This makes no sense at all. I'm trying to figure out the upside for ARM other than they earn more money on their cores. It would certainly put a large damper on the rock star chip designers since they couldn't sell their firms as easily.
The logic really doesn't make sense. If you try to use a design made by someone else with an architecture licence and you have an architecture licence, ARM has to approve it?
I guess they're trying to protect themselves as the sole seller of ARM designs, but even then Qualcomm didn't buy a design, they bought the company. ARM seems to be going overboard here.
And all while RISC-V is nipping at their heels. It's like they want to make it really clear how troublesome licencing proprietary arches is.
If you want ARM, you play by ARM's rules. Or you don't get to use the license. Either ARM is about to learn a valuable lesson about knowledge wanting to be free, or Qualcomm is the about to get beaten with the lawsuit bat.
I think it's ludicrous we even bother with this daftness, but seeing as everything has to revolve around money, it is what it is.
I just can't see Qualcomm's lawyers flubbing this one. They are an extremely aggressive company when it comes to IP and the law. It's not like they don't have top notch IP lawyers. If ARM persists in this stupidity my money is on Qualcomm tearing them a new one.
ARM seems to be the AWS of hardware IP cores. It is very much the case that if someone is basing a design off their base designs, that comes with granting ARM a thumb on the scale, and a place at the negotiating table if two of their (ARM's) customers engage in an M&A. Qualcomm licensed stuff for handsets. Nuvia licensed stuff for servers. The new composite entity is still constrained by both contracts, which requires an IP firewall.
If, on the other hand, Qualcomm wanted Nuvia due to their server core IP, and does not intend to merely consume that company's blackboxed output, then they need to renegotiate terms with ARM, either acquiring a second license for Qualcomm proper, or paying whatever fees are customary for a transferrable license.
In the end, it's unsexy paperwork.
But I get it though. If you dig into semiconductor/hardware lit, like things on logic locking, that entire industry of fabless semi-conductor fabrication has a Jesus Nut in terms of license enforcement. The prospect of physical design/mask theft or unauthorized design reproduction in a way substantially different, and much more tangible than software is. It really annoys me that someone somewhere describes this as "piracy", because it isn't really, it's straight up breach of contract/unauthorized use of physical design masks. Like getting the printing plates from a mint.
Frankly, I've been skeptical of the entire thing, because it does basically rely on manufactured scarcity/info asymmetry created and maintained through legal means, but no one exactly asked either for my input, or seems open to being told not to try to monetize something you can only collect on by essentially parking lawyers at strategic points in the business lifecycle.
From a theoretical system point of view it works. From an execution point of view... Not so much.
Also see a multi trillion dollar company that has to deal with onerous terms for cellular chips from Qualcomm because there is no viable alternative.
In fairness, ARM's ownership of the ARM ISA couldn't be an advantage very long. Much like x86, ARM is just old nowadays. Unless you're supplanting it with a half-dozen ISA extensions and additional hardware, it's a hard sell for anything other than embedded/minimalist applications. Which makes it all the more odd that ARM is picking legal fights instead of making it easier for customers to adopt their (proprietary!!!) instruction set. My untrained, unsolicited opinion is that ARM is just pivoting to support their pre-existing customers and lock-down the money makers.
I disagree: ARM64 is from 2011, it's not that old and it was a significant ISA makeover.
Abandoning all that for a fixed-length 32 bit opcode 64 bit instruction set with just barely better code density than original ARM (and dropping the stuff that made sense in 1985 but doesn't today, such as predication everywhere and load/store multiple) just looked like lunacy to me in 2012 when I read the Aarch64 manual, and looked even worse when RISC-V came out with a 64 bit instruction set using the same basic idea as Thumb2.
If you're doing very wide instruction decode then two instruction lengths is a little harder to handle then one, but not enough to matter, at least at 8- or 16-wide. You can EITHER 1) put a 2:1 32 bit mux in front of half of the decoders if you don't want much extra hardware, OR 2) build 50% more decoders and put bigger 2:1 muxes on the outputs of the decoders. The latter is slightly faster because you have to examine 2 bits out of each 16 bit parcel in the instruction stream to decide which mux input to choose, and option 2) lets you do that in parallel with the decoding instead of having to do it before the decoders.
It's just completely incomparable to doing wide decode on x86_64 where you have to deal with 1-15 byte instructions with an arbitrary number of prefix bytes and several opcode bytes to decode before you know the length of an instruction, and therefore where the next instruction starts. If the first instruction starts at offset 0 then the second instruction can start anywhere from offset 1 to 16, the third instruction can start anywhere from offset 2 to 31, the fourth instruction anywhere from offset 3 to 46 etc.
Focussing on desktop and servers.. It worked.
In a similar fashion RISC-V has things which makes little sense for "high performance" (16-bit aligned 32bit instructions with the C extension, too simple addressing mode) which shows that their main goal is embedded..
Such as?
>(16-bit aligned 32bit instructions with the C extension)
For virtually no cost (see brucehoult reply to my parent post), it brings greatly increased code density, which means more code fits in L1, or same thing, that you can make it smaller and clock it higher. This actually benefits high performance microarchitectures greatly.
