NVIDIA CEO Jensen Huang has officially announced the Hopper graphics architecture that will power the company’s next-generation products. These GPUs, which have 80 million transistors and come out of 4nm production technology, will use the new generation HBM3 memory. Here are all the details…
What does the NVIDIA Hopper architecture promise to gamers?
NVIDIA’s RTX 20 series, aka Turing architecture GPUs, are manufactured at TMSC, a favorite of Apple, Intel, and nearly every other tech company. But for Ampere (the architecture that replaced Turing) GPUs in the GeForce RTX 30 series, the company moved to Samsung’s 8nm production line. However, they showed a great performance increase. As a result, power consumption has also increased.
But with the NVIDIA Hopper architecture, it turns to TMSC again. The biggest reason for this is that this chip will be built using a special version that comes out of the less power-consuming 4nm band instead of TSMC’s known 5nm production line.
AMD’s RDNA 2-based Radeon RX 6000 series GPUs surpassed NVIDIA in low power consumption last year. But the biggest reason was that the production technologies were different. Therefore, since TMSC will produce the chips of both parties next year, it will now be possible to make a clearer comparison.
In addition, AMD became the first devices to use PCIe 4.0 connectivity technology with Radeon RX 5000 series graphics cards. But NVIDIA intends to go one step further with PCIe 5.0. In addition, Intel paved the way for NVIDIA to take this step, as Intel recently brought PCIe 5.0 support to the desktop computer with its 12th generation Core processors.
The Hopper architecture will also bring NVIDIA’s high-bandwidth fourth-generation NVLink technology. The company had restricted multi-GPU usage to GeForce RTX 3090 cards in the RTX 30 series. Considering that this support will continue in the future, we can see a big speed increase in NVLink 4.
GPUs to be used in the data center are also important. At this point, NVIDIA says Hopper’s optimization work has been completed and they’ve seen the best results so far. But he says it’s too early to share any information about CUDA core counts, clock speeds, or GPU die sizes.
We saw that in the RTX 30 cards released last year, a maximum of 54.2 billion transistors can be reached. However, GPUs will be able to reach 80 billion transistors very easily thanks to the new TMSC 4nm technology. Of course, a lot of power will be required to use these transistors efficiently at high performance. Therefore, we should not be surprised if we see 700W TDP values.
Already in the past, it was said that NVIDIA will use 1×16 pin power input instead of 2×8 in the new generation RTX 40 cards. In this way, it will be able to draw up to 600W of power over a single connection. In addition, it can be supported up to 75W via the motherboard if needed. However, when considering the power requirement of overclocked cards and large coolers from other manufacturers such as MSI and ASUS, a 2nd pin input will be needed.
What do you think about this subject? Don’t forget to share your views with us in the comments!