SoftBank's Sale of Arm to Nvidia Collapses, Arm to IPO
reuters.com
reuters.com
It didn't make sense how ARM was worth $22bn in 2016 on LSE, that has a PE100+ when I cant see how it could double its Net profit in the next 10 years, with a saner valuation of P/E 50. It make even less sense to me when Softbank bought it for $32bn. Didn't make any sense when Nvidia want it for $38.5bn, and later at $50bn+.
Now we are back to square one, how many investors, in the open market are willing to buy ARM's stock at a valuation of, let say the same $32bn as Softbank initially paid? Roughly P/E 100, assuming they managed to grow ARM business from 2018 at ~$220M operating income to $300M in 2021. I much rather bet on Intel if I had the money.
And of course many on HN will think ARM is worthless because the future is going to be the Riscy Silver Bullet.
Disclosures: I do not currently hold any stocks or Mutual Funds, Indexes and other financial instrument.
And generally speaking Softbank ( and ARM ) refuse to comment on anything about ARM China.
But that is a good point, it is a potential risk or at least need to be mindful in its valuation.
https://find-and-update.company-information.service.gov.uk/c...
Latest is 2020/21. Profits down to £7m due to higher R&D and other expenses.
- Where a free ISA allows innovation and experimentation in a way that Arm would not permit.
- Where companies can shave cents off mass market products due to lower licensing fees.
Will RISC-V replace Arm in your smartphone though - seems like a big ask at the moment.
Just because there is a port available doesn’t mean the hardware and ecosystem will follow.
> - Where a free ISA allows innovation and experimentation in a way that Arm would not permit.
I would say that the two major innovations of RISC-V are:
- Composability.
- Extensibility (that also fosters experimentation).
Each ARM version X.Y comes with a license and with a condition: «thou shalt taketh either everything or nothing»; RISC-V has neither a license nor strings attached and is «thou may useth only what thou needeth». Anyone who does not need SIMD or the vector processing can do away with them in the silicon. It saves the die size[0] and costs. This is the composability aspect.
Extensions planned out and/or reserved for future use (e.g. 128-bit extension of RISC-V), and the ability to add custom instructions to experiment with (and possible contribute to the RISC-V forum). This is the extensibility aspect.
Other existing open ISA (MIPS, POWER SuperH J-2? – unsure on the status of the last one) do not offer composability and extensibility capabilities, which leads me to reason that it is the major innovation(s) of RISC-V.
[0] Albeit it is somewhat fraught with the danger of creating a highly fragmented RISC-V CPU market where the software has to accommodate every possible permutation of RISC-V extensions and inject an emulation layer for extensions missing in the die.
Then AArch64 was designed very late, but seems sufficient to have (currently) taken the prize for the top supercomputing architecture.
There is architectural flexibility here, in addition to low power.
ARM bigLITTLE exists to address different efficiencies that are exploited in different runtime states. Is big better than LITTLE? Sometimes yes, sometimes no.
If there is no definitive answer within the heterogenous cores of a single ARM implementation, then how can it be compared to RISC-V?
For the same sized single-core chip (x mm squared), running at the same voltage and same clock speed, on the same manufacturing process node, is there a (significant) difference between ARM and Risk-V?
(I'm assuming that core size, voltages and frequencies are the biggest differences between the big and LITTLE cores. And yes of course this is a funky theoretical question so a theoretical answer is perfectly fine :)
"According to SiFive, it is working on the design of a new processor that will be 50% more powerful, and that it will be able to surpass the Cortex-A78 of ARM."
https://techunwrapped.com/risc-v-starts-to-put-pressure-16-c...
Very many. And they are at the CPU implementation level.
The instruction decoding buffer depth, the register file size, the implementation of the out of order execution, the TLB depth, the data bus width, separate hardware assist engines etc are amongst many other conscious and decisive design decisions that can differentiate separate designs. Even at the same (citing verbatim) «voltage and same clock speed, on the same manufacturing process node» the performance can vary wildy.
The M1 CPU family is a very wide implementation in nearly every aspect, with a very large L1 cache (both, I and D, caches), reasonably large L2 cache, a (reportedly) 380 register wide register file, 512-bit wide data bus (in the M1 Max) plus extras, hence why it outperforms competing ARM64 designs running at similar voltages and frequencies. The same is true for POWER CPU's (i.e. compare POWER9 vs POWER10 designs) – POWER10 is 2.6x faster than POWER9 despite running at similar frequencies – due to a substantial design overhaul.
