recent lithographic systems to PRC. They block ARM. They are blocking chemicals for lithography from Japanese companies. This buys time but it won't work long-term.
This will prime the RISC-V pump. But there are rumblings in the US of kind of black-listing RISC V which is really sad.
Actually, it cuts both ways.
Certainly China can cut the US (and others) off just like the US can try to cut off China. -- Len Sorensen