TRELLIS.2 benchmark
NVIDIA RTX 6000 Ada Generation: 9 Works generated through the Kollektive stage contract, each from one generation. Medians of the warm Works (model already loaded); the first Work also loads the model. At the quoted $0.649/h a warm Work costs about $0.0159 of GPU time (nominal, not an invoice).
Where the time goes (warm median)
| Model loading | 0.0 s |
|---|---|
| Generation | 32.4 s |
| Geometry extraction | 37.5 s |
| Material processing | 10.0 s |
| GLB export | 7.7 s |
| Preview | 0.2 s |
| Total per Work | 88.1 s |
Peak GPU memory over the bench: 13235.1 MB reserved, 7868.4 MB allocated.
Every Work
| Source | Model loading | Generation | Geometry extraction | Material processing | GLB export | Preview | Total | Background | VRAM reserved | VRAM allocated | 2048 GLB | 1024 GLB | Raw GLB | Nominal $ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| abo-b07hk8yhb7-cutout (cold) | 94.3 s | 87.0 s | 112.2 s | 11.2 s | 4.1 s | 0.3 s | 309.7 s | source_alpha | 7524.6 MB | 4533.2 MB | 17.4 MB | 17.0 MB | 148.3 MB | $0.0558 |
| abo-b075yptdt5 | 1.7 s | 26.4 s | 45.9 s | 9.0 s | 5.7 s | 0.2 s | 89.3 s | rmbg | 7180.6 MB | 4097.3 MB | 17.3 MB | 16.8 MB | 138.5 MB | $0.0161 |
| abo-b07hk8yhb7 | 0.0 s | 29.9 s | 30.1 s | 8.9 s | 7.1 s | 0.4 s | 77.0 s | rmbg | 6243.2 MB | 4700.7 MB | 16.6 MB | 16.1 MB | 141.6 MB | $0.0139 |
| abo-b083sdzrmb | 0.0 s | 70.3 s | 61.9 s | 10.3 s | 13.0 s | 0.2 s | 156.1 s | rmbg | 13235.1 MB | 7868.4 MB | 19.2 MB | 18.2 MB | 258.4 MB | $0.0281 |
| met-245539 | 0.0 s | 44.0 s | 44.8 s | 9.9 s | 8.4 s | 0.3 s | 107.8 s | rmbg | 8845.8 MB | 5284.4 MB | 16.7 MB | 16.4 MB | 174.8 MB | $0.0194 |
| commons-2015-02-17-shichimi-concept-art-by-david-revoy-jpg | 0.0 s | 34.8 s | 26.6 s | 9.8 s | 8.5 s | 0.2 s | 80.3 s | rmbg | 9667.9 MB | 5873.9 MB | 20.0 MB | 19.0 MB | 195.9 MB | $0.0145 |
| commons-img-9443-guggenheim-museum-bilbao-maman-jpg | 0.0 s | 27.8 s | 16.5 s | 10.9 s | 3.2 s | 0.1 s | 58.8 s | rmbg | 4506.8 MB | 3619.4 MB | 17.3 MB | 16.6 MB | 66.0 MB | $0.0106 |
| nasa-pia00662 | 0.0 s | 20.6 s | 63.3 s | 10.5 s | 5.0 s | 0.1 s | 99.8 s | rmbg | 4806.7 MB | 3619.3 MB | 19.4 MB | 18.5 MB | 73.5 MB | $0.0180 |
| synthetic-teapot | 0.0 s | 42.6 s | 26.1 s | 10.0 s | 9.2 s | 0.2 s | 88.5 s | rmbg | 10003.4 MB | 6820.0 MB | 16.9 MB | 16.3 MB | 223.2 MB | $0.0160 |
What the files are made of
Geometry and the 2048 and 1024 textures, read from the delivered GLBs; playback shares the original's geometry. “Transparent share” is the part of the base-colour texture below 95% opacity (including inpainted margins between UV islands); the exporter switches a material to BLEND above 0.1% of covered texels.
| Source | Faces | Geometry | 2048 textures | 1024 textures | Geometry share of 2048 GLB | Transparent share | Alpha mode |