Holistic system design matters, not the ISA alone.
Which is the problem often with these sort of discussion. ARM of what ( You didn't state its version ) ? RISC-V of what? The ISA in itself? Or the actual implementation in itself. If it is implementation, which implementation? And by difference, what sort of difference, what kind of performance are you specifically looking at? Or you just want a GB5 result? ( One could argue there are compiler differences in GB5 results )
The only undisputed truth would be RISC-V at its bare minimum has smaller instruction than any ARM variation currently on the market. But then its competitor should have been something like Arc, and not ARM.
They may not be as energy efficient as M1 or ARM chips but they certainly show that they are still a competition and invest a lot of money to improve their performance.
The market thinks that in some fixed, relatively near point in the future we will look back at now and say "Intel was already hosed, it just didn't know it".
It may be that the market is currently only expressing interest in the hopes of others that this particular future will come to pass, but that is the only way it will.
AMD's stock can jump like this, because its value is based more closely on its designs.
Intel is not taking that approach. One amazing chip design isn't going to massively change their fortunes, and they are heavily sunk in x86, which is not the chip series the market is most interested in.
One meaningful path forward for Intel is to split in two, precisely so that the x86 design business can be liberated by being fabless, and the fab businesses can get on with making unsexy larger-process chips for everyone in America who needs them.
But Intel seem to be resisting this.
We'll see what the story will be going forward.
ARM has taken out the low-power market that Atom and Pentium classes were aiming at. AMD APUs have long been a favorite of mid-power devices like gaming consoles and are now a desktop enthusiast favorite as well. The i5/i7 class is outmatched in every metric ($/FLOP, W/$, and W/FLOP) by the Ryzen series. IBM and Azure placed their big chip refresh orders with AMD as well, so the hyperscaler market is in danger. The only place Xeons are still superior is the scientific and high computing space, and Threadripper is making inroads there as well.
Intel still has great sales from existing cashflows and OEM agreements, but if you were buying a computer in 2022, why choose Intel? Why would its customers? I think they're going to have to come up with something new and soon.
This isn't really true though. Alder lake performance is better than Ryzen 5x at pretty much all price points. Alder lake's electricity usage is rough, but desktop users don't care. Now ryzen is a year behind intel, and who knows how their next offering will fare, but saying they've outmatched intel in every metric just isn't true.
maybe a couple years ago, at this point Intel is dominating the gaming market, competing in the productivity market despite having fewer cores (i.e. more potent cores), are at least matching Zen3 efficiency (except i9, but you said i5/i7) if not beating (particularly at lower power budgets due to infinity fabric overhead). Ryzen has absolutely lost favor with the enthusiast market due to the significant price increases (in some segments, a 50% increase over last generation) and the refusal to chop prices to remain competitive with intel (justified with "but they can make more money selling Epyc", which is great for them but it doesn't make their products competitive for me as a consumer either).
And AMD has completely abandoned the value market... i3 12100F is listed for $110 at Best Buy (iGPU version is $140), and Intel is in the process of rolling out Celerons for even cheaper rigs (will be nice for HTPC with that iGPU). AMD's cheapest processor has a $300 MSRP and they won't budge on pricing at all given their focus on the server market.
go look at a tech enthusiast forum and nobody is recommending Ryzen anymore. It's overpriced, it no longer has a performance or efficiency advantage (in fact it's rather flopped the other way), and AMD would rather address the server market instead.
this has really shifted in the last 6 months since Alder Lake launched, but really there was a significant niche for recommending 10-series as a budget option given the significant price advantage over AMD. The thing that changed with Alder Lake is that Intel moved back into the performance lead (and matched efficiency) and prices remained the same - leaving them both better and cheaper than AMD. Thing have changed, people don't recommend AMD anymore unless you have a very specific need (like you're buying a 5950X).
Also APUs in particular are rather a flop on desktop. AMD didn't even bother releasing a 7nm APU to the consumer market until... September last year? Up until then it was all 12nm quad-core garbage - and that's GloFo 12nm, so more like Intel 22nm. On top of that, the APUs also had reduced IPC compared to the CPUs of their equivalent architecture (due to less cache). All of the hype around Ryzen was around the non-APU enthusiast chips with no onboard graphics, the APUs were never popular in the desktop segment.
