Agree. It's a terrible time to want a new Nvidia card for gaming or AI projects. I had a 4070 Ti Super in my cart for a while at a cool $700 on Amazon - now it's out of stock and secondhand I can find it for double that now.
If one doesn't require the latest DLSS or need first-class AI support, get an AMD card.
The entire ecosystem is built around CUDA currently. There was someone implementing Cuda on AMD hardware but as I understand it, AMD shut that down.
There are two alternatives to CUDA for AMD right now, but I forget their names.
AMD has STRIX coming, which is a single board embedded ram machine, and is being targeted at AI loads with lots of low latency ram. So we’ll see what happens there.
Thankfully the ONNX runtime supports AMD's ROCm. The performance is nowhere close to Nvidia's TensorRT, but doing inference on an AMD GPU is doable. I haven't ran the numbers, but I wouldn't be surprised if you could get more throughput per dollar on an AMD GPU for at least some scenarios. AFAIK there aren't any great options for training though, so if you want to do anything besides running ONNX models your options are limited.
It's probably false to say they're leaving money on the table. They'd be leaving a lot more money on the table if they allocated chips towards consumer gaming GPUs instead of maxing out the server AI/GPU compute segment. The entire gaming market constitutes like 15%-or-less of their revenue nowadays.
And Nvidia has enough mindshare that they could piss on consumers for the next 3 release cycles and still have more than half the market. I don't like it but it's reality.
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.
Yep, all 10 they shipped sold instantly at 2x MSRP. Prices on 3090's (not a typo) are going up ffs, because there is nothing out there on the NVIDIA front.
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!
Egg production fell because of avian flu, which the US has regulatory restrictions limiting the usage of the vaccine in agriculture compared to other developed nations. Its not an intentional crisis.
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.
NVidia doesn't make the chips, they just design them, so it is a flex. People are literally buying their cards faster than they can be produced. TSMC is building new fabs to make more nvidia chips faster.
NVidia decide how they allocate their wafer starts at TSMC. Consumer chips have lower margin than datacenter parts, so they almost certainly allocated comparatively little volume to consumer chips.
On the one hand, this is a great situation to be in for Nvidia in terms of overall revenue.
On the other hand, this has allowed AMD to grab market share with the RX 9000 series launch, at least in the short term. So the narrow point that Geforce is sold out is decidedly not a flex.
GTC stands for GPU Technology Conference, for anybody else who wasn't familiar with the acronym. Surprisingly they don't have it spelled out anywhere on the homepage.
It means “envy” in Latin and Spanish etc.
Which is why the logo is an acid green evil eye.
And why NVIDIA’s headquarters are two buildings named “ENdeavor” and “Voyager”.
Nvidia has lost the hearts and minds - I'm not just not interested in or excited by anything Nvidia - there's a baseline expectation of their products being overpriced and aimed at stiffing the customer.
what's the better alternative, DIY AMD? No, thanks. I'm interested in the result, not the process. In terms of robotic software there is just nobody close. Any attempt to port it to different hardware will be just a big pain and limited result at best. So NVidia looks to be the way to go till China catches up.
honestly, seeing a company that craps so hard on a core customer group still succeed THAT much just because of some random trends is just sad. and yet people are still willing and continue to vote with their wallets (as the comment below was proving). what a mess
The video glitches out at 2:14:30 where Jensen would be introducing DGX Spark and Station. Does anyone have a link to a working video for that segment?
DGX Spark has the same memory as AMD Strix Halo, a weaker CPU, but perhaps a stronger GPU, except that for now there is no data about the GPU, besides that it might be stronger for AI inference (only FP4 speed is given). For now it is not known whether for graphics it will be better than Strix Halo.
While DGX Spark might be weaker than AMD Strix Halo for anything else except AI Inference, it will still be stronger than any mini-PC made with Intel Arrow Lake H or with AMD Strix Point.
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.
I'm somewhat surprised they are not explicitly mentioning the networking capabilities. There should be two QSFP(?) ports but they don't mention if their speeds and supported protocols (possibly not just Ethernet but also infiniband, but probably not).
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.
You seem to have somehow missed one of the most significant advancements in human history suddenly exploding out of the pages of science fiction and fundamentally altering our existence forever. Minor oversight.
Ah yes, a chat bot that is consistently failing to do such advanced tasks as counting letters in words is definitely life-changing, right next to the metaverse and NFTs.
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.
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.
That's what everybody was predicting based on the memory technology and the pictures. I'm just glad to hear about DGX Station, though I doubt I'll be able to afford one.
how much is it expected, my guess barely fits in 5 digits. Would be nice to have something in between Spark and Station. I.e. some desktop withing $20K.
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.
not a flex