|---|---|---|---|---|---|---|---|
| abo-b07hk8yhb7-cutout | 496,746 | 16.8 MB | 0.6 MB | 0.2 MB | 96% | 0.00% | OPAQUE |
| abo-b075yptdt5 | 484,953 | 16.5 MB | 0.8 MB | 0.3 MB | 95% | 0.00% | OPAQUE |
| abo-b07hk8yhb7 | 489,494 | 15.8 MB | 0.8 MB | 0.3 MB | 95% | 0.00% | OPAQUE |
| abo-b083sdzrmb | 488,719 | 17.8 MB | 1.4 MB | 0.4 MB | 93% | 0.01% | OPAQUE |
| met-245539 | 494,718 | 16.3 MB | 0.5 MB | 0.1 MB | 97% | 0.00% | OPAQUE |
| commons-2015-02-17-shichimi-concept-art-by-david-revoy-jpg | 496,714 | 18.5 MB | 1.4 MB | 0.5 MB | 93% | 0.00% | OPAQUE |
| commons-img-9443-guggenheim-museum-bilbao-maman-jpg | 483,164 | 16.3 MB | 0.9 MB | 0.3 MB | 95% | 0.00% | OPAQUE |
| nasa-pia00662 | 469,571 | 18.0 MB | 1.3 MB | 0.4 MB | 93% | 0.02% | OPAQUE |
| synthetic-teapot | 487,780 | 16.0 MB | 0.9 MB | 0.3 MB | 95% | 0.00% | OPAQUE |
Warm experiments
Models resident on the GPU
The same Source and seed as its matrix Work, with upstream's low-VRAM transfers turned off. Equivalence compares the delivered 2048 GLB with the matrix Work's.
| Source | Matrix (low VRAM) | Resident | Saved | VRAM reserved, matrix | VRAM reserved, resident | Equivalence |
|---|---|---|---|---|---|---|
| abo-b075yptdt5 | 87.6 s | 116.1 s | -32% | 7180.6 MB | 23699.9 MB | {"faces": [484953, 483889], "identicalBytes": false, "sameTopology": false, "textureMse": 1580.794423699379, "texturePsnrDb": 16.142049656087075} |
| met-245539 | 107.8 s | 93.6 s | 13% | 8845.8 MB | 25409.1 MB | {"faces": [494718, 495244], "identicalBytes": false, "sameTopology": false, "textureMse": 168.08661895990372, "texturePsnrDb": 25.875472193688513} |
| synthetic-teapot | 88.5 s | 77.2 s | 13% | 10003.4 MB | 27133.0 MB | {"faces": [487780, 488250], "identicalBytes": false, "sameTopology": false, "textureMse": 1992.7266152501106, "texturePsnrDb": 15.136326395042353} |
Where a warm Work spends GPU time
Profiled warm Work: 103.7 s wall (profiling adds overhead), 108.6 s of GPU operator time; host-to-device copies 5.8 s, device-to-host 5.6 s.
| Operator | Calls | GPU | CPU |
|---|---|---|---|
| aten::copy_ | 75148 | 14.16 s | 0.30 s |
| void flash_fwd_kernel<Flash_fwd_kernel_traits<128, 128, 32, 4, false, false, cutlass::bflo | 3900 | 14.05 s | 0.00 s |
| flash_attn::_flash_attn_varlen_forward | 3240 | 13.91 s | 0.20 s |
| aten::addmm | 16977 | 11.98 s | 0.38 s |
| Memcpy HtoD (Pageable -> Device) | 11897 | 5.84 s | 0.00 s |
| Memcpy DtoH (Device -> Pageable) | 11560 | 5.62 s | 0.00 s |
| ampere_bf16_s1688gemm_bf16_64x128_sliced1x2_ldg8_relu_f2f_tn | 6840 | 4.74 s | 0.00 s |
| ampere_bf16_s16816gemm_bf16_128x64_ldg8_relu_f2f_tn | 1620 | 3.15 s | 0.00 s |
| ampere_bf16_s1688gemm_bf16_128x128_ldg8_relu_f2f_stages_32x1_tn | 4590 | 3.11 s | 0.00 s |
| aten::mul | 32988 | 2.80 s | 0.24 s |
| Buffer Flush | 71 | 1.98 s | 0.02 s |
