https://petapixel.com/2023/06/12/sphere-studios-big-sky-cine...
Bits like a fiber-to-media recording system and custom GPU-based processing systems, really awesome project
https://petapixel.com/2023/06/12/sphere-studios-big-sky-cine...
Bits like a fiber-to-media recording system and custom GPU-based processing systems, really awesome project
Big Sky is shooting 120FPS at 10-bit (or 60FPS at 12 bit). That's 408Gbps/51GBps.
A RX 7700 XT has 432Gbps memory bandwidth and a 4070 504Gbps... maybe that could render this? Unless you need to both write to memory and then read from it to render, and now you are up to a flagship RX7900 XTX or 4090 (960/1008Gbps). Which with their 24GB memory can fit a whopping 0.47s of video in memory.
Now consider the NVMe storage array you ll need to get video at 408Gbps. PCIe 5.0 NVMe drives can supposedly hit 14GBps, so "only" 4 drives going full tilt should be adequate to let you play this back at realtime speeds.
Here's a pro-level card, for instance: https://www.amd.com/en/products/professional-graphics/amd-ra...
That says 864 GB/s, which is pretty good.
So any of these would be more than fine. That makes a lot more sense, considering what it takes to make a complex 3d scene.
That's not that impressive, Blackmagic cameras have been able to run on fiber since 2018 [1].
The bandwidth for the Sphere cam however, that is insane. 500 Gbit/sec, that used to be the domain of entire ISP's core switches, and now that kind of performance is (effectively) handheld-capable.
[1] https://www.blackmagicdesign.com/media/release/20180201-02
It’s an almost unfathomable number but at the same time it is well within the realm of commodity hardware. We are living in the future.
And then you realize that the latest datacenter Ethernet switch chips have 51 Tbit/sec agregate non-blocking capacity (128 ports of 400Gb/s, or similar setups), and a single AI cluster can use hundreds of these switches.
That's insane.