The APUs are nice if you are building a SFF rig with no discrete GPU... but they have a lot of compromises - reduced IPC, PCIe 3.0, and up until September they didn't have any that weren't limited to PCIe 3.0x8 (5000G series moves this to 3.0x16). They have no HDMI 2.1 support, no HDMI VRR support, no AV1 support, and the worst hardware encoder (i.e. Teams videoconferencing quality/etc) on the market. If you really want a gaming-focused SFF rig with iGPU only, they are the best available, but they were never popular, and AMD never really pushed them the way they did the enthusiast chips.
(HEDT is another segment that's been abandoned... AMD still has not released Zen3 processors, it's been 16 months now, and who knows whether AMD will try the same "release it for pro but not for consumer" crap again like they did with the APUs.)
I expect them to be at least $400bn to $600bn market cap by the end of 2030. Which is 2x or 3x of today's price.
Not exactly super high growth, but if you are a money manager with too much cash ( which is often the case today ) it is not a bad investment.
So Apple Silicon is not the same as someone else licensing ARM IP.
Apple may have been a founding member, but they don’t own ARM, and I’m assuming that ARM still wants to make a profit off of its licenses to Apple.
This is getting into fantasy chip designer territory, but what are the chances that Intel makes a hybrid x86-ARM chip?
Nvidia would have expanded the market from their own products to licenses.
> ksec 67 days ago
> The unwritten rule of HN:
> You do not criticise The Rusted Holy Grail and the Riscy Silver Bullet.
RISC-V literally has no "openness" advantage over ARM other than that there's no licensing fee. Apple could make a RISC-V chip and use it in their iPhones and it would be just as locked down as their ARM chips.
Those facts will:
- Disprove once more that story of the market efficiency/price discovery bullshit,
will also confirm that:
- Benjamin Graham, Value Investing and Warren Buffet strategies were made for other times. The times before we allowed that Central Banks hijack our income streams and personal wealth.
Disclosure: As I write this, I also do not currently hold any stocks, Mutual Funds, Indexes and other financial instruments. Did in the past.
> Reflexivity theory states that investors don't base their decisions on reality, but rather on their perceptions of reality instead.
Meme stocks, Covid crash, Tesla stock, the SP500 for the last decade, NFTs. None of it is priced correctly because humans are at heart greedy, fearful, irrational. I think the trick is to figure out how to use that to make money.
I also sort of disagree with Value investing as a bygone era. It is just I have yet to see any one who does value investing actually understand anything about Tech. Because it is not something simple enough like Coca Cola or McDonald, so they are much harder to evaluate.
But Yes, it will be fun to revisit this thread then. :)
The FTC objecting to Nvidia and ARM's deal on an antitrust basis [0] makes Arm toxic to Intel and AMD.
[0] https://www.ftc.gov/news-events/press-releases/2021/12/ftc-s...
I feel Jensen and Nvidia still, very strongly, wanted to do the deal. The deal increased in value to $87bn only because NVDA's stock price increased that much. The real costs to Nvidia did not increase.
ARM is a legitimate competitor to x86, which Nvidia wanted to control.
I get what you’re saying, but in practice, NVDA was essentially issuing x shares to buy ARM. They could do the same thing, sell them, and keep the cash. The rising stock price very much did raise the “real” cost to NVDA, it just wasn’t a simple cash charge.
Put alternatively, NVDA basically said to ARM “you’re worth x% of our company.” While that may be true at one date, a deal struck at the new market cap would be unlikely to come to the same conclusion.
That said, Nvidia was still trying hard to make the deal work, even at that price. Jensen strongly believed the deal would be worth many multiples of the acquisition cost.
The counterpoint ofc being (with Facebook/Meta being a great example), just because a CEO was once a visionary doesn't mean they are infallible.
From the source:
https://nvidianews.nvidia.com/news/nvidia-to-acquire-arm-for... "The number of NVIDIA shares to be issued at closing is 44.3 million" [this is the pre-split amount]
They've always managed to display ambivalence about it. We should be glad they never owned ARM and hopefully never will; they'd have killed it with indifference.
It was such a hot stock in 2014-15. Softbank kinda bought it at the peak. Still, Apple has proven ARM on the desktop is phenomenal, and Ampere has done the same in the datacenter. Both have swiftly made compiler support accelerate dramatically. I would love for my next desktop to be arm-based, and it would be great if it didn’t NEED to be a mac. Broad adoption of ARM would be a huge win for sustainability. I’m bullish on their business, even if their business model isn’t vertically scalable.
I think that RISC V also has a bright future, but its still a bit far out before its a viable competitor. Still there have been some things that look really promising such as what is coming out of SiFive:
https://www.sifive.com/press/sifive-raises-risc-v-performanc...
