CES 2025 is underway here in Las Vegas and I’m on the ground covering all the latest computing developments, including the highly anticipated announcement of Nvidia’s GeForce RTX 5000 series graphics cards, including the GeForce RTX 5090.
It’s been a little over two years since Nvidia released its last flagship desktop GPU, so anticipation has been building for quite some time now.
Fortunately, we won’t have long to wait before we get out first look at what Nvidia has in store for 2025, and I’ll be here bringing you all the announcements as they happen from Mandalay Bay’s Michelob Arena in Las Vegas.
How to watch the Nvidia CES 2025 keynote
While I encourage you to follow along with my live report, you can also watch Nvidia CEO Jensen Huang’s keynote presentation along with me at the YouTube link embedded below.
OK folks, we’re coming up on the 15 minutes from the start of Nvidia’s CES 2025 keynote, where CEO Jensen Huang will take the state at Madalay Bay’s Michelob arena. We’re expecting some major news tonight, so for those who’ve been waiting to hear about Nvidia’s next-gen consumer graphics cards, you don’t much longer to wait.
And if you really want to hear all about data center AI and Omniverse stuff, I’m sure Jensen will get around to those as well.
The biggest thing I’m expecting tonight is Nvidia’s GeForce RTX 5090, RTX 5080, and possibly the RTX 5070 Ti.
Following up Nvidia’s Lovelace GPUs, the Blackwell-based RTX 5000 series is expected to be substantially more powerful, with the rumor mill putting the RTX 5080 around 10% faster than the RTX 4090, currently the best graphics card on the consumer market.
That, of course, would put the RTX 5090 in a class entirely on its own, and there’s no telling where its performance will ultimately end up. That said, if speculation is on the mark, it should feature 32GB GDDR7 VRAM with a memory bandwidth of 1.52TB/s on a 512-bit memory bus, making it truly the world’s first 8K gaming graphics card.
OK, here we go.
OK, I DO want an exoskeleton. Those things look cool as hell.
CTA President Gary Shapiro is introducing Jensen Huang.
OK, I unironically love ‘Never gonna give you up.’ I used to rollerskate to that song as a kid.
OK, the Nvidia segment of the keynote is about to begin, but it sure is taking a while. Nvidia is normally quicker to launch than this.
OK, NOW we’re kicking things off.
Tokens, tokens, tokens. It’s no surprise that we’re just jumping right into AI, but yeah, it’s remarkable how much Nvidia has transformed almost overnight.
LOL, Jensen’s jacket is bedazzled.
That was a very short Virtua Fighter demo.
OK, the first mention of GeForce, so here we go.
That was a pretty impressive demo.
Nvidia RTX Blackwell is official, and that is a very pretty looking graphics card.
OK, so Jensen is holding the RTX 5090.
OK, RTX 5070 at $549, RTX 4090 performance. Whoa.
OK, RTX 5090 starting at $1,999. RTX 5080 for $999. RTX 5070 Ti for $749. Yes, absolutely. This is what I want to see.
Don’t get me wrong, these are still expensive graphics cards, but given the fears of a $1600 RTX 5080, this is a very pleasant surprise.
OK, so very little on specs, but I want to know about this AI management processor.
If the shader cores can also carry the weight of AI workloads, as Jensen stated, then we’re getting way better DLSS on these cards.
Also, the RTX 5000 series will be available starting in January, though we don’t know which will be coming first.
Nvidia GeForce RTX 5000 series mobile GPUs are also coming, with the RTX 5070 mobile featuring RTX 4090 performance, though I’m guessing Jensen means RTX 4090 mobile performance.
So we have fully entered into the data center segment of the keynote. While GeForce graphics cards got at least a bit more time and attention than Lovelace got, it’s clear that these cards aren’t as important to Nvidia as the data center business.
And yeah, that shield bit was a bit…well, it was something.
I want to say, I find all of this AI discussion interesting from an academic perspective, but I think there’s a lot of expectations for these data centers, and no one is mentioning that the power requirements for these are pretty much going to put a cap on what they can do, since we only have so much electrical power available on a grid at any one time.
Oh man, I just had a dark thought. Can you imagine training the AI agent that’s taking your job? That’s grim.
Coding assistants are the death knell for the junior software developer. So much for ‘learn to code’.
OK, the virtual human thing is still giving major uncanny valley, but it’s less severe than it used to be.
OK, sorry about that folks, we were dealing with some technical difficulties, but we’re back to the action, and that action is all about tokens. It’s tokens all the way down.
OK, so the end of data for training models is another major bottleneck for AI, and what Jensen is talking about here with Cosmos is generating new data that subsequent models can be trained on (synthetic data), since these models have already consumed all the existing data it could be trained on.
However.
I wonder how Cosmos will avoid model collapse.
The worry with synthetic data is that you’ll end up with a Habsburg AI, one that’s effectively inbred on its own data to the point where it becomes a useless abomination. Google the Habsburg monarchs of Europe if you want to see why this is such an apt description of the problem.
OK, we just got word on the Blackwell GPU availability.
The RTX 5090 and RTX 5080 will go on sale on January 30, 2025, for $1,999 and $999, respectively.
The RTX 5070 Ti and RTX 5070 will be available in February for $749 and $549 respectively. UK and Australia pricing wasn’t given, but we’ve reached out to Nvidia for clarification.
Nvidia RTX 5090, RTX 5080, and RTX 5070 Ti laptops will be available starting in March, while the RTX 5070 laptops will be available starting in April.
Nvidia’s autonomous vehicle chip would get annihilated in NYC traffic, I can guarantee that, though it might work in a lot of other cities.
Hmmm. Sythetic input data has been shown to quickly degrade the quality of the model you’re training (model collapse) which i haven’t heard mentioned once. I wonder how nvidia plans on tackling that problem.
I do have to say, if I have to deal with AI model collapse in the product I’m using, I’d really rather not have to deal with that at highway speeds.
And that’s a wrap on Nvidia’s CES 2025 keynote. Now, if you’ll excuse me, I have to go bother Nvidia PR to try and track down some spec sheets. Stay tuned for more from CES 2025 thoughout the week, including more details on the Nvidia Blackwell GPUs.
