Nvidia kicked off its CES 2025 keynote with a rehashing of the company history, starting with NV1 and arcade machines and evolving into an AI powerhouse. The goal is to apply machine learning to every application possible, all powered by Nvidia GPUs — the house that GeForce built. And that's what we're really here to see: The next generation of GeForce hardware, powered by the Blackwell architecture.
Jumping right into the heart of the matter, Nvidia kicked off the GPU announcements with the RTX 5070 at $549. That will leverage AI in various ways to deliver, according to Nvidia, RTX 4090 performance for one third the price. It will also allow for higher levels of performance in laptops, with a mobile RTX 5070 using half the power of an RTX 4090 while matching it in performance (we're assuming Nvidia is talking about a mobile 4090, though that wasn't immediately clear).
The rest of the announced 50-series lineup consists of the RTX 5090 at $1,999 as the halo part of this generation, with 3,400 AI TOPS of performance. The RTX 5080 will deliver a bit more than half the AI performance at 1,800 TOPS, with a far more attractive price of $999 — inheriting the price of the outgoing RTX 4080 Super. The RTX 5070 Ti comes next at $749 with 1,400 TOPS, and then finally the RTX 4070 with 1,000 TOPS at already noted $549 price point.
It's an impressive start to the next generation GPU announcements, but we need to understand how Nvidia plans to deliver these upgrades. There's also plenty that we don't yet know (officially) about these GPUs. But let's start with what we do know. Many of the core specifications have been listed by Nvidia now, and we've updated the table as appropriate.
| Graphics Card | RTX 5090 | RTX 5080 | RTX 5070 Ti | RTX 5070 |
|---|---|---|---|---|
| Architecture | GB202 | GB203 | GB203 | GB205 |
| Process Node | TSMC 4NP | TSMC 4NP | TSMC 4NP | TSMC 4NP |
| Transistors (Billion) | 92 | ? | ? | ? |
| Die size (mm^2) | 744 | ? | ? | ? |
| SMs | 170 | 84 | 70 | 48 |
| GPU Shaders (ALUs) | 21760 | 10752 | 8960 | 6144 |
| Tensor Cores | 680 | 336 | 280 | 192 |
| Ray Tracing Cores | 170 | 84 | 70 | 48 |
| Boost Clock (MHz) | 2407 | 2617 | 2452 | 2512 |
| VRAM Speed (Gbps) | 28 | 30 | 28 | 28 |
| VRAM (GB) | 32 | 16 | 16 | 12 |
| VRAM Bus Width | 512 | 256 | 256 | 192 |
| L2 Cache | 128? | 64? | 64? | 48? |
| Render Output Units | 240? | 112? | 96? | 64? |
| Texture Mapping Units | 680 | 336 | 280 | 192 |
| TFLOPS FP32 (Boost) | 104.8 | 56.3 | 43.9 | 30.9 |
| TFLOPS FP16 (INT8 TOPS) | 1676? (3352) | 900 (1801) | 703 (1406) | 494 (988) |
| Bandwidth (GB/s) | 1792 | 960 | 896 | 672 |
| TBP (watts) | 575 | 360 | 300 | 250 |
| Launch Date | Jan 2025 | Jan 2025 | Feb/Mar 2025? | Feb/Mar 2025? |
| Launch Price | $1,999 | $999 | $749 | $549 |
You'll note that there are some question marks still, mostly on the GPUs below the RTX 5090 level. Nvidia provided the full specs for the top GPU, but transistor counts and die sizes haven't been provided or hinted at for the other models. (Memory speeds and bandwidth details have been updated; they weren't listed previously.)
Given the stated AI TOPS of performance, the first thing to note is that Nvidia has doubled the AI compute operations relative to Ada Lovelace — at least for INT8 workloads. Our Blackwell RTX 50-series overview has had rumored specifications for a while, and based on the AI TOPS and other specs, we're assuming the rest of the tensor core performance doubled as well.
Nvidia also showed "RTX Blackwell" offering 125 TFLOPS of FP32 graphics compute via the shaders, which is 1.5X more than its Ada generation counterpart, while the AI performance will be 3X as high. But that's for the full-fat GB202 chip, not for the trimmed down solution used in the RTX 5090. There could potentially be a higher spec RTX 5090 Ti or maybe even Titan down the line, or if not that, certainly there's the potential for a higher tier AI / data center part using a fully enabled GB202 die.
So far, none of the announced RTX 50-series parts offers anything beyond standard 16Gb (2GB) GDDR7 modules. That means while the 512-bit interface on the RTX 5090 provides a whopping 32GB of VRAM, the 256-bit interfaces on the 5080 and 5070 Ti still end up with a 16GB GPU, and the 192-bit interface on the 5070 will 'only' provide 12GB of VRAM.
That's generally enough right now for more games, but with the voracious appetite for VRAM that AI workloads have, we wouldn't be at all surprised to see future variants of each GPU paired with 24Gb (3GB) memory modules, providing a 50% boost to VRAM capacity at all levels. Maybe that will be part of the mid-cycle refresh next year.
We don't know if the fifth generation tensor cores used in Blackwell will double the throughput for other number formats. Considering the multipurpose use cases for these GPUs — they'll go into gaming cards, yes, but also professional GPUs and data center AI solutions — we suspect all aspects of the tensor cores got upgraded. We've updated the FP16 performance as appropriate, though the question marks indicate we don't know these figures for certain.