And Intel just announced a $1B fund for companies building on their Open Foundry RISC-V alliance. https://www.tomshardware.com/news/intel-1b-fund-risc-v
Big things are happening for small instruction sets… watch this space.
Why would you want that? The Intel+AMD duopoly seems like nirvana compared to the Ampere and M1.
In fact, Ampere and M1 are a hint of the most dystopian future of computing I can imagine.
OK I'll bite. What aspect of Ampere is locked down? They are nothing like the bad closed-source hacks that we see in small development boards.
In terms of firmware openness, yeah, you don't get the firmware source, but you don't get to touch the machines physically either so you wouldn't have a chance to run a custom build anyway :D
(While Ampere Altras are supported by upstream EDK2, have fun)
What does being able to buy them have to do with.. anything?
The concept of renting a VM is not new. The introduction of processors that a VM hosting company designs exclusively for use on that service is, and the comments in the thread above seem to imply a fear of this causing some kind of lock-in or whatever. What I'm saying is, this is completely unfounded fear. Graviton instances are very much standard machines and you can run the exact same OS images on there as you can on an Altra or a HoneyComb LX2K or a MACCHIATObin. (Yeah, the OS needs a driver for Amazon's network card, but mainline Linux, FreeBSD and NetBSD have that.)
Oh, we should be so grateful...
Or perhaps be happy with what we got and not sacrifice everything for some very short term 5% benefits?
Take one guess.
Nothing becomes any more closed if you switch the ISA of the CPU used in your VPS to aarch64. The only closed ecosystem here is AWS itself, but what takes power away from you is using non-EC2 services like DynamoDB.
In fact the Arm ecosystem is way more open, because there are SoCs that can run completely blob-free firmware (e.g. from Marvell), which is just not happening with any modern Intel/AMD system.
A firmware-blob is nothing in comparison for the vast majority of cases.
BTW - any SoC designer that uses RISC-V can do the same.
It should not be celebrated. Or we will fix the mistake that is the PC forever.
https://en.wikipedia.org/wiki/Server_Base_System_Architectur...
There's also plenty of ARM hardware out there that really doesn't bother with lockdowns. Raspberry Pi, for example, will happily boot whatever you put on the SD card[2]. There's also plenty of other dev boards and chips that are similarly not locked-down. The reason why people think of ARM as "more locked-down" is because it's more often used in locked-down hardware, but there's no architectural requirement for it to be this way.
The real big problem with ARM is that every chip is very different from one another. SoC designers have a habit of reassigning addresses and IDs on every new chip, with no thought towards backwards compatibility. So even if an ARM chip has no technical protection against third-party operating systems, any Linux kernel you build is unlikely to boot unless the SoC vendor has bothered to upstream support for it. Ironically, Apple is one of the few ARM vendors that actually does maintain backwards compatibility where feasible.
[0] The first admin account is granted possession of a local signing key that lets that user sign kernels. This means that you need to be able to log into that account under macOS in order to install Linux - evil maids need not apply.
[1] Legend has it that if you say "M1" three times in a HackerNews comment, you'll summon the primary developer of Asahi Linux.
[2] You do have to provide a GPU boot blob; no clue if there's a project to reimplement that.
There was! And it even booted Linux in some capacity: https://github.com/christinaa/rpi-open-firmware
> every chip is very different from one another
Eh, the usual embedded SoCs are not that different from each other — ARM GIC, ARM timer, lots of Synopsys Designware crap for SDMMC/XHCI/PCIe/etc.
For many SoCs it's totally feasible to make standards-compliant firmware, e.g. for the Rockchip RK3566 there is https://github.com/jaredmcneill/quartz64_uefi
And SoCs from the networking world (Marvell, NXP) are typically supported by upstream EDK2.
Are there several different manufacturers of motherboards for it?
There isn't even an ecosystem to talk about. The fascination about the M1 is quite disturbing. Unless you want to be locked into the apple store it has no value whatsoever. The energy put into running Linux on it is really impressive. But from a broad perspective nothing positive can ever come out of it.
And it's still a closed ecosystem.
Sorry for Simon Segars who must have had a tough job essentially implementing Son's ideas. And those ideas seem to have been pretty poor - Arm China debacle, focus on IoT and selling to Nvidia have all clearly harmed Arm's prospects.
It's clear by now that selling CPUs is a mature market but that Arm has room to grow and there is still lots of money to be made in hardware if you innovate - eg in accelerated computing. It's also true both that Arm's model works and that it has more competition than ever from RISC-V.
