Nvidia has approached Softbank and is considering buying ARM Holdings

[uzzi38]

Will update with articles as they come out (they haven't yet but news is public). For now just a couple of Tweets:

https://twitter.com/x/status/1286003982079533056

https://twitter.com/x/status/1286004990327132161

This is a real significant move for Nvidia, having a huge role in ARM core designs could have major implementations down the line. Most importantly, it secures them a platform and ecosystem down the line.

Oh and I guess Nvidia becomes the defacto standard for GPU IP for mobile instead of Mali. That too.

EDIT: Bloomberg article here: https://www.bloomberg.com/news/arti...ompany-arm-is-said-to-attract-nvidia-interest

 

[DrMrLordX]

Guess we'll never know what NV would have done with complete control of ARM, Ltd. Good. Let them compete with everyone else.

 

[NTMBK]

Apparently, Nvidia is creating a new R&D group in Israel and hiring hundreds of people to accelerate CPU development in light of the changing HPC market. I'm guessing this is the fallback option given that their attempt to acquire ARM is likely going to be rejected. If this truly is the fallback option, it goes to show that buying ARM was firstly about buying the CPU design team and less about saving ARM.

https://videocardz.com/newz/nvidia-establishes-a-new-cpu-rd-group-in-israel

Interestingly, the original article never mentioned the HPC angle- simply that they are working on CPUs. I wonder whether we'll see custom CPUs from NVidia in the consumer markets too- it would be nice to see them produce a competitive laptop SoC, or a good console CPU.

 

[Frenetic Pony]

I don't get it, maybe Jensen is just plain losing it. How is CPU a good bet today? To use a business term, it's become an ocean so red you can tell there's blood blinfolded just by the smell. ARM, Apple, AMD, Amazon, Intel, Qualcomm, Microsoft, what's that RISC V startup, ahh Rivos. They're all out there doing CPUs as well, why the hell would you throw good money away competing against 4-5 dedicated vendors and 3 other major customers that aren't interested because they built their own chips. It's nonsense.

 

[itsmydamnation]

I don't get it, maybe Jensen is just plain losing it. How is CPU a good bet today? To use a business term, it's become an ocean so red you can tell there's blood blinfolded just by the smell. ARM, Apple, AMD, Amazon, Intel, Qualcomm, Microsoft, what's that RISC V startup, ahh Rivos. They're all out there doing CPUs as well, why the hell would you throw good money away competing against 4-5 dedicated vendors and 3 other major customers that aren't interested because they built their own chips. It's nonsense.

you dont understand NV's problem then, Jensen isnt stupid , that much should be obvious.

 

[RnR_au]

you dont understand NV's problem then

What is NV's problem?

 

[beginner99]

I don't get it, maybe Jensen is just plain losing it. How is CPU a good bet today? To use a business term, it's become an ocean so red you can tell there's blood blinfolded just by the smell. ARM, Apple, AMD, Amazon, Intel, Qualcomm, Microsoft, what's that RISC V startup, ahh Rivos. They're all out there doing CPUs as well, why the hell would you throw good money away competing against 4-5 dedicated vendors and 3 other major customers that aren't interested because they built their own chips. It's nonsense.

It was a good bet for AMD.

Intel soon has its own full stack of CPU and GPU and other accelerators. Intel can sell you a HPC server with all intel parts (optane memory, optane ssds on top).
AMD already has this stack of CPU and GPU and now they also have FPGAs (xilinx) as accelerators.
Ver likley both intel and AMD will add features limited to their platforms. Just look at zen3 and AMD gpus with the texture streaming or how it's called. That is just the first tiny step.

Nvidia only has GPUs and CUDA. See intel and AMD are threatening NVs HPC (=CUDA) dominance. Libraries like tensorflow start supporting more and more other things than CUDA. If you setup a new python data science envrioment with new libraries you can already see intels footprint with oneapi references everywhere. they are putting in a lot of effort into software (in contrast to AMD which still haven't learnt). The easier it is to not need CUDA (=nvidia GPU) the less sales nvidia gets.
Hence NV needs it's own full stack including the CPU which somehow work well together with their GPUs and CUDA environment. This is also why NV buying ARM would have been a threat as they could have somehow forced CUDA down your throat if you want to use an ARM server CPU.

 

[igor_kavinski]

What is NV's problem?

AMD has their own CPU. Intel has their own CPU. Both have GPUs that compete with Nvidia GPUs. They both will introduce features in their laptops that make their own iGPUs and dGPUs work together in harmony and accelerate some task like video encoding etc., making them seem very attractive to end users. Also, supercomputers. AMD and Intel will try to get their own CPUs and GPUs into whatever supercomputing contracts they win. Nvidia desperately needs a potent CPU.