What's interesting is that if we plug in the known clock speeds and core counts, we get most of the final specs. The 125 TFLOPS figure is also accompanied by a maximum 4,000 INT8 TOPS, while the RTX 5090 scales that down to 3,400 TOPS — 85% of a fully enabled GB202, give or take. That 125 TFLOPS figure for a hypothetical chip thus scales down to ~107 TFLOPS on the RTX 5090, which makes sense.
Current rumors put the GB202 at up to 192 SMs, while the RTX 5090 will only have 170 enabled. With a 2.41 GHz boost clock, we also wonder if that's as conservative as the RTX 40-series boost clocks. Most of the 40-series GPUs tend to boost roughly 200 MHz higher than the listed boost clocks, so if things stay the same with Blackwell, we can expect clocks more in the 2.6–2.7 GHz range.
The 1.8 TB/s of bandwidth figure does match up perfectly with the previously rumored 28 Gbps GDDR7 memory running on a 512-bit memory interface. Will the 5080, 5070 Ti, and 5070 also use 28 Gbps GDDR7, or will they clock it a bit higher? Nvidia released the full memory specs now, and the RTX 5080 at least will push beyond 28 Gbps, running at 30 Gbps. The 5070 Ti and 5070 will stick with the same 28 Gbps as the 5090.
What will Nvidia do with double the AI compute on all of its gaming GPUs? Naturally, it plans to have new features and software solutions that will take advantage of the capabilities. With the RTX 5070 offering 1,000 TOPS of compute, nearly the same performance as the RTX 4090 (1,320 TOPS) at one third the price, that opens the doors for more computationally demanding tasks.
One of the most likely use cases will be AI-based texture compression. We've heard about this in the past, and the idea was demonstrated running on previous generation hardware... but not at extreme framerates. Neural Texture Compression (NTC) back in May 2023 ran at less than half the speed of standard BTC (Block Truncation Coding) compression. But 18 months later, with boosted AI compute and more training? It's conceivable that we could have NTC running at the same speed as traditional BTC.
Given the concerns we and others have had with GPUs running out of VRAM on modern games, it's not too surprising that NTC would be one of the major new features of the Blackwell generation of hardware. Boasting higher image quality with one third the memory use, if utilized it could make even 8GB graphics cards far more viable — and the 5070 with 12GB wouldn't be as concerning.
There's just one slight problem: Many games are cross-platform titles that run on consoles powered by AMD GPUs. How many games are going to support Nvidia's new texture compression technology if it requires even just an RTX graphics card? And if it requires an RTX 50-series card, that number will be far smaller. But Nvidia has enough sway to move the gaming market in ways that AMD and Intel can't.
It's also possible that Nvidia could get around this by having an NTC setting in the drivers or as part of DLSS that would work with games, even if they don't explicitly support the feature. That would be the, in our view, ideal situation, because it would mean potential VRAM effective capacity benefits for a whole host of games. We're also curious about whether NTC will be locked to the RTX 50-series, or if it will also be available on other RTX GPUs.
We have a lot of questions, in other words, and we'll find out more in the coming days. This is likely all part of the DLSS 4 family of technologies, which were shown in a game demo during the keynote. Under the broad heading of "neurally rendered graphics," we saw various new technologies at work: DLSS 4, RTX Neural Materials, DLSS: CNN, DLSS: Transformer, Text-to-Animation, and RTX Neural Faces — all with full ray tracing, naturally. However all of these new technologies are enabled, certainly they appear to hold a lot of promise as the demo video looked pretty amazing.
Besides the desktop GPUs, Nvidia also announced the mobile lineup product names. There will be matching RTX 5090, 5080 5070 Ti, and 5070 laptops with availability in March 2025. While the model names match the desktop line, performance will be significantly lower, and we expect the other specifications will also see similar cuts.
The RTX 5090 laptop GPU will offer 1,850 AI TOPS and start at $2,899. That means it's basically equal to the desktop RTX 5080. The mobile 5080 drops to 1,350 AI TOPS, slightly less than the desktop 5070 Ti. Mobile 5070 Ti will have the same 1,000 TOPS as the desktop 5070, and then the vanilla 5070 laptop GPU will offer up to 800 AI TOPS — which is probably a tease of the upcoming RTX 5060 Ti desktop part.
The rest of the keynote, as you might expect, spent a lot of time talking about AI use in all sorts of other areas — vehicles, medical, warehouses, robotics, etc. It's all stuff we've been hearing repeatedly from Nvidia for the past several years, and it's all interesting, but it's not really our core focus. There's so much happening in the realm of AI, and at times it feels a lot like the crypto and NFT hype we were hearing about back in 2020–2021. Except, this time it doesn't appear that we'll see an end to Ethereum mining that will quiet things down.
Nvidia also showed off its new "AI supercomputer" that packs a Grace Blackwell GB10 superchip into a mini PC. Called Project Digits, it runs the full DGX software stack, with 20 Grace CPU cores, 1 PFLOPS of FP4 performance, 128GB of memory, and a 4TB SSD into what should be a more affordable and portable solution than existing DGX servers, as well as something that can easily sit on a desk.
And that wraps up the keynote. The most exciting stuff was obviously the RTX 50-series announcement, and there's still a lot that we still don't know. That will all be revealed in the coming days, and we're anticipating the full RTX 50-series launch to begin before the end of the month. Stay tuned.