A lot still to play for in my view and I think that the IPO price will reflect that.
I wouldn’t be surprised if there is some lobbying going on behind the scenes from the UK government to try and ensure it happens. Some strategic tax brakes or regional investment.
With the governments (somewhat controversial) investment in OneWeb last year they have shown they have a uk tech strategy. Although I think that was Dominic Cummings doing and he’s well out of the picture.
Starling bank on the other hand, is a great prospect. It's still founder-controlled, UK-focused, and would be able to join the FTSE 250/100 quite easily where the S&P is beyond its reach.
The LSE could really do with this as a boost and to break the cycle. But this will IPO in the US.
“There are currently 37 private $1b+ businesses headquartered in the UK—a further 14 have exited, whilst 2 are no longer active
13 UK companies became unicorns in H1 2021—the largest concentration of new unicorns to date”
You inspired me to have a look at how many “Unicorns” are in the US for comparison. Depending on source it’s between about 400-480 in the US. So about 12x as many with about 4.8x the population.
Deliveroo much less sustainable.
This is the cultural difference between ARM and Nvidia.
ARM does care enough about their IP that you'll need to pay them for ARM's effort to develop their IP, but beyond that they will not restrict you unless it'll put ARM into legal trouble (basically export restrictions). Hey, you could get everything now with a one-time license payment if you are willing to do that. It's not FSF open, but it's certainly better than a lot of hardware companies.
Nvidia is trying to extract everything into each pipeline. You want to use their hardware? You have to agree to only use this to precise conditions that is purely for market segmentation, which conflicts with ARM's culture. Compounding this with the fact that almost all acquisitions done by Nvidia that was more open was closed down (try talking to Nvidia about Mellanox), if the ARM acquisition proceeded, it'll be a bad day for everyone else.
F*** Oracle definitely, they can't even bother to have a proper SystemD service on systems which uses SystemD, relying on scripts that are definitely not redeeming them.
ARM makes prototype designs that you can extend, M1 is an absolute success of that model.
Regardless: you say they've failed as a business, your metric is how much they earn from being nearly everywhere.
I would pointedly disagree: they're everywhere, that's success, they're everywhere and profitable. There's absolutely no need to rent seek... this is unsustainable capitalist thinking.
They're able to innovate and turn a profit with their "modest" income, why do you need them to do more than that? How do you know they would do better with more money?
Intel (yes, apple to oranges) has many orders of magnitude more money, but most of it goes into bureaucracies.
See also: Microsoft, Oracle
Do you think that once they IPO pressure from their stockholders will cause rent seeking behavior? Even if they sold to nvidia frankly.
B: "it is a failure as a business. If you employed underhand tactics you could be even bigger!"
This is the phylosophy that gives us DRM on cars, 100 page lisence agreement with TVs, and fraud on wallstreet.
Obviously not true, but I can see how some might think this if their main comparators are companies that get to $1trn valuations in 30 years. Not every company can or should be a FAANG.
To be fair, AMD does super annoying lock in/binning things too, though not with software locking.
For example, there was a Radeon that would have been a crazy awesome server rack mounted GPU for Deep learning 4U blades. You couldn't put it in a server because it was mounted on a chassis that was 4mm too large for PCI spec.
(Btw I hate hate hate Nvidia so this is not pointless whataboutism)
They also have moved to locking their CPUs to the motherboards, to restrict second-hand sales of surplus CPUs/etc. So far it's only locked to the brand, so you can put it in any Dell system but only a dell system, but it's still sufficient to cause a huge amount of headaches for the surplus sector. I was looking at some cheap Epyc 7402Ps on ebay until I saw... "locked to dell, only runs on Dell". Welp. And the other cheap listing, is that one locked too? Who knows, and that's the point AMD is going for, they just have to inject enough uncertainty to change consumer behavior. Most people aren't going to be willing to trace the provenance of a used CPU, they will just pay 20% more for a new one. A limited number will be sold as complete systems, but as the market shrinks and people stop considering secondhand purchases, eventually they will just be thrown in the trash. Mission accomplished.
And that's coming for consumer CPUs too. It's already here for AMD's "ryzen pro" desktop line, it won't be long until it hits walmart prebuilts too.
That's not "the point". The purpose of the feature is legitimate, but you're right that the consequences suck hard for the second hand market.
The thing you're talking about is a security feature where the CPU would be locked to only boot when the firmware is signed by a particular signing key. If you're a high-value target such as a cloud vendor or a government / military, attacks on the firmware are a legitimate threat, and this ensures that the firmware cannot be silently tampered with. This is a big deal to them and that's the reason why AMD implemented this.