| GridSample3dFunction | 3 | 1.97 s | 0.25 s |
| indice_weighed_sum_fwd_kernel | 4352 | 1.96 s | 0.00 s |
| void at::native::vectorized_elementwise_kernel<4, at::native::bfloat16_copy_kernel_cuda(at | 23028 | 1.49 s | 0.00 s |
| void at::native::unrolled_elementwise_kernel<at::native::direct_copy_kernel_cuda(at::Tenso | 13905 | 1.13 s | 0.00 s |
| aten::native_layer_norm | 7265 | 0.97 s | 0.07 s |
| void at::native::(anonymous namespace)::vectorized_layer_norm_kernel<float, float>(int, fl | 7265 | 0.97 s | 0.00 s |
| aten::add | 23803 | 0.93 s | 0.16 s |
| aten::index | 14754 | 0.90 s | 0.15 s |
| void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel<at::native: | 9744 | 0.88 s | 0.00 s |
| void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native: | 9507 | 0.86 s | 0.00 s |
| aten::fill_ | 3923 | 0.84 s | 0.02 s |
| void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array | 3502 | 0.84 s | 0.00 s |
| void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<c10::BFloat1 | 10080 | 0.81 s | 0.00 s |
| aten::gelu | 2376 | 0.81 s | 0.02 s |
GPU utilisation per Work
Share of each Work's wall time in which any kernel ran (nvidia-smi, 500 ms samples). It shows idle time, not how full the GPU was while busy.
| Work | Kind | GPU busy |
|---|---|---|
| trellis2-abo-b07hk8yhb7-cutout | matrix | 26.0% |
| trellis2-abo-b075yptdt5 | matrix | 33.3% |
| trellis2-abo-b07hk8yhb7 | matrix | 43.6% |
| trellis2-abo-b083sdzrmb | matrix | 60.0% |
| trellis2-met-245539 | matrix | 43.7% |
| trellis2-commons-2015-02-17-shichimi-concept-art-by-david-revoy-jpg | matrix | 47.1% |
| trellis2-commons-img-9443-guggenheim-museum-bilbao-maman-jpg | matrix | 40.9% |
| trellis2-nasa-pia00662 | matrix | 25.8% |
| trellis2-synthetic-teapot | matrix | 52.0% |
| trellis2-abo-b083sdzrmb-warmup | warmup | — |
| trellis2-abo-b083sdzrmb-warmup | warmup | 37.6% |
| trellis2-abo-b083sdzrmb-profiled | profile | 48.1% |
| trellis2-abo-b075yptdt5-resident | residency | 60.6% |
| trellis2-met-245539-resident | residency | 55.4% |
| trellis2-synthetic-teapot-resident | residency | 56.1% |
| trellis2-met-245539-concurrent | concurrency | 57.8% |
Amazon Brand – Rivet Mid-Century Modern Cone Nightstand Table Desk Lamp with LED Light Bulb - 15 Inches, Black and Gold (background removed by flood fill)


Amazon Brand – Stone & Beam Parson Armless Dining Room Chairs, 40 Inch Height, Set of 2, Beige


Amazon Brand – Rivet Mid-Century Modern Cone Nightstand Table Desk Lamp with LED Light Bulb - 15 Inches, Black and Gold


Amazon Essentials Onore Fashion-Sneakers, Evelina, 8 Toddler


Part of the left foot of a colossal marble statue


2015-02-17 shichimi-concept-art by-David-Revoy.jpg


"IMG 9443"Guggenheim Museum Bilbao - "Maman".jpg


360-degree b/w Monster


Controlled synthetic ceramic teapot