See M1's L1 size for an example of why ARMv9 AArch64 is mind-blowingly bad. ARM's literally lost its marbles.
>(too simple addressing mode)
More complex addressing modes were evaluated and discarded, as they only brought harm, no benefit.
This is a pattern in RISC-V, where decisions were made after careful weighting, rather than some individual's favored way of doing things.
Presumably similarly troublesome to licensing SiFive's proprietary IP?
ARM architecture licences are used when you want to do a largely complete re-implementation of the arch, so the instruction set is really the only issue.
Anyone who wants to can design and sell their own RISC-V implementation (as Nuvia and Qualcomm paid ARM [1] to do with the ARM ISA) without asking anyone for permission, without telling anyone if they don't want to, and without paying anything to anyone.
[1] just, apparently, not enough
When the iTunes Music Store was introduced back in 2003, not only was it a better alternative because it was more convenient. But you also didn’t risk being sued.
Before anyone chimes in, “by pirating I can get DRM free music”, Apple has sold DRM free music on iTunes since 2008.
Also, it’s not like Qualcomm has any alternatives to ARM. No the entire mobile market is not going to move to RISC-V.
Maybe the logic is different for ARM. Maybe they really badly negotiated licenses with Qualcomm and they need to work around that, or they want to protect Cortex core licensing from some hypothetical Snapdragon with a good CPU. However, by doing that you are undermining your own reputation. Nobody wants to be known as Mr. Litigious.
[0] Or at least, Steve Jobs said so in his open letter on DRM. They had telemetry in iTunes and knew how full iPods were versus how many songs people bought.
His exact words on his “Thoughts on Music” letter were.
https://macdailynews.com/2007/02/06/apple_ceo_steve_jobs_pos...
> Today’s most popular iPod holds 1000 songs, and research tells us that the average iPod is nearly full. This means that only 22 out of 1000 songs, or under 3% of the music on the average iPod, is purchased from the iTunes store and protected with a DRM. The remaining 97% of the music is unprotected
I wouldn't want to bet too much on that. If anything, the effective stranglehold on phone SoCs that Qualcomm has gives them more power on this issue. Qualcomm is pretty much the only vendor powerful enough that they could tell Google that continued performance improvements to phone SoC performance would be contingent on switching ISA to RISC-V. Much of the hard work in getting Android to support RISC-V has already been done by Alibaba.
Samsung isn't big enough to force a switch on their own, especially not it if would result in a mixed architecture product line. But if Qualcomm wanted to?
I don't think such a switch is _likely_, but it is _possible_, and ARM screwing around with one of its largest and highest-profile customers (especially one known for aggressive legal action and a very short-term notion of compatibility) is a good way to make such a switch look more appealing.
Really, it seems to me that ARM really has only one possible good outcome from this: a settlement where Qualcomm pays them a bit of money. If they actually win, that maybe benefits ARM's in-house design teams and TLA licensing business in the short term, but it basically guarantees that ALA licensees in the server space look at switching ISAs.
Google sold only 1.2 million Pixel phones last quarter in the US
https://9to5google.com/2022/05/19/pixel-6-us-market-q1-2022/
Compared to Apple’s 20 million
While Google doesn’t break out sales of Pixels. The best estimate I could find is around 2.5 million a quarter globally.
https://9to5google.com/2022/05/11/the-pixel-6-sold-more-than...
Or about the number that Apple sells in 3 days.
The US market is 3rd largest behind India and China or 4th if you count the EU as one market.
But in the US, iOS has about a 60% market share and it dominates the high end. People who are buying cheap $261 or less Android phones aren’t looking for performance.
https://www.statista.com/statistics/951537/worldwide-average...
Not to mention that Google is non existent in China - the worlds largest market.
The market for high performance Android phones in the US is particularly small. Android phones are so far behind in performance now that even the low end iPhone SE benchmarks better than high end Android phones.
This is the same reason that no chip maker even tries to make processors that are as performant for Android Watches as the Apple Watch. Android devices have been a commoditized race to the bottom for a decade.
This entire lawsuit, one could argue, centers around the belief by some vendors (Qualcomm, Nuvia, etc) that in-house designs - even those made by small teams - can significantly outperform the designs that ARM produces and makes available via TLA licensing, and that (implicitly) the pace of ARM's own development is not fast enough or reliable enough to be competitive in the markets these vendors want to operate in.
The last decade is littered with OEMs trying to produce “high end” Android phones. Not enough people would buy them to make it a sustainable business.
I’m not making a value judgment on them. Most people are fine with a sub $100 unsubsidized Android phone like the ones selling in developing countries and in the US.
Each Arm ALA (Architecture License Agreement) grants the licensee only the right to design CPUs for a specific purpose.
If the licensee has a license for a server CPU and it attempts to sell the designed CPU as a laptop CPU, Arm claims that the original license is terminated and everything designed under that license must be destroyed, unless a new license is negotiated with Arm, allowing the licensee to design and sell laptop CPUs.
In the filing, Arm says: "Qualcomm’s Arm licenses do not cover products based on or incorporating Arm-based technologies developed by third parties under different Arm licenses, such as the now-terminated Nuvia licenses".