Jensen Huang keynote at CES 2025 [video](youtube.com) |
Jensen Huang keynote at CES 2025 [video](youtube.com) |
Blackwell is still based on N4, a derivative of 5nm. We know we will have N3 GPU next year, and they should be working with TSMC on capacity planning for N2. Currently Blackwell is pretty much limited by capacity of either TSMC, HBM or packaging. And unless the AI hype dies off soon ( which I doubt will happen ), we should still have at least another 5 years of these GPU improvements. We will have another 5 - 10x performance increase.
And they have foreshadowed their PC entry with MediaTek Partnership. ( I wonder why they dont just acquired Mediatek ) and may be even Smartphone or Tablet with Geforce GPU IP.
The future is exciting.
Market access, shares and relationship to all vendors, most of them never worked with Nvidia. Ready made 5G solutions. Know how in product category that is not as simple as buying IP and making a chip. Shipping Nvidia CUDA GPU IP to all other domains. For acquisition benefits they share the development cost of leading node and volume. Which is part of the wider SemiConductor industry strategy aka Broadcom.
But yes. I guess China would be the first one against it. I never quite understand why Two Company in country A and B would need Country C 's approval for them to merge or be acquired.
A lot of people would think so. In reality had Smartphone ( iPhone ) not taken off. The PC market alone would not be able to sustain the cost of innovation and Moore's Law would have ended at 14nm and we would have to wait 3 to 4 years for every node. Smartphone along with all the adjacent work on Internet infrastructure scaled us to 3nm. I wrote about both of these in 2014 and was expecting some slow down by 3nm. Now AI will provide the Capex to scale us to 1nm and beyond.
We aren't seeing nearly the same gains on VRAM bandwidth as we are on compute bandwidth
I know stock prices is driven by AI hype but how much does it actually effect the bottom line of Nvidia? I think GPU improvement happens regardless of AI.
Datacenter revenue alone is ~10x of gaming. The datacenter revenue is thought to have literally ~100x the earnings all up (H100 and 4090 have similar transistor counts but the H100 sells for over $30k while the 4090 sells for $2k which indicates huge margins).
Gaming is pretty much insignificant for nvidia. That’s why nvidias stock has 10x’ed recently and their PE looks better now than it did 5 years ago despite that stock increase. They found a new market that dwarfs their old market.
Of course, the GPUs for the datacenter and for gaming are the same designs, so my read is that in gaming NVIDIA makes up for lack of actual performance for traditional rendering by pushing technologies that can utilize tensor cores like AI upscaling, frame prediction, ray tracing + denoising, etc.., that don't actually contribute to game graphics as much as they could have if they did an architecture tailored to gaming needs instead, with the technologies that they have. It's also sexier in theory to talk about exclusive AI-powered technologies proprietary to NVIDIA than just better performance.
Gaming almost doesn't even register in Nvidias revenue anymore.
But I do think Jensen is smart enough to not drop gaming completely, he knows the AI hype might come and go and competitors might finally scrounge up some working SDKs for the other platforms.
It wouldn't be approved by country that Mediatek resides in. For the same reasons they won't approve a TSMC sale either.
AI is bewitching.
Claim it will be quite good at AI (1 Tflop of fp4), but sadly don't mention the memory bandwidth. It's somewhere in the range of awesome to terrible depending on the bandwidth.
[1] Nvidia's Project Digits is a 'personal AI supercomputer'
If that's not just hot hair from NVIDIA, I totally get the business decision from AMD but man, would love some more competition in the higher end.
