You wanna build a Supercomputer? - NVIDIA NVLink
0 up · 0 down · 0 ratings
Description
So, you want to build a supercomputer. Step number one, get the world's fastest GPU and then get seven more of them. That's no good. You want double the density. So, you're going to be getting liquid cooling. You can't just be managing all of those GPUs. You need them to all be seen as one, which is why you need this network switch and this back plane. Sounds really, really great if it's your NAN server. Jeez. Which is why you need Nvidia's Kyber Midplane. This is a 72 layer PCB which allows you to quadruple the number of GPUs inside of each rack. We've got the NVLink switchblade right here. Those go on either side of your Kyber mid plane for a total of 288 GPUs in your rack and 365 terab of fast memory. But of course, you don't just have one rack. You have racks on racks and racks of GPUs, which is why we have to come over here and switch it. Previously, you'd connect all of your racks together with something like this pitily 51.2 terabit switch. Transceivers consumes 30 watts of power. That's just a bunch of crap. So, instead, we have this 115 terabit switch. IS go directly in here and into what they call their CPO switch system. This allows for not just way more reliability, but it also can eliminate 40 megawws of power if you have a 100,000 GPUs.
So, you want to build a supercomputer. The speaker starts by highlighting the need for the world’s fastest GPU and a set of seven additional units to double the density. He explains that liquid cooling is essential and that all GPUs must be seen as one, which is why a network switch and a back plane are required. The Kyber Midplane, a 72‑layer PCB, quadruples the number of GPUs in each rack, allowing for a total of 288 GPUs and 365 terabytes of fast memory. The speaker then shows how the NVLink switchblade connects to either side of the Kyber midplane, and how a 115‑terabit switch replaces a smaller 51.2‑terabit model, providing more reliability and eliminating 40 megawatts of power for a 100,000‑GPU system.
The comments reveal a strong enthusiasm for the video, with over 10,000 likes and several high‑liked comments praising the system’s capability to run games like Arma 3 and Crysis at high frame rates. Viewers appreciate the detailed explanation of the NVLink switchblade and the Kyber Midplane, noting that the 115‑terabit switch is a significant upgrade over previous models. Some commenters humorously reference gaming and AI markets, while others discuss the potential for high‑quality music and Minecraft performance. Overall, the audience feels the video is informative and engaging, with a clear focus on hardware performance and system design.
Topics · unboxing · review · pc · tech · nvidia · performance · gpu · build
Questions answered
- What is the purpose of the Kyber Midplane in the supercomputer build?
- The Kyber Midplane is a 72‑layer PCB that quadruples the number of GPUs in each rack, allowing for a total of 288 GPUs and 365 terabytes of fast memory.
- How many GPUs can be supported by the NVLink switchblade?
- The NVLink switchblade can connect to either side of the Kyber Midplane, supporting a total of 288 GPUs in a single rack.
- What is the total memory capacity of the system?
- The system provides 365 terabytes of fast memory across all GPUs in the rack.