Jensen Huang – Nvidia GTC 2025 Keynote

The top end should at least be stocked because a lot of us are using cards locally for AI that eventually runs in the cloud.

So yeah they did leave $2600 from me on the table that is now becoming more likely to be spent on a bootleg 48GB 4090 than a 5090 and if I get that they won’t see money from me for many years till they beat 48GB in consumer form factor.

"You PC people got us where we are now - so screw you, we don't need you anymore because the AI bubble will go forever!"

AMD has documentation for its assembly language as I remember. Or one needs more documentation to run AI?

Egg prices didn’t raise because of a flu.

Egg prices raised because the government asked for the killing of a hundred million of chickens - for fear of spreading bird flu. It may have been the right call, we’ll never know.

Small difference, but important.

Here, CNN is obscuring the fact that the chickens killed were not tested or confirmed to have flu, but some around them might have so they had to go too.

https://www.cnn.com/2025/01/28/business/chickens-avian-flu/i...

> Egg production fell because of avian flu

Well... Yes but no.

https://www.thebignewsletter.com/p/hatching-a-conspiracy-a-b...

But the datacenter GPUs are sold out too even at ~90% profit margin. I heard somewhere lead time is in years.

This made me curious, so I just took a look on eBay.

My 4070 Ti that I bought (new, oops) this past December has appreciated 20 - 25%! At least according to what people have them listed for... no idea if anyone is actually buying them at those prices.

Of course even if I managed to sell it, everything else has gone up in that time, so it's not like I'd get to make money on the deal. Pretty wild, nonetheless!

It was a full computer, $5200 for an i9-14900KF, 64 GB, 4TB NVMe machine with a 5090.

success is the combination of luck and preparation. NVIDIA has been well prepared + anticipating deep learning to take off for a long while.

Perhaps you are confusing Strix Halo with Strix Point.

The GPU of Strix Halo is many times faster than the GPU of the biggest Orin AGX.

The Strix Halo has 25% more "CUDA cores" (2560 vs. 2048), which work at a clock frequency that is more than double (2.9 GHz vs. 1.3 GHz) and which can have for some operations a double throughput even at the same clock frequency. The memory throughput is also higher by 25%.

The GPU of DGX Spark will have to be 4 or 5 times faster than Orin to match Strix Halo as a GPU. This is not at all certain because NVIDIA has stressed only AI/ML applications without saying anything about graphics.

For operations with floating-point numbers, big integer numbers or arrays, a Cortex-X925 matches 3/4 Zen 5 full cores at the same clock frequency, while Cortex-A725 matches 1/4 Zen 5 full cores.

So DGX Spark is equivalent with at most 10 * 3/4 + 10 * 1/4 = 10 Zen 5 cores versus 16 Zen 5 cores of Strix Halo.

In reality DGX Spark will be even slower, because it will have a lower clock frequency (especially the Cortex-A725 cores) and a worse cache memory.

For irregular code that does only operations with pointers and integers, the advantage of AMD Strix Halo will be significantly less, but even in that case the 10+10 cores of DGX Spark are unlikely to match more than 15 Zen 5 cores at the same clock frequency and less than that at the real clock frequencies.

On the other hand, like I have said, DGX Spark should be faster than Intel Arrow Lake H, the best that Intel can offer in a mini-PC of this size.

Pointing out this one error is reductive and greatly simplifies the profound effect that LLMs (and research around them) are having across a wide range of industries, unlike the metaverse and NFTs.

Metaverse is changing the world. Just as you could see in this presentation - Data centers are first being built as digital twins in Omniverse and later in real world.

It is a profound change, but it will take time for significant number of digital twins being build. However digital twin is a CAPEX in a way. You build 3D model of your warehouse only once and then use it for years to manage operations, robots and everything else.

It means that Disney, company which is a CGI/movie/cartoon producing company(Pixar created OpenUSD format) can use the same tools and make physical robot walk on stage.

If the venue, where the presentation took place also has readily available 3D model made with OpenUSD in SimReady level of details, Pixar could just download this model to their cartoon tools and start training virtual model of the robot in the virtual model of the venue before the event and troubleshoot any issues with the program.

Isn't it mindblowing?

Ah I had a typo. The vids are from 2024 and 2025. It seems to me the 2024 robots are quite clunky moving as you kind of expect from robots, whereas the 2025 one moves quite well with comparable dexterity to an animal. Which according to the talk is down to nvidia's new tech. I mean you see robots move like that in movies but I think it's the first time I've seen it in real life. Apparently they are going to put them in the Disney theme parks.

It's significant in the scheme of things in that LLMs have got quite good at text chat but for AI do to real world things like build you a house or fix your car it's got to get good at physical robot stuff too.

The unified memory model w/ CUDA support should be the huge differentiator for this hardware.

I think a better comparison might be the Mac Studio ultra.

Based on the chip count in the promo picture it looked like it was going to be a 256 bit interface, but it'd be unfair to say that it was known at that point. I certainly wish it was better.

well, "RTX Pro Blackwell" desktop and workstation version looks to be the answer for midrange. Some should be available starting in April.