Ask HN: Pull the curtain back on Nvidia's CES keynote please - https://news.ycombinator.com/item?id=42670808 - Jan 2025 (2 comments)
This has been stated on HN in most robotics threads, but the core of what they show, once again, is content generation, a feature largely looking for an application. The main application discussed is training data synthesis. While there is value in this for very specific use cases it's still lipstick ("look it works! wow AI!") on a pig (ie. non-deterministic system being placed in a critical operations process). This embodies one of the most fallacious, generally unspoken assumptions in AI and robotics today - that it is desirable to deal with the real world in an unstructured manner using fuzzy, vendor-linked, unauditable, shifting sand AI building blocks. This assumption can make sense for driving and other relatively uncontrolled environments with immovable infrastructure and vast cultural, capital and paradigm investments demanding complex multi-sensor synthesis and rapid decision making based on environmental context based on prior training, but it makes very little sense for industrial, construction, agricultural, rural, etc. Industrial is traditionally all about understanding the problem or breaking it in to unit operations, design, fabricate and control the environment to optimize the process for each of those in sequence, and thus lowering the cost and increasing the throughput.
NVidia further wants us to believe we should buy three products from them: an embedded system ("nano"), a general purpose robotic system ("super") and something more computationally expensive for simulation-type applications ("ultra"). They claim (with apparently no need to proffer evidence whatsoever) that "all robotics" companies need these "three computers". I've got news for you: we don't, this is a fantasy, and limited if any value add will result from what amounts to yet another amorphous simulation, integration and modeling platform based on broken vendor assumptions. Ask anyone experienced in industrial, they'll agree. The industrial vendor space is somewhat broken and rife with all sorts of dodgy things that wouldn't fly in other sectors, but NVidia simply ain't gonna fix it with their current take, which for me lands somewhere between wishful thinking and downright duplicitous.
As for "digital twins", most industrial systems are much like software systems: emergent, cobbled together from multiple poor and broken individual implementations, sharing state across disparate models, each based on poorly or undocumented design assumptions. This means their view of self-state, or "digital twin", is usually functionally fallacious. Where "digital twins" can truly add value is in areas like functional safety, where if you design things correctly you avoid being mired in potentially lethally disastrous emergent states from interdependent subsystems that were not considered at subsystem design, maintenance or upgrade time because a non-exhaustive, insufficiently formal and deterministic approach was used in system design and specification. This very real value however hinges on the value being delivered at design time, before implementation, which means you're not going to be buying 10,000 NVidia chips, but most likely zero.
So my 2c is the Physical AI portion is basically a poorly founded forward-looking application sketch from what amounts to a professional salesman in a shiny black crocodile jacket at a purchased high-viz keynote. Perhaps the other segments had more weight.
"one small step at a time, and one giant leap, together."
5090 for $2000, about $500 more than 4090 when it was announced.
Also nitpick: the opening video said tokens are responsible for all AI, but that only applies to a subset of AI models...
Personally? It's a no for me dawg, DLSS unfortunately doesn't actually replace the need for the raw GPGPU crunch.
For the average layman and a consumer? Nvidia will be selling those GPU's like hotcakes.
DLSS is absolutely the answer for gaming for the masses. DLSS will continue to improve and is far, far cheaper at creating intermediary frames than rendering them for real.
I buy my cards for AI first and gaming second, though. So DLSS is little use to me. 5090 is a bit of a weak improvement over the 4090 for me, but here we are.
In other words, the same performance at somewhere in the range of 2x-3.5x lighter workload. In other words, vastly different performance under same workload. It is veeery on the edge of false advertisement for them to omit the information in parentheses until later in the presentation. :)
That line should have been “4090 frame rates on budget hardware thanks to DLSS 4” or something like that.
Anyway, my point with those lines was to highlight just how defensive AMD was vs how offensive NVIDIA was.
Never trust vendor performance claims (these specifically rely on 3x frame generation), and never assume cards will be available at MSRP.
That's obviously a marketing hoax. Spray DLSS and other tricks and you can make this kind of claims while the raw power is clearly on the side of the 4090
0: https://www.techradar.com/computing/gpu/amd-announces-new-ra...
Update: typical HN behavior. someone is downvoting all my comments one by one...
You cant downvote older comments or comments older than 1 / 2 days.
>Industrial is traditionally all about understanding the problem or breaking it in to unit operations, design, fabricate and control the environment to optimize the process for each of those in sequence, and thus lowering the cost and increasing the throughput.