But yes, most people don't need or want this. It's entirely optional, it's supposed to be something the customer configures, but Lenovo turned it on by default. Pressure needs to be put on Lenovo to stop doing so.
What incentive would Nvidia have to develop new ARM features that would assist competitors in competing with them?
The Fujitsu's A64FX Processor for example somewhat competes with Nvidia datacenter GPUs. What incentive would Nvidia have to continue to develop the Arm Scalable Vector Extension?
The only good thing from this sale would be the pivot to RISC-V.
Nvidia is a castle because of the moats they've built. Not because they're leaps and bounds ahead in terms of silicon and technical wizardry. They shunned OpenCL, they'll probably shun Vulkan for GPGPU and make sure that people are locked to CUDA and its ecosystem. They have no intention to play fair.
We're going to see how Mellanox will evolve under their command.
[0]:https://www.reddit.com/r/realAMD/comments/qujwas/amd_drops_o...
this comment is factually untrue and anyone can check it for themselves.
https://en.wikipedia.org/wiki/OpenCL#Conformant_products
https://en.wikipedia.org/wiki/Vulkan#Intel,_Nvidia_and_AMD
it would be absolutely insane not to support either of those features. An absolute shit-ton of games use Vulkan these days, there is zero chance of that being deprecated in any timespan not measured in decades. Everyone is still supporting OpenGL after all, and that's like what, an early 90s API?
there is a larger point to be made here about some of the reflexively negative responses people have when the N-word is brought up. There are certainly criticisms you can make of NVIDIA, but people have this intense and reflexively negative emotional response that tends to slide these discussions into flamewar territory.
I know I'm about to hear about how Linus gave them the finger (cause he's a totally balanced and wholesome person), but providing a closed-source linux driver and not supporting wayland (which I think they did eventually anyway?) aren't the end of the world like people make them out to be. Pick the product that fits your needs, if not move on, you don't need to become emotionally attached and you can keep your positions fact-based.
I agree that it's my mistake, however there was a gap between this support periods and levels, and I've been burned by nVidia's own driver on an nVidia 800 series card. It's not out of spite or prejudice. I stand corrected, but my experience also stands. Also nVidia's practices for treating technologies they don't like is still valid.
Also, being compatible doesn't mean that these technologies are first class citizens on the driver/hardware level, and they may not be allowed to use the card up to its full potential.
> it would be absolutely insane not to support either of those features. An absolute shit-ton of games use Vulkan these days, there is zero chance of that being deprecated in any timespan not measured in decades. Everyone is still supporting OpenGL after all, and that's like what, an early 90s API?
There's support and there's support. I'm not talking about games and graphics stuff. I was doing some research about nVidia's "Vulkan Compute" support, and seen this:
"Vulkan compute on AMD and Intel might be okay but we already know of ways Nvidia puts the clamp on Vulkan compute (doesn't make use of multiple DMAs for transfer) and given their history with OpenCL and the vested interest they have in developers preferring CUDA, they're really not to be trusted." [0][1]
The guy writing seems to know what he's talking about (reading about the history gives a lot of nice low level details about how these things work), and this is what I was exactly saying about "Ditching Vulkan Compute in favor of CUDA, like OpenCL". "Yeah, it's supported, but it might not be fast, sorry."
> there is a larger point to be made here about some of the reflexively negative responses people have when the N-word is brought up. There are certainly criticisms you can make of NVIDIA, but people have this intense and reflexively negative emotional response that tends to slide these discussions into flamewar territory.
I'm personally no fan of flamewars. I also don't like to disseminate wrong information knowingly (see above). Also, I work in a HPC center where we use nVidia hardware, and personally develop scientific software. So, I'm not distant to either tiers of nVidia hardware. So, maybe not having prejudice about people one's replying to the first time is a good thing, no?
> I know I'm about to hear about how Linus gave them the finger (cause he's a totally balanced and wholesome person)...
No, you're not.
TL;DR: You can't write multi-vendor, GPU acceleration code which uses every card to its highest potential, even if there's standards. You need to write at least two copies of the code, one being CUDA. Why? nVidia doesn't like other technologies to work as fast on their cards. That's it.
[0]: https://www.reddit.com/r/vulkan/comments/jwhnfe/comment/gcq5...
