NVIDIA discussion [2024]

  • Thread starter Deleted member 2197
  • Start date
https://wccftech.com/nvidia-h20-ai-accelerators-next-us-ban-rumor-stops-taking-new-orders-china/

a potential H20 AI GPU ban are looming in the markets, due to which, Team Green has reportedly halted the sale of its accelerator to Chinese companies, as reported by Taiwan Economic Daily. It is said that NVIDIA has stopped accepting new orders for the H20 accelerator and hasn't specified a reason behind it, but it's highly likely that the firm is prepping for a potential revision in the US trade policy.
 
How I imagine that meeting preparation went:

"How do we get them to spend tens of billions more on things we're going to make a lot of money from?"

"Say they'll make money and it'll help against the big bad!"

Next thing you know they'll be saying doctors recommend OpenAI and saying that it's good for you. It's cool to see the limits of what's possible being pushed but >100x great power consumption from today (about 30-40MW for Fugaku and Aurora) seems incredibly unsustainable and something that'll backfire in the next 10-15 years.
 
"Tesla intends to spend at least $10B on Al in 2024, on several compute projects: the supercluster in Austin, Texas (planned to house 50,000 H100 GPUs and 20,000 of Tesla's own D1 chip), an inference computer, and Dojo.

Of the $10bn, Musk estimated spending $5b on internal development, $3-4bn on NVDA GPUs alone, and the rest on residual infrastructure.

 
An alliance between BlackRock, Microsoft, Global Infrastructure Partners (GIP) and MGX was announced in September 2024 with a goal of building the backbone of future AI infrastructure.

The new Global AI Infrastructure Investment Partnership (GAIIP) announced that it planned to raise $80 billion to $100 billion to build data centers and, perhaps more importantly, the supporting grid energy infrastructure to power them.
...
An initial $30 billion of private equity capital announced by GAIIP is planned to be raised by investors, corporations and asset owners, with a goal of enabling an increase to $100 billion, which can include debt financing.

Development plans include new energy plants and building new AI infrastructure, with an expectation that GAIIP will benefit from new policies that encourage private investment in the utility sector.

NVIDIA is on-board as an official technical advisor, with plans to design GAIIP’s infrastructure to integrate with NVIDIA’s AI factories and full-stack computing platforms so that performance can be maximized for AI and HPC workloads.
...
NVIDIA’s expertise in AI factories – which it defines as data centers "tailored to AI workloads" – would help guide the design and expansion objectives of GAIIP’s infrastructure in the face of mounting demands from generative AI, deep learning, and other cutting-edge artificial intelligence technologies.

But this isn’t just an NVIDIA game. The partnership has committed to an open architecture and inclusive design, and wants to facilitate engagement of a large number of partners, making sure that a very large ecosystem exists that can access its data centers and infrastructure.
 
Pharma you post at 7:12 makes no sense
The title claims Super Micro is Nvidia's third largest customer
then the first line claims Nvidia is a client of Super Micro, so who's the customer Super Micro or Nvidia ?
 
Pharma you post at 7:12 makes no sense
The title claims Super Micro is Nvidia's third largest customer
then the first line claims Nvidia is a client of Super Micro, so who's the customer Super Micro or Nvidia ?
It makes sense but shows how convoluted the industry has become.
Typically... customer = product, client = service.
Corporate speak will blur the lines and make their own definitions when it suits them.
 
They're both right, I think you might have just misunderstood.

For SuperMicro, their biggest customer is NVIDIA. Meaning nobody buys more SuperMicro than NVIDIA does. Remember, NVIDIA actually has to buy servers, they don't make them, and apparently they buy their servers from SuperMicro. Also keep in mind SuperMicro does literally everything from building their own ASICs and motherboards to building entire physical datacenters, and literally everything in between.

For NVIDIA, their 3rd biggest customer is SuperMicro. Meaning only two other entities buy more NVIDIA than SuperMicro does. And having recently attended a conference with SuperMicro, their biggest profit comes from selling entire racks prestacked with servers and NVIDIA GPUs. Remember above when I mentioned SuperMicro builds datacenters? You can just call them, and say "I want a 2GW 250,000sqft datacenter and I want you to build it for me" and they can do the entire thing.

Nothing about those two statements is incongruent.
 
@Davros ... do are you referring to the 2:12 pm post and is this the sentence?
Bear in mind that Super Micro Computer is NVIDIA's third-largest customer. What's more, SMCI's biggest client is NVIDIA, and its second-biggest client is an entity that is backed by NVIDIA.
I assume they are referring to the symbiotic relationship between Super Micro and Nvidia.
 
When we talk about NVIDIA, especially in the modern day, many of the newcomers in the industry believe that Team Green has been an AI-focused firm since the start, and only real enthusiasts are aware of the tales of NVIDIA and the GPU segment.

What started with NVIDIA's GeForce FX series of desktop GPUs has now turned into a platform that has not only dominated the desktop GPU markets but has also proven to be massive competition for the likes of AMD. Based on data shown by 3DCenter, NVIDIA has managed to dominate the GPU market share since 2002, and the trend continues to be the same.

NVIDIA-AMD-GPU-Market-Share.png

The competition was indeed strict in the initial years, back when AMD introduced the Radeon X100 series, which managed to narrow down the gap in market share due to the debut of technologies such as HDR rendering and CrossFire support, but since then, it's all NVIDIA, and even now, the firm is quite dominant when it comes to consumer adoption, despite putting dGPU as a secondary priority.

The pivotal moment for NVIDIA and its huge market share was during the crypto-mining hype, as we see market share figures soar up to 80% in that timeline.
...
Well, it won't be wrong to call NVIDIA's business a "Jensen masterstroke", given that in all business domains tapped by the firm, it has dominated it. Whether it be AI, robotics, or even the consumer GPU segment, we have seen NVIDIA's presence as a vital one for the progression of the respective markets, and Team Green's status is certainly equivalent to that of the Cupertino giant Apple, but for the computing markets.
 
There hasn't been any Shield discontinuation notice. Wonder if the next Shields will use Orin or Thor?
 
There hasn't been any Shield discontinuation notice. Wonder if the next Shields will use Orin or Thor?
I wonder what could they improve besides AV1/AV2 support, and maybe even better upscaling (despite already being best-in-class afaik)?

I don’t think NVIDIA has any incentive to improve gaming performance on Shield because of Geforce NOW but I did just realise GFN now supports AV1 which won’t work on current Shield which is unfortunate.

Orin/Thor are much more expensive SoCs, they’d only make sense for a high-end product, which historically hasn’t had that much demand in that market and wouldn’t replace the current Shield SKU. The kind of differentiators they offer wouldn’t be that appealing for the price IMO. And a rumoured Windows on ARM SoC would be as bad or worse cost-wise…

I love my Shield and would like an upgraded version but realistically I’m struggling to see how NVIDIA could justify making an entirely new SoC just for it.
 
Couldn't we expect to see a repeat of the current situation where the next Shield and Switch 2 share a chip i.e. the mooted T239? These products are not competitors.
 
Using the switch 2 chip would make sense, but I'm curious if the shield line will continue. The current model hasn't had a software update in a year now. It's running a version of Android TV that's 2-3 major versions out of date. It's still great, but idk if they have quietly ended support or what.
 
Back
Top