The thing is, you can do that using SayCan [0] [1]. You give your robot a natural language instruction and that instruction will be broken down to those unit operations and each unit operation is then performed by a robotics transformer model predicting the next action to move the mobile base and end-effector in cartesian space.
What Nvidia is going is just selling you the tools to do it yourself and one of the key parts is adapting simulation video data back into real world camera data. Robotics transformers work just as well as conventional LLMs in practice, but due to the lack of large scale data, they are limited to what they have seen in their training data, with okay performance on unseen tasks. Now here is the thing though, running a robotics transformer in real time is not feasible so far. Even Google's on-robot hardware was running the control loop at 3 Hz, which isn't fast enough.
[0] https://robotics-transformer1.github.io/img/saycan_rt1_demo3...
[1] https://say-can.github.io/img/demo_sequence_compressed.mp4
Yes, of course, TMTOWTDI. Respectfully, you sound like an academic or someone with only software-side experience, having less business side comprehension. It's common also with recent grads who often seem to view robotics through a ROS-influenced lens. Stated succinctly, generalist approaches such as robot arm work cells are usually low throughput, high cost, and have numerous limitations (eg. arms can't lift heavy or large items very well). While programming time may be sold as short, the reality is you're building in to a high capex business an avoidable dependence on a software stack which has a half-life of six months where people are hard to acquire and you could buy an objectively superior solution in terms of throughput, accuracy, spatial efficiency and maintainability without these problems.
Case in point: if you look at the build process for any relatively high complexity modern product (phones, cameras, etc.) they generally avoid using arms because they're relatively slow, weak, mandate large motion envelopes with OH&S implications and must be tediously integrated through vendor-linked APIs. There are better building blocks, and they don't involve tokenized fudge-factors. Yes, you could program them this way, no it generally won't add value. Arms are excellent for things like repeat precision welding (due to the high number of axes of motion offering superior approach and tracking options for target parts and sub-assemblies), but control in such cases comes from a combination of prescriptive part models and real time sensors for physical path following and not from AI fudge.
When you have a retail price so far below "street" price, it just makes it harder to obtain and scalpers take a bigger cut. Raising the price to something more normal at least gives you more of a chance at the big-box store.
Let the market set the price and only control how many cards you auction and when
> If Apple M4 Ultra gets close to 1.8 TB/s bandwidth of 5090
If past trends hold (Ultra = 2x Max) it'll be around 1.1 TB/s, so closer to the 4090.
More tokens/sec on the dual 5090, but way bigger model on the M4.
Plus the dual 5090 might trip your breakers.
Supposedly, this image of a Inno3D 5090 box leaked, revealing 32GB of VRAM. It seems like the 5090 will be more of a true halo product, given the pricing of the other cards.
https://www.techpowerup.com/330538/first-nvidia-geforce-rtx-...
It would not matter if C == Lichtenstein, but China is currently the world's biggest semi-conductor market ...
“Coming soon to auction at a price set by users, no idea what that will be though, good luck!” is much less compelling for consumers trying to plan their purchases in advance.
It's amazing we still get significant improvements in every generation.
128GB of VRAM for $3000.
Slow? Yes. It isn't meant to compete with the datacenter chips, it's just a way to stop the embarrassment of being beaten at HPC workstations by apple, but it does the job.
That said, if you read between the lines, it looks like the new 5090 is about ~40% faster than the 4090 at rasterisation. That’s a solid inter generational improvement.
Buy their datacenter GPU if you need more VRAM.
32GB (for 5090) / 24GB (for 4090) ≃ 1.33
$1500 × 1.33 ≃ $2000
I just think that for home AI use, 32GB isn't that helpful. In my experience and especially for agents, models at 32B parameters just start to be useful. Below that, they're useful only for simple tasks.
a) they are never going to be happy,
b) it's actually a significant step up given the bulk are dual-card users anyway, so this bumps them from (at the high end of the consumer segment) 48GB to 64GB of VRAM, which _is_ pretty significant given the prevalence of larger models / quants in that space, and
c) vendors really don't care terribly much about the home / hobbyist LLM market, no matter how much people in that market wish otherwise.