For example, if you look at the TOP500, #1 (A64FX) gets about 14.78 TFLOP/watt, #5 (EPYC/A100) gets 27.37 TFLOP/watt. Most of the A100 based solutions are similar performance per watt. Nvidia is already beating ARM in power efficiency. This looks closer to the 3dfx acquisition from Nvidia's prospective.
> The Fujitsu's A64FX Processor for example somewhat competes with Nvidia datacenter GPUs. What incentive would Nvidia have to continue to develop the Arm Scalable Vector Extension?
"the enemy of my enemy is my friend".
the more R&D that goes into the ARM ecosystem, the tougher things get for x86. At the end of the day, all that R&D spending is making NVIDIA's position stronger, it's a force multiplier for their own efforts.
I fundamentally don't understand why anyone would think the ARM acquisition was about anything other than making Jensen kingmaker over one of the two most important processor IPs in the world - and he gains absolutely nothing by becoming king and then slaying all his subjects. That would be an amazingly shortsighted decision from one of the most far-sighted tech CEOs in the business.
Selling some more Tegras is peanuts in comparison and that bump would never last in the long term. The money is in Qualcomm and Fujitsu and IBM's R&D budgets working in synergy with your own, and in the ability to leverage the ARM licensing model to push CUDA into the last places it hasn't reached.
The fact that NVIDIA was even making this offer at all pretty much means they were looking at loosening up their IP licensing IMO. As a black box, sure, but I don't see a world in which NVIDIA would buy ARM and either not license GeForce as the graphics IP, or would choose not to license ARM at all ("selling a few more Tegras is not worth $40b"). If you accept those two givens, then NVIDIA would have had to provide GeForce IP as a black-box SIP license.
It doesn't matter, as the damage is done.
Once the industry caught a whiff of NVIDIA's intent to purchase ARM, they looked for alternatives. That's RISC-V.
Today, it's been years and the RISC-V efforts within the industry have advanced considerably. These won't be dropped now, irrespective of ARM's fate.
It is not other people's say unfortunately.
And a bad track record of preventing interoperability and cross-vendor standardization, to lock customers in.
Sorry had to do it.
Someone saw further down the road than I did.
ARM processors are also in many, many things besides Apple products.
To be clear I'm not suggesting this will happen or not, I'm just curious if it would be allowed?
Even if they bought a controlling amount of shares in public markets that would be subject to regulation.
I'll be looking to pick up some shares as a long-term investment. I think ARM is one of the best things to happen to computing, and anything that threatens the x86 monoculture/hegemony is a good thing in my view.
The companies that have their own invoice printer will “win”, the others will suffer.
Cloudflare reached 212 in November. Now it's down to 109.
Gitlab reached 125 in November. Now down to 73.
Coinbase was at 342 in November. Now it's 203.
You mean the PC standard?
I love ARM and the power-efficient mobile devices it enabled in the past two decades. That being said, I am deeply concerned with the lack of coherent hardware standards in ARM devices causing fragmentation and necessitate custom kernels, ROMs, and firmware.
If it weren't for PC, Linux might have died long ago.
Its a bit like the way you need more than an x86 CPU, you need an "IBM PC Compatible" system[1] with various buses and peripherals in the right memory locations, BIOS/UEFI firmware etc. for generic PC OSes to work. SBSA standardises something like an equivalent of an "IBM PC" for ARM Servers.
[1] Or whatever you would call it nowadays, as legacy BIOS booting is fading away modern systems might not count as that.
There are more ARM cores in the world by multiple orders of magnitudes at this point. ARM is the bully now, not Intel anymore.
Pedantic nit: not quite an IPO (Initial Public Offering) in its true sense since the original ARM Holdings was already a public limited company (PLC) before Softbank bought it from its shareholders in 2016. The trading entity and group parent is Arm Ltd, a private limited company formed in 1990 [0].
There is however a currently dormant Arm Holdings Ltd [1] incorporated in 2018 which may be converted to a public company so it would be the first time this legal entity has been publicly listed.
[0] https://find-and-update.company-information.service.gov.uk/c...
[1] https://find-and-update.company-information.service.gov.uk/c...
[0] https://www.reuters.com/article/us-dell-ipo-idUSKCN1OR14E
[1] https://www.wsj.com/articles/dell-returns-to-public-equity-m...
[2] https://www.theverge.com/2018/12/28/18159305/dell-stock-mark...
It's actually gone; MIPS owners have abandoned MIPS to focus on RISC-V.
Do you have a source for this? RISC-V always seemed like it was a production ISA from the get-go
"Provide a realistic but open ISA that captures important details of commercial generalpurpose ISA designs and that is suitable for direct hardware implementation."
