Qualcomm announces Snapdragon 8 Gen 1 with eight cores and 8k-raw capability – Computer – News

AnandTech has a very extensive article.

Striking, compared to, for example, the Dimensity 9000, is that Qualcomm does take advantage of the option iom two Cortex-A510 merge cores while sharing L2 cache and the NEON/SIMD pipeline. This saves surface area on the chip, but the SIMD performance can be lower. On the other hand, SIMD is often very intensive and still runs on the more powerful cores.

The new ISP (branded as “Snapdragon Sight”) seems like a very serious upgrade. After years of 14-bit color depth, we now see the upgrade to 18-bit. Most camera sensors for smartphones still use 12-bit color depth, and some 14-bit, so it’s not immediately useful for single shots. Computational photography, however, often uses multiple photos aligned and superimposed, for example HDR and night mode. This also creates extra dynamic range, and it’s nice if you can store that somewhere. 18-bit gives more than enough space, and will also ensure that we will see more 14- and maybe even 16-bit sensors in smartphones in the future.

Furthermore, there seems to be even more hardware in the ISP to speed up certain processes, such as:
neural-net controlled 3AA (auto-exposure, auto-focus, auto-white-balance) and face detection block
multi-frame noise reduction and image stacking block
distortion correction block (now also able to correct for chromatic aberrations)
hardware video Bokeh engine (up to 4K video)
Qualcomm specifically mentions that the AI ​​functions run directly on the ISP and do not have to be moved to the Hexagon DSP. This can save latency and increase throughput.

Also: AV1 decoding is still not supported. Netflix, Youtube and many other platforms are investing heavily in AV1. The MediaTek Dimension 1000 supported AV1 decoding already in April 2020, for a 2022 SoC it is really ridiculous that AV1 decoding is not in it. That just means lower quality Netflix and Youtube videos that consume more data, because VP9 streams are used.

Furthermore: Samsung 4nm is used. Samsung 5nm had a density of 134 million transistors per square millimeter (MTr/mm²), while TSMC with 5nm already at 185 MTr/mm² sat. Both Samsung’s 4LPE and TSMC’s N4 are evolutions of their 5nm processes, the chance that Samsung makes up for the almost 40% lag in transistor density is very small. The Dimensity 9000 will therefore have a significant lead, using TSMC’s N4 node.

–