GPU-accelerated SCENIC

Gregor Lueg

2026-09-24

Intro

This vignette walks through the GPU-accelerated version of the SCENIC gene regulatory network inference in bixverse. If you have not read the CPU SCENIC vignette first, do so. That vignette explains the biology and the CPU code path. This one focuses on what the GPU version buys you and how the API differs.

Behind scenic_grn_sc_gpu() sits a wave-scheduled, multi-tree GPU driver that dispatches ExtraTrees and RandomForest fits to the wgpu backend via cubecl. GRNBoost2 stays on the CPU because gradient boosting doesn’t map cleanly onto the same multi-output kernel design.

The bixverse.gpu package provides:

• scenic_grn_sc_gpu() for SingleCells and MetaCells objects. Streaming and non-streaming drivers on the SingleCells path; a single in-memory driver on the MetaCells path.
• A single GPU-specific knob, wave_byte_budget, that caps VRAM use of the per-wave histogram tensors. Default is 4 GiB. Shrink on 8 GB adapters shared with other workloads; raise on 16 GB+ adapters to let the scheduler pick a wider wave.

Everything else, the params, the target-gene shortlisting, the downstream motif enrichment, is imported straight from bixverse. The GPU version reuses the same params_scenic(), the same scenic_gene_filter_sc(), the same ScenicGrn result class, so downstream code (identify_tf_to_genes, tf_to_genes_correlations, tf_to_genes_motif_enrichment) is identical.

library(bixverse)
library(bixverse.gpu)
library(data.table)
#> Warning: package 'data.table' was built under R version 4.5.2
library(magrittr)
#> Warning: package 'magrittr' was built under R version 4.5.2

Note

Vignettes were built locally on a MacBook Pro M1 Max. The GH runners were just too slow and do not have proper GPU support. This gives an idea of speed on a decent, but older machine.

Rebuilding a processed PBMC3k object

We reconstruct the same SingleCells object used in the CPU SCENIC vignette. The chunk is folded because we care about SCENIC, not QC.

Rebuild the PBMC3k object (click to expand)

pbmc3k_path <- download_pbmc3k()
tempdir_pbmc <- file.path(tempdir(), "gpu_scenic")
dir.create(tempdir_pbmc, showWarnings = FALSE, recursive = TRUE)

sc_object <- SingleCells(dir_data = tempdir_pbmc)
mtx_io_params <- get_cell_ranger_params(pbmc3k_path)

sc_object <- load_mtx(
  object = sc_object,
  sc_mtx_io_param = mtx_io_params,
  mtx_streaming = FALSE,
  .verbose = FALSE
)
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.

setnames_sc(sc_object, table = "var", old = "column1", new = "gene_symbol")
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.

var <- get_sc_var(sc_object)
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.
ensembl_to_symbol <- setNames(var$gene_symbol, var$gene_id)
symbol_to_ensembl <- setNames(var$gene_id, var$gene_symbol)

gs_of_interest <- list(
  MT = var[grepl("^MT-", gene_symbol), gene_id],
  Ribo = var[grepl("^RPS|^RPL", gene_symbol), gene_id]
)
sc_object <- gene_set_proportions_sc(
  sc_object,
  gs_of_interest,
  streaming = FALSE,
  .verbose = FALSE
)
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.

qc_df <- sc_object[[c("cell_id", "lib_size", "nnz", "MT")]]
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.
metrics <- list(
  log10_lib_size = log10(qc_df$lib_size),
  log10_nnz = log10(qc_df$nnz),
  MT = qc_df$MT
)
directions <- c(
  log10_lib_size = "twosided",
  log10_nnz = "twosided",
  MT = "above"
)
qc <- run_cell_qc(
  metrics = metrics,
  cells_to_keep = get_cells_to_keep(sc_object),
  directions = directions,
  threshold = 3
)
sc_object[["outlier"]] <- qc$combined
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.
cells_to_keep <- qc_df[!qc$combined, cell_id]
sc_object <- set_cells_to_keep(sc_object, cells_to_keep)
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.

sc_object <- find_hvg_sc(sc_object, hvg_no = 2000L, .verbose = FALSE)
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.
sc_object <- calculate_pca_sc(sc_object, no_pcs = 30L, .verbose = FALSE)
sc_object <- find_neighbours_sc(
  sc_object,
  neighbours_params = params_sc_neighbours(knn = list(k = 15L)),
  .verbose = FALSE
)

Gene filtering

Same as the CPU workflow. The gene filter is a cheap min-counts / min-cells scan; there’s no GPU version because there’s no need for one.

scenic_genes <- scenic_gene_filter_sc(
  object = sc_object,
  scenic_params = params_scenic(min_counts = 50L),
  .verbose = FALSE
)

length(scenic_genes)
#> [1] 5430

Transcription factor list

Aerts lab curated list, mapped to Ensembl.

tf_dt <- data.table::fread(
  "https://resources.aertslab.org/cistarget/tf_lists/allTFs_hg38.txt",
  header = FALSE,
  col.names = "tf"
)
tf_dt[, gene_id := symbol_to_ensembl[tf]]
tf_dt <- tf_dt[!is.na(gene_id)]

nrow(tf_dt)
#> [1] 1100

GRN inference on the GPU

Same signature as scenic_grn_sc() from bixverse, with the extra wave_byte_budget knob. Everything else, scenic_params, tf_ids, genes_to_take, works the same way.

