Stefem:

Why they would need to use InfiniBand which is an internode (to connect computers) communication protocol when they have NVLink that is a intranode (to connect processors) protocol?

Most of the Supercomputers already have HUGE InfiniBand switches.. Supercomputers are not one huge MB where you can plug in more and more GPU's.. connecting nodes is, or being able to replace 'legacy nodes' is where compatibility is important.. and where the big bucks are. *on second thoughts.., if you're asking "why not use NVlink to connect nodes" then I doubt NVlink can sustain its speed over a 3-5M cable, but maybe it can/will now Nvidia own Infiniband. Also Infiniband transmits much more than just GPU traffic.. again maybe even more now Nvidia own the IP.

Stefem:

Why? what debacles? They are in the NINTENDO Switch and the X360 had an AMD's GPU, the real reason AMD won in both consoles is that they offered prices so low to be virtually unprofitable for them with the purpose of pushing NVIDIA out of that market (which is a nice one since there is no competition at all), consider that despite the PS4 alone sold more than 100M unit they almost gone bankrupt and they had to sell their headquarter building to survive...

To be fair Nvidia didn't play nice with the GPU in the Original xBox and when Sony came begging for a GPU for the PS3 after their internal development didn't turn out.. AMD got kudo's for delivering what 'price-conscious' customers wanted... but who can blame Nvidia asking to be paid for the crazy money they spent, creating and pushing the GPU world forward? Rumour has it nvidia wanted a win in the console world no matter what the cost.. so that's why we have been gifted a great console in the form of the 'switch' 😀

Stefem:

There are lots of funny comments, I understand that many like to see the tech industry in a overly romanticized way but things aren't so simple to be divided in good and evil, it's more like different shade of gray and there are also wolfs in sheep's clothing...

Where did I say otherwise? If you actually read my post, you'd see that my perspective on the matter is very gray.

It's interesting you are worried for Qualcomm when they was them that have been fined for anti-competitive practices for selling SoC bundled with modems at artificially low prices to push emerging competition (NVIDIA and modem made by Intel to name the biggest) out of the market...

Where did I say I was worried about Qualcomm?

Anyway, I would really like to know what kind of "proprietary nonsense" NVIDIA would plan to add to an already proprietary ISA (yes, ARM ISA is proprietary), maybe some proprietary extension?... 🙄

ARM can be licensed. Assuming Nvidia keeps the licensing (which they might not, if they intend to monopolize the market), they can do what they did with their GPUs: make a technology needlessly exclusive to their platform, forcing customers and developers to "pick a side". If you buy Nvidia, great - you have nothing to worry about. But it's very anti-consumer to tether people to a single company. Some examples on the GPU side of things were CUDA, PhysX, G-Sync, OptiX, etc. None of those had to be made proprietary. I do not fault Nvidia for things like SLI, DLSS, and certain media codecs, since those things are too heavily tied to their hardware.

You probably don't know but NVIDIA is one of the main RISC-V supporter and the first to use it in a high performance mass produced product (Volta, Turing and I think even Pascal use RISC-V for GPU internal controller)

How in any way is that relevant? RISC-V is open-source; Nvidia can do whatever they want with it. That has nothing to do with whatever they plan to do with ARM.

Dudes I'm all for you having good conversation, but you guys need to not do multiple posts in a row! Edit!

The more I think about this, the more I think Nvidia is going to make massively parallel, multi-chip modules that will run CUDA on ARM cores. Just think about what has been done with ARM cores so far - eighty cores with four-way symmetric multi-threading. I suspect they will make an ARM-based equivalent of a streaming multiprocessor with eight-way SMT and sixty-four cores for 512 threads on a chip, possibly 4x64 for 256 threads. That will be on one module and the MCM will contain up to eight of them for 4096 (or 2048) threads in a single package. The really cool thing about this is it would not be a co-processor like a GPU is so no additional data transfer would be necessary. Just imagine the super computer that could be built with those and how low its power consumption would be. I think this is going to be really impressive and it will change the direction of HPC and super computers.

Nvida to buy arm, they would lock it down make some proprietary crap out of it and then try to sell it like it's their actual developed baby.

Agonist:

Nvida to buy arm, they would lock it down make some proprietary crap out of it and then try to sell it like it's their actual developed baby.

isn't softbank's model already collecting royalty and license fees

well it's not surprising to me that Nvidia would want to but ARM, but it's almost inconceivable that they'd be able to outbid Apple or Qualcomm. no one has more cash on hand than Apple (even though some of it is parked in Eire), no one is more reliant (now) on ARM design given their stated ditching of Intel. Qualcomm is a distant #2, but has lots of other people's money from legal fees and a willingness to go to court that's way above the litigious standard of the tech industries.

Microsoft attempted to reneg the contracted price they paid for the gpu's and nvidia beat them in court over it.