NVIDIA discussion [2025]

[Kaotik]

Yeah there's a number of online AIs out there which run on Google's Tensor chips. As a simple example, https://pi.ai is one such model which all inference nodes are powered by Tensor. I think the actual training is still done on NV hardware tho...

In Google's case it's apparently both Google, training is for sure

https://www.cnbc.com/2024/07/29/apple-says-its-ai-models-were-trained-on-googles-custom-chips-.html

 

[Florin]

Does this mark a shift in strategy for Apple? What’s Siri running on now?

I believe it was widely publicised that Apple opted for Google's TPUs as their main AI partner, both for the server side and training for on device processing.
However I've also seen rumors that Apple has actually quietly been using a lot of rented cloud Nvidia GPUs for a while now.

The choice to actually buy hardware itself is certainly a shift. But then again one billion barely qualifies as a toe in the water compared to major players like Microsoft and Meta.

I reckon Apple is just biding its time and plan-B-ing it until its own AI chip is expected to be good to go (2026, as far as we know).

There isn't any love lost between Apple and Nvidia - Apple's vertical integration is anathema to Nvidia, and Nvidia's success in negotiating and then sticking to profitable pricing for something as lowly as components is not something Apple can abide.

 

[Cyan]

things are changing. Intel is starting to lean towards ARM, and AMD announced the other day that they are going to get into the ARM like race.

 

[Scott_Arm]

things are changing. Intel is starting to lean towards ARM, and AMD announced the other day that they are going to get into the ARM like race.

Casey Muratori just published an interview with the chief architect of Zen, which talks a bit about ARM vs x86 in terms of how they affect CPU design. The impression I get is that it's not as significant as people make it out to be.

An Interview with Zen Chief Architect Mike Clark

Zen is one of the most important microarchitectures in the history of the x86 ecosystem.

www.computerenhance.com

 

[Bludd]

Casey Muratori just published an interview with the chief architect of Zen, which talks a bit about ARM vs x86 in terms of how they affect CPU design. The impression I get is that it's not as significant as people make it out to be.

An Interview with Zen Chief Architect Mike Clark

Zen is one of the most important microarchitectures in the history of the x86 ecosystem.

www.computerenhance.com

afaik modern x86 cpus are basically only CISC on the outside

 

[DegustatoR]

When asked about future generation Nvidia GPU architectures like Feynman, which is expected two generations from now (2028), Huang mentioned that if Nvidia transitions to a process technology that relies on GAA transistors, it should bring a 20% increase in performance.

Our own Jarred Walton was at the Q&A, and says Huang seemed to downplay the importance of process node changes, emphasizing that the slowdown in Moore's Law means brand-new process technologies going forward are only likely to bring around a 20% improvement — in density, power, and/or efficiency. It wasn't a definitive statement on what node Nvidia might intend to use, though the answer was in response to an analyst question looking for his comments about the potential for Nvidia to use Samsung Foundry in particular.

Huang also noted that while improvements enabled by leading-edge process technologies are welcome, they're no longer transformative. "We'll take it," he said, but indicated other factors were more important. As AI systems scale, the efficiency of managing vast numbers of processors is becoming more important than the raw performance of each processor. Data centers are increasingly looking at performance per watt — that "we're at the limit of physics."

Nvidia's Jensen Huang expects GAA-based technologies to bring a 20% performance uplift

Is that enough for GPU production to jump to a cutting-edge node?

www.tomshardware.com

 

[pcchen]

afaik modern x86 cpus are basically only CISC on the outside

Instructions wise that's true, but there are some design decisions of x86 (and by extension, x86-64) which are not common among RISC CPUs. For example, x86's memory ordering model is pretty strict, basically all memory writes are considered visible to all subsequent instructions, while RISC ISA have various degrees of strictness in this regard. Apple implemented a special mode in their M-series CPU in order to emulate x86 efficiently, as otherwise all memory write instructions needs to have a memory fence instruction attached. ARM binaries run without this special mode to get better performance.

 

[eastmen]

things are changing. Intel is starting to lean towards ARM, and AMD announced the other day that they are going to get into the ARM like race.

This happens with amd every decade it seems now

AMD Announces Opteron A1100 Series 64-bit ARM Processors for the Datacenter

AMD marks a major step towards delivering choice and innovation in the data center with the launch of the AMD Opteron A1100 System-on-Chip (SoC), formerly codenamed "Seattle". Jointly with its software and hardware partners, AMD is accelerating time-to-deployment of ARM-based systems and driving...

www.techpowerup.com

 

[DavidGraham]

The first data center of project Stargate will house 400K of NVIDIA GPUs. Making it one of the largest AI clusters known to man.

https://twitter.com/x/status/1905625172109033965

On other news, NVIDIA H20 supply is exhausted due to a huge surge in demand.

Nvidia H20 chips are in short supply amid surging demand in China, local AI server maker H3C warned, Reuters reports. In a note to clients, H3C said its “supply chain faces significant uncertainties,” with H20 inventory nearly exhausted. H3C will prioritize stable, long-term…

— Dan Nystedt (@dnystedt) March 28, 2025

 

[trinibwoy]

The first data center of project Stargate will house 400K of NVIDIA GPUs. Making it one of the largest AI clusters known to man.

It doesn’t say it will house 400K GPUs just that it has the capacity for it.