ExtraTrees

extratrees is the recommended default on GPU. Random split thresholds mean each level dispatches a single split-scoring kernel, and the driver saturates the device with wide waves.

scenic_res_et <- scenic_grn_sc_gpu(
  object = sc_object,
  tf_ids = tf_dt$gene_id,
  genes_to_take = scenic_genes,
  scenic_params = params_scenic(
    learner_type = "extratrees",
    gene_batch_size = 64L
  ),
  .verbose = TRUE
)
#> SCENIC GPU: 5430 target genes, 466 TFs, 2163 cells (streaming: FALSE)

scenic_res_et
#> ScenicGrn (GRN results)
#>   No genes:                 5430 
#>   No TFs:                   466 
#>   Applied learner:          extratrees 
#>   TF to gene generated:     FALSE 
#>   CisTarget res generated:  FALSE

RandomForest

randomforest uses exhaustive split evaluation. It’s a bit slower per tree than ExtraTrees but often lands on tighter importance rankings.

scenic_res_rf <- scenic_grn_sc_gpu(
  object = sc_object,
  tf_ids = tf_dt$gene_id,
  genes_to_take = scenic_genes,
  scenic_params = params_scenic(
    learner_type = "randomforest",
    gene_batch_size = 64L
  ),
  .verbose = TRUE
)
#> SCENIC GPU: 5430 target genes, 466 TFs, 2163 cells (streaming: FALSE)

scenic_res_rf
#> ScenicGrn (GRN results)
#>   No genes:                 5430 
#>   No TFs:                   466 
#>   Applied learner:          randomforest 
#>   TF to gene generated:     FALSE 
#>   CisTarget res generated:  FALSE

GRNBoost2

Not available on GPU. The wrapper fails fast with a clear message before touching the device.

scenic_grn_sc_gpu(
  object = sc_object,
  tf_ids = tf_dt$gene_id,
  genes_to_take = scenic_genes,
  scenic_params = params_scenic(learner_type = "grnboost2"),
  .verbose = FALSE
)
#> Error in `method(scenic_grn_sc_gpu, bixverse::SingleCells)`:
#> ! GRNBoost2 (gradient boosting) is not supported on GPU. Use bixverse::scenic_grn_sc() for the CPU implementation.

Fall back to bixverse::scenic_grn_sc() if you need it.

Streaming vs non-streaming

On disk-backed SingleCells objects both drivers exist. Non-streaming loads the full sparse target column vector into host RAM up front and dispatches wave by wave. Streaming reads targets in I/O chunks and dispatches per chunk, capping peak host memory at roughly one chunk of sparse columns.

Set streaming = FALSE to force the non-streaming path when host RAM isn’t the limiting factor:

scenic_res_streaming <- scenic_grn_sc_gpu(
  object = sc_object,
  tf_ids = tf_dt$gene_id,
  genes_to_take = scenic_genes,
  scenic_params = params_scenic(
    learner_type = "extratrees",
    gene_batch_size = 64L
  ),
  streaming = TRUE,
  .verbose = TRUE
)

Leave streaming = NULL to let the same auto_streaming() heuristic that the CPU version uses pick for you.

TF-to-gene refinement

Because the result is a ScenicGrn object, the CPU downstream steps drop straight in.

scenic_res_et <- identify_tf_to_genes(
  scenic_res_et,
  n_sd = 1,
  .verbose = TRUE
)
#> Extracting TF to gene associations via per-gene threshold (mean + 1.0 * SD).

scenic_res_et <- tf_to_genes_correlations(
  x = scenic_res_et,
  object = sc_object,
  cor_filter = 0.01,
  .verbose = TRUE
)
#> Calculating the pairwise correlations between the TFs and genes
#> Removing TF <> gene pairs with cors <= 0.010
#> Removing self loops (TF controlling its own expression

tf_gene_dt <- get_tf_to_gene(scenic_res_et)
tf_gene_dt[, tf_symbol := ensembl_to_symbol[tf]]
tf_gene_dt[, gene_symbol := ensembl_to_symbol[gene]]

head(tf_gene_dt[order(-importance)], 5L)
#>                 tf            gene importance pairwise_cor tf_symbol
#>             <char>          <char>      <num>        <num>    <char>
#> 1: ENSG00000139187 ENSG00000113088  0.2567620    0.2531497     KLRG1
#> 2: ENSG00000171223 ENSG00000120129  0.2535830    0.3743879      JUNB
#> 3: ENSG00000139187 ENSG00000161570  0.2012684    0.3895632     KLRG1
#> 4: ENSG00000138795 ENSG00000166681  0.2005688    0.1963190      LEF1
#> 5: ENSG00000221869 ENSG00000115828  0.1988880    0.2871167     CEBPD
#>    gene_symbol
#>         <char>
#> 1:        GZMK
#> 2:       DUSP1
#> 3:        CCL5
#> 4:     NGFRAP1
#> 5:        QPCT