 

[scannall]

AMD has their own CPU. Intel has their own CPU. Both have GPUs that compete with Nvidia GPUs. They both will introduce features in their laptops that make their own iGPUs and dGPUs work together in harmony and accelerate some task like video encoding etc., making them seem very attractive to end users. Also, supercomputers. AMD and Intel will try to get their own CPUs and GPUs into whatever supercomputing contracts they win. Nvidia desperately needs a potent CPU.

While I agree that they do need their own CPU, they didn't need to buy ARM to do that. Wanting to buy ARM was so they could strongarm ARM designers into using CUDA and other locked in tech from Nvidia.

 

[itsmydamnation]

What is NV's problem?

Memory access

everyone needs to learn, moving data hard , executing data easy.
NV's memory domain is to limited.
you have Wafer scale chips on one side (lots of memory) , FPGA's on another ( customise your execution/data flow) , then intel / AMD on the sea of servers front ( one big memory domain).

NV has had massive growth from AI , they want to keep it that way.

 

[Frenetic Pony]

Hmm... damned if they do, even more damned if they don't is the hypothesis, all vendor lock in and integrated SOC future?

I mean, Cerebras doesn't seem bothered by not having a CPU. And Nvidia still does quite well enough on the Desktop gaming market, I'm not sure SOCs are going to be dominating that anytime soon, especially with features like resizable BAR and low level APIs giving some of the integrated SOC advantages to desktop. Even worse for desktop gaming, programmers build in wide ranging compatibility from the start, so I don't see SOCs dominating gaming either in Desktop or Laptops anytime soon.

I'm not sure the vendor lock in libraries bullshit is the end of the world. And how long will the HPC craze last? Is just a political fueled flash in the pan about chest thumping, or is it really here to stay? Well... damned if they do, even more damned if they don't. Would've been much better for them if they could buy ARM. But hell maybe they can buy Rivos. No country will put up political pressure to keep them from getting a startup, and they'll get a CPU team working on the hot new open ISA.

 

[moinmoin]

I mean, Cerebras doesn't seem bothered by not having a CPU.

With all due respect, Cerebras are small minions in this. You are aware that Nvidia has over 50x the market cap of Cerebras?

And Nvidia still does quite well enough on the Desktop gaming market

Sure, but for Nvidia data centers toppled desktop gaming in revenue recently, and the latter won't ever get back to the top again after the acquisition of Mellanox.

I'm not sure the vendor lock in libraries bullshit is the end of the world.

Cuda is a long-living pioneering API that's ripe to capitalize on right now. Every competitor is scrambling to essentially replace Cuda. And Nvidia has to find hedges for the time Cuda is no longer dominating as it is right now. Tight vertical integration a la Apple appears to be Jensen's goal. Having control over the CPUs and overall platform is a necessity for that.

 

[LightningZ71]

That leaves NVidia with either having to come up with their own ISA entirely, or adopting an open ISA like RISC-V so that they aren't beholden to ARM/Softbank or the x86 license holders.

 

[Ajay]

With all due respect, Cerebras are small minions in this. You are aware that Nvidia has over 50x the market cap of Cerebras?


Sure, but for Nvidia data centers toppled desktop gaming in revenue recently, and the latter won't ever get back to the top again after the acquisition of Mellanox.


Cuda is a long-living pioneering API that's ripe to capitalize on right now. Every competitor is scrambling to essentially replace Cuda. And Nvidia has to find hedges for the time Cuda is no longer dominating as it is right now. Tight vertical integration a la Apple appears to be Jensen's goal. Having control over the CPUs and overall platform is a necessity for that.

As you have just said. Nvidia, in high performance compute (which encompasses more than traditional big iron HPC systems), is in need of a vertical solutions that give them total control over their own ecosystem. Then they can build in features that specifically enhance the competitiveness of their own platform without waiting to see if AMD or Intel will finally incorporate some similar standards based protocol/phy. Nvidia would also have the sales advantage of presenting customers complete, in-house developed server solutions - as they are doing now on a smaller scale. They can’t do that very well with x86 as they would compete with large, well established firms like HP.

Still, breaking out of a more boutique vendor, yet with enough differentiation to earn higher margins - could expand NV's revenue substantially. Of course, without a high performance, IO heavy ARM CPU, none of this can happen. They don’t have to match x86 CPU performance across as wide an application space as x86, that will help.

