Skip to content

Large volumes and browser caches

Neuroglancer and Viv have different volume-loading models. Their memory figures are estimates or engine cache accounting, never a measurement of free VRAM.

Engine guidance

Neuroglancer's API provides separate gpuMemoryLimit, systemMemoryLimit, concurrentDownloads and adaptive prefetch controls. It loads chunks and can display a volume whose full uncompressed size exceeds its resident cache. A ready view does not mean that the complete dataset has downloaded. The viewer reports queued/downloading chunks and allocated GPU chunk bytes separately while further detail loads. The compact streaming message is at the bottom of the image so partially loaded volumes remain visible and interactive. Clipping controls use smaller aligned rows and one-based plane positions at the selected quality instead of percentages.

Viv's volume loader assembles a contiguous volume per channel. Its 3D layer converts data to Float32 textures. For uint16, the full texture estimate is twice the uncompressed source size, before other GPU allocations. Whole-volume texture dimension limits still apply to Viv.

Quality therefore shows full typed data size for streamed Neuroglancer sources and estimated full GPU texture size for Viv. Neuroglancer permits manually selecting an above-budget level, with an explicit slow-loading/crash warning; automatic selection stays within budget. A higher level number is coarser.

Chrome's per-tab GPU column is not a reliable total for a volume: Chrome also uses a shared GPU process. Windows Task Manager includes allocations beyond this viewer. Neither number is directly comparable to the full-level estimate. See Chromium's GPU architecture.

Cache lifetime and loading feedback

Switching between 2D and Neuroglancer retains the existing scene and its CPU/GPU chunk caches. New downloads and prefetching pause while that scene is hidden. Returning restores the current navigation; changing source/field or switching to Viv disposes the scene. Retention uses memory, and cache eviction can still require new downloads. A page reload discards these in-memory caches.

Viv's packed-volume cache follows the selected budget, capped at 1 GiB of CPU memory. This is independent of GPU texture storage and preserves all three channels of the astrocyte level-4 volume (about 478 MiB total). The existing 200 MiB per-channel server preparation guard remains in force. Larger levels continue through the regular loader. A single volume overlay shows preparation, download and texture-upload stages; an error replaces the progress overlay.

Sharded Zarr and 2D tiles

Partial range reads bypass Chrome's HTTP disk cache and use a private 32 MiB, 60-second byte cache keyed by URL and Range. This avoids reproduced ERR_CACHE_OPERATION_NOT_SUPPORTED failures after reload/panning. Metadata stays fresh, errors are not cached, and a source/context change clears range entries.

Storage requests are bounded through body consumption, not just response headers. Transient fetch/body failures have bounded retries; cancellation is not an error to retry. Shared small shard-index reads use the image lifetime rather than the first tile's abort signal, so cancelling one tile cannot poison other tile reads.

Large physical images with small logical chunks use 256px screen tiles without changing stored chunks, pixel values or calibration. Native-size statistics tiles remain separate to avoid increasing automatic-contrast I/O. Missing chunks retain the Zarr fill-value semantics; genuine transfer/decode errors are not replaced with zeros.

Exact 2D planes and larger decoded 3D blocks

For large physical sources with small logical chunks, the optional signed raster_url capability enables a bounded pixel-tile API. A 256px screen tile previously required 64 logical 32px chunk reads per channel, including unused Z planes from each 3D chunk. The server now reads the requested XY plane from the original array and returns exact typed pixels. There is no conversion to RGB, contrast baking, resampling or source rewrite. Browser decoding and image-binding cache rules remain unchanged; four plane transfers may run concurrently. Statistics still sample small native tiles. Older servers without this capability, HTTP-Zarr sources and other chunk layouts keep their existing routes.

Large physical v3 Neuroglancer intensity sources use bounded 128-cubed decoded blocks instead of 64-cubed blocks (up to 8 MiB at float32). This amortizes HTTP and shard-index work. Preparation no longer walks all chunk files for uncached intensities: the read-only storage adapter validates each requested metadata or chunk path against the signed mount/routes, including partial reads and symlinks. Label fingerprint/cache rules, native v2 streaming, HTTP providers and their streaming grids remain unchanged. Zarr-Python 2 and 3 are supported.

A local bounded comparison on image 20401 at x=1588.89, y=4173.02, zoom=-0.503, z=700 used fresh browser contexts and the same 12-second post-controls window. Before: 1,010 completed requests / 26.5 MiB, still showing the blurred pyramid. After: 224 / 14.9 MiB, with the requested detail tiles complete about 6.8 seconds after navigation (controls appeared at 2.7 seconds). This is a local observation, not a guaranteed load time. No full high-resolution volume was loaded. In the same short Neuroglancer level-2 switch probe, the new route showed partial 3D content and about 340 MiB of GPU chunk cache within six seconds; the old route was still preparing. The v3 grid alone does not imply an eight-fold speedup.

Local investigation fixture

IDR0048 image 9846151, the OME-Zarr 0.5 astrocyte sample, has three uint16 channels, 1402 Z planes and XY-only downsampling. Logical chunks are 32 cubed inside 32 x 512 x 512 shards. Its level 2 contains about 7.48 GiB of raw channel data; Viv's full Float32 textures would require about 14.97 GiB. This dataset advertises pixel coordinates, not a verified micrometer calibration. It advertises no NGFF segmentation labels.

The persistent-Chrome reproduction previously recorded 52 non-cancellation fetch failures while reloading/panning. The corrected range path recorded zero in the same profile. Tests use bounded 2D and coarse 3D probes; they do not wait for full high-resolution volume loading. Public data and machine-specific probes are not included in automated repository tests. Synthetic tests cover exact v2/v3 edge pixels, cancellation, range identity, cache reuse, engine memory models and real WebGL section/clipping rendering.