MetaCells path

The GPU driver also dispatches on MetaCells. Counts already live in memory as a sparse CSC block, so the streaming flag has no effect (and is silently ignored). Build the metacells the same way as the CPU vignette:

mc_object <- generate_bt_meta_cells_sc(
  sc_object,
  sc_meta_cell_params = params_sc_bt_metacells(
    target_no_metacells = 200L
  ),
  .verbose = FALSE
)
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.
#> duckdb is storing downloaded extensions and secrets under ~/.duckdb:
#> ℹ /Users/gregorlueg/.duckdb
#> This persists across sessions and is shared with the DuckDB CLI and other clients.
#> ℹ Run duckdb(shared_home = FALSE) to use a temporary directory instead.
#> ℹ See ?duckdb_storage for details and alternatives.

scenic_res_mc <- scenic_grn_sc_gpu(
  object = mc_object,
  tf_ids = tf_dt$gene_id,
  scenic_params = params_scenic(
    learner_type = "extratrees",
    min_counts = 50L,
    min_cells = 0.02,
    gene_batch_size = 64L
  ),
  .verbose = TRUE
)
#> The mean leafs per sample is set quite high for meta cells. Reducing to 10L.
#> No target genes supplied, running CPU gene filter...
#> SCENIC gene filter: 7566 / 11139 genes pass.
#> SCENIC GPU: 7566 target genes, 703 TFs, 200 cells

scenic_res_mc
#> ScenicGrn (GRN results)
#>   No genes:                 7566 
#>   No TFs:                   703 
#>   Applied learner:          extratrees 
#>   TF to gene generated:     FALSE 
#>   CisTarget res generated:  FALSE

Same result class, same downstream methods. identify_tf_to_genes(), tf_to_genes_correlations() and tf_to_genes_motif_enrichment() all work unchanged.

CisTarget motif enrichment

Nothing changes on the CisTarget side. Same rankings, same annotations, same result. Not evaluated by default because the reference files are around 1.5 GB.

paths <- download_cistarget_hg38()

rankings <- read_motif_ranking(paths$rankings)
annotations <- read_motif_annotation_file(paths$motif_annotations)

scenic_res_et <- tf_to_genes_motif_enrichment(
  x = scenic_res_et,
  motif_rankings = rankings,
  annot_data = annotations,
  gene_id_to_symbol = ensembl_to_symbol,
  cis_target_params = params_cistarget(nes_threshold = 3.0),
  only_high_conf_tf = FALSE,
  .verbose = TRUE
)

cistarget_dt <- get_cistarget_res(scenic_res_et)
head(cistarget_dt[order(-nes)])

Choosing between CPU and GPU

Numbers first. M1 Max (8P+2E cores, 32 GPU cores, unified memory), 10k cells, 1000 TFs, 4000 target genes in roughly 63 batches, 250 trees, max_depth = 10, min_samples_leaf = 50:

Learner	CPU	GPU	Speed-up
ExtraTrees	113.09s	56.66s	2.00x
RandomForest	87.02s	48.10s	1.81x

That is end to end, gene filtering and batching included. Look at a single 64-target batch against one CPU core and the GPU side is 21.3x (ExtraTrees) and 23.6x (RandomForest). The end-to-end number is lower because the CPU version parallelises across all cores, so you’re really comparing 32 GPU cores against 10 CPU cores. RandomForest ends up the fastest absolute option here: subsample_rate = 0.632 means each tree only sees 63% of the cells.

Fidelity against the CPU implementation is per-target Pearson 0.987 for RandomForest and 0.975 with bootstrapping on. Same trees, different scheduling, so the importances agree to within sampling noise.

Where the GPU still loses:

• Small data (a few thousand cells, a few hundred targets). Fixed setup cost eats the win. Stay on CPU.
• GRNBoost2. Its single-target histogram-subtraction algorithm shares no structure across targets and gains nothing from a GPU port, so scenic_grn_sc_gpu() rejects it outright. Use bixverse::scenic_grn_sc().

RandomForest used to belong on that list. Its per-wave tensors needed roughly 8.4 GB at a million cells even at wave size 1, and the driver refused rather than thrashing. The fused-kernel rewrite moved the histogram into threadgroup memory and collapsed that to about 10 MB, so RF now scales like ExtraTrees and runs at cell counts where it previously would not start.

What breaks first at very large cell counts is the packed feature tensor, one byte per (cell, TF) at four bins per u32 word, so 1 GB at 1000 TFs and a million cells. The driver checks that against the adapter’s per-binding limit up front. If you hit it, subsample cells via n_subsample or trim the TF list.

wave_byte_budget is the knob for the middle ground. It caps VRAM for the per-wave histogram and cumulative tensors. Shrink it on an 8 GB adapter you’re sharing with a display; raise it on 16 GB+ to let the scheduler pick a wider wave.