"Be simple to subset for educational purposes and to reduce complexity of bringing up new implementations."
https://www2.eecs.berkeley.edu/Pubs/TechRpts/2011/EECS-2011-...
"Yunsup Lee, Krste Asanovi´c, David Patterson, and I conceived RISC-V in the summer of 2010 as an ISA for research and education at Berkeley. Two of our research ventures, RAMP Gold [92] and Maven [60], had just wound down. These projects were based around the SPARC ISA and a lightly modified MIPS ISA, respectively, and we sought to unify behind a single architecture for the next round of projects."
https://www2.eecs.berkeley.edu/Pubs/TechRpts/2016/EECS-2016-...
There's been a few changes since its inception in academy and the ratified standard unprivileged spec[0].
Then they realized that they need to push this more aggressively, establish a foundation and create a real standard.
You can listen to the 'State of the Union' talks from the early conferences if you want a source.
Because for the mass consumer market needing decent compute power, there is only ARM and x86-64 left, everything else has gone downhill in actual usage - and ARM is way easier to implement with a lot more vendors available than the x86-64 world which is basically Intel vs AMD.
The only competition that remains is Loongson for Chinese government purposes, but no one outside of China and maybe Russia will touch that stuff voluntarily.
It's meant to be used in areas of critical infra and national security, every major nation that is not joint with US at the hip is thinking or planning chip independance.
Otherwise your economy could be rouined because Donald Trimp 2 had too much cocaine in the morning and tweeted something.
The older architectures were much earlier (MIPS64 1991, SPARC64 1995, Itanium 2001, AMD64 2003, AArch64 2011).
Yes, AArch64 runs the top supercomputer, and there is enthusiasm; don't mistake this for market share.
Arm (the corporation) can make some money for themselves by coming up with increasingly cool Arm cores to sell (a mission Nvidia could have aided immediately and with high impact, but... nevermind) but the instruction set itself is not especially important anymore. There's a lot less path-dependency for companies in the embedded arena than there is among, for example, desktop computer companies.
ARM is on mobile, data center and desktop/laptop.
Some in depth look at various features here:
https://userpages.umbc.edu/~vijay/mashey.on.risc.html
Also, getting the job done with fewer instructions meant requiring less RAM in embedded roles where it thrived before entering the market from below in classic disruptive innovation style, the same way that x86 entered the server market. The other RISCs were mostly high end workstation or server chips suffering from low volumes. I actually ported some software from an SGI MIPS box to x86/Linux once. It was a nice system and I see why people liked it but it just couldn't keep up.
Thumb2 in 32bit ARM was competitive in code density.
AArch64 code density is abysmal, dramatically worse than RV64GC.
These eccentricities eventually became a liability.
https://www.jwhitham.org/2016/02/risc-instruction-sets-i-hav...
https://www.bloomberg.com/news/articles/2018-10-21/explainin...
Why does it take so long? I would be pretty surprised if they aren't already operating with audited GAAP financials, etc etc. Or is the extended date a hedge against market conditions, to avoid having to announce a delay.
M&A departments in New York or London must have been (or still be) pulling all-nighters.
exciting times ahead of us
I'm still unclear why Nvidia would want to control Arm. If they wanted an architecture they could extend as they wished, they can use RISC-V (which they are already doing in a small way).
I think nvidia is worried about being shut out of the x86 space so there's nowhere to sell their products. So they had a long term play to guide ARM development as an alternative platform. I could see that platform being useful as a hedge, but also giving nvidia a better way of competing in consoles and if they were really successful maybe eventually turning Intel/AMDs x86 advantage into a weakness.
Back when Intel killed off all third party motherboard chipsets nvidia won a lawsuit against Intel where they forced Intel to ensure a pcie bus for GPUs on all their platforms. That provision expired a few years ago. If nvidia is smart (and I believe they are) they've never let go of that worry that APUs might shove them out of the market. Their best bet is to remain so much better than the competition they cannot be ignored, but it would be wise to have a backup plan.
Nvidia have some really good high performance Arm cores that they've developed and sold (Denver, Carmel). Continuing to invest in that area makes some sense for Nvidia today, but it would make a lot of sense if the products they're paying to develop were official Arm products sold through the Arm IP sales channel. In other words, it would increase Nvidia's return on investment on work they're already doing.
(on the other hand, the amount they would have ended up paying for Arm might be fully justified by the value of the company but it was really, really high)