Apple's proven it can be done in the consumer space. Nvidia is a large enough company to shoulder the burden and time time to make this happen, IMHO. Will they be able to pull it off?

Anyway, my 2 cents.

 

[Ajay]

That leaves NVidia with either having to come up with their own ISA entirely, or adopting an open ISA like RISC-V so that they aren't beholden to ARM/Softbank or the x86 license holders.

I don’t think so, look at how far Apple has pushed ARM. The infrastructure in place at software tool vendors and fabs resources dedicated to ARM implementations is just too much of an advantage to overlook.

 

[Heartbreaker]

Wanting to buy ARM was so they could strongarm ARM designers into using CUDA and other locked in tech from Nvidia.

They would have VERY limited opportunities to do any arm twisting, since if they acquired ARM they would be under intense scrutiny, to maintain all the guarantees they would need to make to have any hope of doing an acquisition.

While I agree that they do need their own CPU, they didn't need to buy ARM to do that. Wanting to buy ARM was so they could strongarm ARM designers into using CUDA and other locked in tech from Nvidia.

They don't have a CPU design team though, and building a world class CPU design team is not easy. Look at both Samsung and Qualcomm that tried to design better ARM CPU cores but essentially gave up (Qualcomm recently purchased Nuvia (Ex Apple CPU designers) for over a Billion just to get an ARM design team - so they will obviously try again).

That is one factor. Another is they get full access to all the ARM customers as a massive sales channel to upsell other NVidia IP into (this one was stated by Jensen when asked).

Beyond that, owning ARM would mean they now also own their CPU-GPU stack like Intel and AMD. Even if NVidia could build acquire a world class team, the still wouldn't own the IP.

Owning ARM means they also have a vested interest in funding accelerating ARM into Servers/PCs, to a greater degree.

 

[Saylick]

Buying ARM was always going to, and is, scrutinized to all hell. They really should've bought Nuvia when they had the chance. 1B for Nuvia is far more bang for buck than trying to buy ARM, especially when you consider that if the acquisition fails they pay SoftBank over a billion anyways in fees. Nuvia's goal of gunning for the server market would've aligned with Nvidia's, but now that they've been bought by Qualcomm, that effort is being diverted to mobile SOCs. I just think Nvidia got too greedy (in usual Nvidia fashion) by trying to acquire the whale when they could've had that 1 lb lobster and nurtured it into a 3-4 lb main course.

 

[Doug S]

While I agree that they do need their own CPU, they didn't need to buy ARM to do that. Wanting to buy ARM was so they could strongarm ARM designers into using CUDA and other locked in tech from Nvidia.

Totally agree. Apple was able to do what they did buying tiny companies like PA-Semi and Intrinisity, they didn't need to buy ARM. There is zero need for Nvidia to buy ARM if they want to sell CPUs to go along with their GPU, unless they want to try to restrain the ARM core licensing market in some way. They could have bought Nuvia instead for a tiny fraction of what they're paying for ARM if they wanted a jump start in designing their own high end ARM cores.

I checked historical prices for 2008 around the time Apple made those other purchases - ARM's price was about 7% of what Nvidia offered so a few billion would have got the deal done. They didn't buy ARM because it wasn't necessary for their CPU design goals.

The really telling thing about how much ARM is worth is what ARM's CEO has said to UK regulators - basically that ARM couldn't get nearly as much money if they IPOed instead of sold off to Nvidia. If they aren't worth that much on the open market, where is Nvidia going to extract their extra value from them to make it worth the hefty offer? Has to be via higher revenue from licensees, which means raising their prices and/or selling them more stuff (i.e. bundling Nvidia GPUs with the core licenses)

 

[Thibsie]

Buying ARM was always going to, and is, scrutinized to all hell. They really should've bought Nuvia when they had the chance. 1B for Nuvia is far more bang for buck than trying to buy ARM, especially when you consider that if the acquisition fails they pay SoftBank over a billion anyways in fees. Nuvia's goal of gunning for the server market would've aligned with Nvidia's, but now that they've been bought by Qualcomm, that effort is being diverted to mobile SOCs. I just think Nvidia got too greedy (in usual Nvidia fashion) by trying to acquire the whale when they could've had that 1 lb lobster and nurtured it into a 3-4 lb main course.

Agreed. This is, IMO, a big fail. Jensen's fail.

 

[DrMrLordX]

They don't have a CPU design team though,

Actually they do. Whom do you think designed Grace, among other things?

 

[Heartbreaker]

Actually they do. Whom do you think designed Grace, among other things?

Perhaps I should have said, "top notch" CPU team.

