Consider using smaller chunk size (64^3) and jpeg encoding

[jbms]

You may find 64^3 chunks provide better visualization performance than 96^3, since for cross section views you end up loading 50% more data for the initial display with 96^3 chunks. You may also find it helpful to use jpeg compression for uint8 data --- I see some volumes are using n5 with gzip compression --- I know we already discussed about the lack of n5/zarr jpeg compression support in Neuroglancer and tensorstore, though.

[d-v-b]

I admit I picked 96^3 somewhat arbitrarily in an attempt to balance loading / visualization performance against the cost of a lot of tiny objects in storage. I should do a more thorough comparison.

As for datatypes, I think all the raw uint8 data are stored with jpeg compression in precomputed format, albeit unsharded (until tensorstore has a better API for saving sharded volumes with parallelism over shard files :) ). For some of the analysis volumes (predictions) I opted for the lossless compression in case someone wants to download those volumes for re-analysis.

jpeg compression for n5/zarr would be great, I have a stalled PR that would enable this for n5: zarr-developers/zarr-python#577. As soon as that gets some traction I will probably open an issue / PR on neuroglancer to support this.