Nvidia gives RTX 40 series a Super refresh as AI PC hype takes off
Is it a bird? Is it a plane? No, it's GPUs with slightly better specs
CES Nvidia revamped its RTX 40 series GPU lineup at CES on Monday with three "Super" cards, which boast higher performance and memory upgrades over their predecessors.
Nvidia's $599 RTX 4070 Super and $799 4070 TI Super see some of the biggest improvements.
Digging into the spec sheet, the somewhat confusingly named 4070 Ti Super — there are now four 4070 SKUs to keep track of — features a 300 MHz higher base clock and 768 more CUDA cores compared to the standard 4070 Ti.
The biggest improvement actually involves memory. The 4070 Ti Super packs 16 GB of GDDR6x memory attached to a wider 256-bit bus. Four extra gigabytes of memory might not sound like much to write home about, but as on-device generative AI applications grow more prevalent, more memory is essential for running larger or higher precision models.
For instance, a 16 GB card is a minimum amount of memory required to run Meta's LLama 2 13B model without resorting to tricks. The larger memory bus, however, should help to speed up model response times things like chatbot responses or image generation.
Turning into the least expensive of the new cards, the 12 GB RTX 4070 Super doesn't get more VRAM or a wider memory bus, but it does feature 20 percent more CUDA cores than the original 4070, topping out at 7,168. For reference, that puts the card within spitting distance (512 cores) of the RTX 4070 Ti.
Nvidia hasn't shown how these cards perform compared to their non-super siblings as yet, preferring instead to draw comparisons to its older hardware.
- Nvidia readies downgraded chips for China, but will anyone want to buy them?
- Apple sets new 16,000-foot iPhone drop test after 737 fuselage fail
- Avoiding AI-capable PCs will be impossible by 2027
- AMD brings its AI engines to the desktop with Ryzen 8000G APUs, RX 7600 XT graphics cards
If Nvidia is to be believed, and we recommend you take these claims with a grain of salt, the RTX 4070 Ti Super is about 60 percent faster than the 3070 Ti without relying on AI upscaling or frame generation trickery. Meanwhile, Nvidia says the less powerful RTX 4070 Super is comparable in terms of performance to the RTX 3090.
Both cards are slated to hit shelves later this month, beginning with the RTX 4070 Super on January 17, followed by the 4070 Ti Super on January 24.
RTX 4080 Super gets a marginal performance boost
At the top of the heap is the $999 RTX 4080 Super. The 16 GB GPU, while the most powerful of the updated cards, is also the most modest, boasting five percent more CUDA cores at 10,240 up from 9,728 on the original 4080 and a 1.6 percent higher boost clock. However, beyond this the card appears to be largely unchanged.
In terms of performance, Nvidia says you can expect about 40 percent higher performance than the much older 3080 Ti even with its third-gen Deep Learning Super Sampling (DLSS 3) upscaler turned off. Switch it on, however, and team green claims the newer card can push twice as many frames as the last-gen card.
The good news is Nvidia isn't charging a premium for the extra CUDA cores or megahertz of frequency. In fact, the card has a suggested retail price of $999, about $200 less than the RTX 4080 when it launched in late 2022.
Of course, those prices are just that, a suggestion. We'll have to wait until the card hits store shelves on January 31, to see what partners actually charge for the 4080 Super.
While more powerful, the 4080 Super still comes up short of Nvidia's flagship 4090 with its 24 GB of vRAM, 384-bit memory bus, and whopping 16,384 CUDA cores. For its successor, it appears we'll have to wait for the launch of Nvidia's Blackwell architecture later this year for more details. ®