3 ms·
Here is an interesting thought experiment. With PCI-7.0 coming in 2025/2026, product launch by 2028. And assuming no further bottleneck, we assume even a linear
by ksec 4y ago
Here is an interesting thought experiment. With PCI-7.0 coming in 2025/2026, product launch by 2028. And assuming no further bottleneck, we assume even a linear increase of memory bandwidth would be handled with HBM3 or HBM4. We should be able to hit 6.4Tbps by 2030. In a single 1U unit. With next generation codec VVC, or even a generation beyond that. Serving an average 10Mbps files, a single box is capable of serving 640K customers. A single Rack could serve 25 Million Customer. i.e 10 Rack would be able to serve all of the current Netflix subscribers world wide, simultaneously.
Although I would assume the CPU being the bottleneck with 8x the required processing power.
- ksec 4y agoThis got me thinking, at what point would the benefits of these optimisation are no longer meaningful. At 225Million subscriber, assuming somehow, it has a peak concurrent user of 70% total, around 160M. This figure should already be way above the actual number because they are spread across the Globe in different Time Zone. At 10Mbps Netflix would need a peak bandwidth of 1600M Mbps, or 1600Tbps. At 1.6Tbps that is only 1000 box required worldwide. Or 2000 if you include redundancy. Once you factor in you need a box with every ISP, Local Peer Exchange or Regional ISP worldwide, and there are hundreds of those. Shrinking it further wouldn't make much of a different. Hopefully that means Netflix now has the incentive to bump bitrate to 20Mbps or higher.