Skip to content

tenstorrent/tt-metal

Repository files navigation

ttnn logo

TT-NN is a Python & C++ Neural Network OP library.


Grayskull (GS) Models

Model Batch End-to-end throughput [1] Device throughput [2] Target
ResNet-50 (fps) 20 5,500 7,700 10,000
BERT-Large (sen/s) 12 370 406 410
Falcon7B-decode (t/s) 32 135 135 140
ViT (fps) 8 860 1570 2000
T5 small (sen/s) 140
Bloom (sen/s) 70
U-Net coming soon

[1] - Observed from the host. Includes dispatch overhead and kernel execution time. For LLMs, token-to-token decode throughput is reported.

[2] - Ignoring host overhead. Kernel execution time only. For LLMs, token-to-token decode throughput is reported.

Wormhole (WH) Models

Note

All model demos in this table function on both N150 and N300 Wormhole cards, unless otherwise stated.

Furthermore, all performance numbers here are run or based off an N300 Wormhole card.

Model Gen. Token [3] Batch End-to-end throughput [1] Device throughput [2] Target
Falcon7B 129th 32 13.3 t/s/u - 425 t/s 15.4 t/s/u - 493 t/s 26
Mistral-7B 129th 32 9.9 t/s/u - 317 t/s 11.0 t/s/u - 352 t/s 25
Mamba-2.8B any 32 11.6 t/s/u - 370 t/s 16.5 t/s/u - 528 t/s 41
BERT-Large (sen/s) [4] 8 270 340 400
Stable Diffusion 1.4 512x512 (sec/img) [5] 1 6 5 3
ResNet-50 (fps) 16 4,300 5,550 7,000

[1] - Observed from the host. Includes dispatch overhead and kernel execution time. For LLMs, token-to-token decode throughput is reported.

[2] - Ignoring host overhead. Kernel execution time only. For LLMs, token-to-token decode throughput is reported.

[3] - Generating the i'th token in a sequence while the kv_cache is filled with i-1 rows.

[4] - This model demo does not work on N150. It does work on N300.

[5] - This model demo does not work on N300. It does work on N150.

TT-QuietBox & TT-LoudBox (2x4 mesh of WHs) Models

Model Technique Gen. Token [3] Batch End-to-end throughput [1] Device throughput [2] Target
Falcon7B Data Parallel 129th 256 7.4 t/s/u - 1901 t/s 15.5 t/s/u - 3968 t/s 26 t/s/u
LLaMA-2-70B Tensor Parallel 129th 32 10.4 t/s/u - 333 t/s 16.6 t/s/u - 532 t/s 20 t/s/u
LLaMA-3-70B Tensor Parallel 129th 32 10.4 t/s/u - 333 t/s 15.8 t/s/u - 506 t/s 20 t/s/u
Falcon40B Tensor Parallel 129th 32 work-in-progress 10.0 t/s/u - 320 t/s 36 t/s/u
Mixtral7Bx8 Tensor Parallel 129th 32 15.1 t/s/u - 483 t/s 27.1 t/s/u - 868 t/s 33 t/s/u
ResNet50 Data Parallel coming soon

Using TT-NN ops and tensors

import ttnn
import torch

with ttnn.manage_device(device_id=0) as device:
   a = torch.ones((5, 7))
   b = torch.ones((1, 7))

   a = ttnn.from_torch(a, device=device, dtype=ttnn.bfloat16, layout=ttnn.TILE_LAYOUT)
   b = ttnn.from_torch(b, device=device, dtype=ttnn.bfloat16, layout=ttnn.TILE_LAYOUT)

   output = a + b
   output = ttnn.to_torch(output)

print(output)

TT-Metalium logo

TT-Metalium is our low-level programming model, enabling kernel development for Tenstorrent hardware.

Getting started

Get started with simple kernels.