Their designs don't have seem to have stood out from ARM designs:
Carmel in Xavier is the latest tested that I found:

AnandTech Forums: Technology, Hardware, Software, and Deals

Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

www.anandtech.com

In the integer workloads, the Carmel CPU cores' performance is quite average. Overall, the performance across most workloads is extremely similar to that of Arm’s Cortex-A75 inside the Snapdragon 845, with the only outlier being 462.libquantum which showcases larger gains due to Xavier’s increased memory bandwidth.

This is a shipping CPU with ASIL-C functional safety features that we have in our hands today. The only competition in this regard would be Arm’s new Cortex A76AE, which we won’t see in silicon for at least another year or more. When taking this into account, it could possibly make sense for NVIDIA to have gone with its in-house designs, however as Arm starts to offer more designs for this space I’m having a bit of a hard time seeing a path forward in following generations after Xavier, as competitively, the Carmel cores don’t position themselves too well.

And indeed it sounds like Grace will use ARM Neoverse Cores not, in house NVidia cores. Which is similar to Samsung and Qualcomm that dabble but didn't stand out from ARM designs and went back to ARM designs.

Being in control of ARM, they could push Server features they want faster, as it sounds like the only benefit of Carmel, was "ASIL-C functional safety features" they wanted, and now they are back to ARM cores:

Introducing Grace

A CPU Purpose-Built for Servers Running Terabytes of Data

www.nvidia.com

A fast and efficient CPU is a critical component of system design to enable maximum workload acceleration. The NVIDIA Grace CPU integrates next-generation Arm Neoverse™ cores to deliver high-performance in a power-efficient design, making it easier for scientists and researchers to do their life’s work.

 

[DrMrLordX]

Perhaps I should have said, "top notch" CPU team.

That would seem more accurate. Too bad for them they didn't buy out Nuvia.

 

[Heartbreaker]

That would seem more accurate. Too bad for them they didn't buy out Nuvia.

They wen't for the home run, instead of the base hit.

 

[HurleyBird]

That would seem more accurate. Too bad for them they didn't buy out Nuvia.

Nvidia along with AMD, Intel, Microsoft, Mediatek, etc... there's a long list of companies that should have made a play for Nuvia. From a cost-benefit standpoint, the benefit of a CPU design team that can compete with Apple is, at least, measured in tens of billions of dollars. If you think that there's even a 10% chance that Nuvia does that, then they're still a steal at $1.4B.

 

[Doug S]

Nvidia along with AMD, Intel, Microsoft, Mediatek, etc... there's a long list of companies that should have made a play for Nuvia. From a cost-benefit standpoint, the benefit of a CPU design team that can compete with Apple is, at least, measured in tens of billions of dollars. If you think that there's even a 10% chance that Nuvia does that, then they're still a steal at $1.4B.

But you don't get that long term "tens of billions" of value from them unless you convince the principals to stay beyond their lockup period. It remains to be seen if Qualcomm will be able to keep them there. There's even the additional problem that they had to redesign their core to be appropriate for Qualcomm's goals instead of Nuvia's goals, which is why it is arriving over a year later than Nuvia originally claimed.

Qualcomm may not even get a second generation core out of them if they decide not to stick around. So while it may be worth $1.4 billion to give them a leg up for a couple of generations, if they don't keep a team able to maintain an advantage over ARM designed cores then it certainly isn't worth "tens of billions of dollars". That's only true if they keep them around for many years. Since they left Apple over being unsatisfied with the prospect of designing mobile, laptop and desktop cores when they wanted to do a server design, it doesn't seem too likely to me they will suddenly be satisfied doing the same non-server stuff at Qualcomm.

 

[DrMrLordX]

But you don't get that long term "tens of billions" of value from them unless you convince the principals to stay beyond their lockup period. It remains to be seen if Qualcomm will be able to keep them there. There's even the additional problem that they had to redesign their core to be appropriate for Qualcomm's goals instead of Nuvia's goals, which is why it is arriving over a year later than Nuvia originally claimed.

Qualcomm may not even get a second generation core out of them if they decide not to stick around. So while it may be worth $1.4 billion to give them a leg up for a couple of generations, if they don't keep a team able to maintain an advantage over ARM designed cores then it certainly isn't worth "tens of billions of dollars". That's only true if they keep them around for many years. Since they left Apple over being unsatisfied with the prospect of designing mobile, laptop and desktop cores when they wanted to do a server design, it doesn't seem too likely to me they will suddenly be satisfied doing the same non-server stuff at Qualcomm.

Well then they shouldn't have sold out in the first place.