Nvidia claims 1000x graphics power over 5 years but 4060 Ti isn’t even 1.1x
As Nvidia shows off massive servers and bigger GPUs, the 1000X graphical performance claim doesn't apply to every GPU, clearly
Share this article...
WePC is reader-supported. When you buy through links on our site, we may earn an affiliate commission. Prices subject to change. Learn more
Nvidia had a lot of enterprise hardware to show off in its Computex 2023 keynote. And in it, Jensen claimed that Nvidia has improved 1000x in five years unironically a week after the launch of the RTX 4060 Ti, which hasn’t exactly even improved over the previous generation and its predecessor.
The 4060 Ti reviews and benchmarks have shown how not that much improvement comes with the new generation of GPU. So the 40 series card only improves by a maximum of 10FPS over the 30 series card. It also diminishes over the higher resolutions and makes it less future-proof even with the limited memory bandwidth. And at some tests, it even falls behind as well, so how can Nvidia claim such improvements?
Well, the claim is more likely aimed toward the enterprise GPUs instead and what they can achieve in AI and ray tracing. As it has announced the production of the Grace Hopper chip and all-in-one power computer. This can turn into a true supercomputer in one that allows for combination to create great power and efficiency for large language models as the focus of its talk, and in fact, becoming more like Apple with its ecosystem.
Grace Hopper superchip the AI supercomputer
With the new platform, Nvidia has provided a whole solution to supercomputers. With the Grace Hopper platform, it is an all-in-one option to do so. The reason for Nvidia’s attempted failed ARM acquisition, it combines an ARM based Grace CPU and Nvidia Hopper GPU. Offering improved efficiency that can be combined with plenty of others to create a whole network
The platform pushes Infiniband and Nvlink to combine these processors into a whole network. With the ability to connect up to 256 GPUs, that are connected via the Nvlink 900 GB/s connection they can share the memory and not split the workload and rather share it out for more efficient processing.
With each GPU connected to 96 GB of fast HBM 3 memory, with the CPU with LPDDR5X up to 512 GB, there is a lot of memory available on them as well that can be split amongst the one giant supercomputer.
This is where the power is and might explain the claim of the giant graphical power increase from Nvidia. But it is limited to custom solutions and definitely not intended for gaming, since the solution is aimed at those large language models such as powering the next AI works. And the big Nvidia Omniverse capabilities that are aiming to improve work with AI with the likes of ACE.
