Crown Citadel Group Ciru Inference Lab llm.ciru.ai / research Research index
Technical illustration of a graphics card powering a luminous neural bonsai

Crown Citadel Benchmark Report · July 15, 2026

RTX 4080 SUPER × Bonsai 27B

Successful target-only and DSpark inference benchmarks for the Ternary Bonsai 27B GGUF model on a single 16 GB NVIDIA GPU.

Single GPU 27B parameters CUDA 12.4 262K validated

Benchmark headline

262,144Maximum validated target-only context with KV4.
817.36Prompt tokens/s at the full 262,016-token pass.
57.71Generation tokens/s at the full-window pass.
141.27Peak decoded DSpark generation tokens/s.

GPU, model, and runtime

GPU
NVIDIA GeForce RTX 4080 SUPER
16,376 MiB VRAM · compute capability 8.9
Target model
Ternary Bonsai 27B Q2_0
Ternary-Bonsai-27B-Q2_0.gguf · 6.66 GiB
DSpark drafter
Ternary Bonsai 27B DSpark Q4_1
Ternary-Bonsai-27B-dspark-Q4_1.gguf · 1.81 GiB
Runtime
PrismML llama.cpp · Windows CUDA 12.4
Build b1-62061f9 · commit 62061f9

Target-only throughput

The KV4 lane validates long-context operation. The F16 lane captures the high-throughput context ladder using the standard llama-bench profile.

KV4 full-context scaling

Prompt / contextPP tok/sTG tok/sPeak VRAMMinimum free
16,3841,963.3166.189,158 MiB6,888 MiB
32,7681,793.5464.939,499 MiB6,547 MiB
65,5361,588.4666.6510,030 MiB6,016 MiB
131,0721,229.4966.9311,478 MiB4,568 MiB
262,144817.3657.7114,329 MiB1,717 MiB

KV4 VRAM scaling

F16 context ladder

Prompt tokensPP tok/sTG tok/sPeak VRAMMinimum free
5122,157.87 ± 35.9468.22 ± 0.088,543 MiB7,503 MiB
4,0962,111.83 ± 4.2868.04 ± 0.178,758 MiB7,288 MiB
8,1921,925.5261.148,972 MiB7,074 MiB
16,3841,851.6662.739,474 MiB6,572 MiB
32,7681,673.0559.8710,633 MiB5,413 MiB
65,5361,609.6568.3112,729 MiB3,317 MiB
98,3041,407.0466.2014,783 MiB1,263 MiB
110,5921,339.0368.1015,576 MiB470 MiB
Ternary signal paths and bonsai-shaped circuitry on a dark silicon substrate
One card. 27 billion parameters. Target-only KV4 and GPU-resident DSpark profiles.

DSpark speculative throughput

Decoded generation speed and draft acceptance for the successful GPU-resident DSpark profiles.

Workload throughput

GPU-resident KV4 context

ContextWorkloadDecoded TGDraftedAcceptedAcceptancePeak VRAMMinimum free
1,024Code · temp 0.7117.907 t/s1449465.278%13,165 MiB2,881 MiB
2,048Code · temp 0.7113.789 t/s1449465.278%14,089 MiB1,957 MiB
2,048Code · greedy123.237 t/s1329773.485%14,346 MiB1,700 MiB
2,048Arithmetic · temp 0.7141.272 t/s11210190.179%14,374 MiB1,672 MiB
GPU DSpark KV4 contextPP tok/sTG tok/sAcceptanceFree after request
16,3841,185.8498.1157.89% · 88/1521,806 MiB
20,4801,056.7097.5257.89% · 88/152871 MiB
22,5281,274.5095.8957.89% · 88/152472 MiB

Validated configuration

Target-only · F16 throughput

Standard llama-bench profile

-ngl 99 -fa on
-ctk f16 -ctv f16
-t 8 -b 512 -ub 512
Target-only · maximum context

262K KV4 profile

-ngl 99 -fa on
-ctk q4_0 -ctv q4_0
-c 262144 -np 1
-b 512 -ub 512
Target-only · balanced headroom

131K KV4 profile

-ngl 99 -fa on
-ctk q4_0 -ctv q4_0
-c 131072 -np 1
-b 512 -ub 512
DSpark · decoded speed

GPU-resident drafter

--spec-type draft-dspark
--spec-draft-n-max 4
-ngl 99 -ngld 99
-c 1024

Sampler: temperature 0.7 · top-p 0.95 · top-k 20 · seed 42.

DSpark · practical API context

20,480-token GPU KV4 profile

--spec-type draft-dspark
--spec-draft-n-max 4
-c 20480 -np 1
-b 128 -ub 128
-ngl 99 -ngld 99
-ctk q4_0 -ctv q4_0
-ctkd q4_0 -ctvd q4_0
Model files

Target + paired drafter

Target:
Ternary-Bonsai-27B-Q2_0.gguf

Drafter:
Ternary-Bonsai-27B-dspark-Q4_1.gguf
Model card ↗ DSpark guide ↗