( â ) , . , , , state-of-the-art bleeding edge , , , " ". , GPU, , , ( 2 â 2.5 GHz AMD 4+ GHz 64 ). , , !
3 :
- ;
- , ;
- SLA ;
- " " ;
:
- ?
- ?
- ?
. STT , Facebook AI Research , , . , " ".
, 3 , "":
- Throughput ( ) â . "";
- Real Time Factor (RTF) ( ) â , . Real Time Speed (RTS) 1 / RTF, , 1 ;
- Latency () â - ;
FAIR, :
- FAIR Intel Skylake c 18 ( 2 , - );
- end-to-end C++ "" ;
- wav2letter++;
- , - ;
:
- "" (RTS * ) . , ;
- FAIR , .. , TDS-;
- - ;
- , , - state-of-the-art ( , );
- FAIR , "" 40 , 0.26 RTF ( ). , , ;
- 40 * 1 / 0.26 18 . , 1 1 - 8-9 ;
:
- : 750 ( );
- : 8-9 , 40 18 ;
- : 500 1000 , 750 â ;
- C++ -;
â ? , .. . . , - . â , , .
end-to-end , FAIR, , . , , end-2-end ( ).
FAIR, .. , C++ . , .
. , , . , " " "" ( ). ( ):
FP32 | FP32 + Fused | FP32 + INT8 | FP32 Fused + INT8 | Full INT8 + Fused | New Best | |
---|---|---|---|---|---|---|
1 | 7.7 | 8.8 | 8.8 | 9.1 | 11.0 | 22.6 |
5 | 11.8 | 13.6 | 13.6 | 15.6 | 17.5 | 29.8 |
10 | 12.8 | 14.6 | 14.6 | 16.7 | 18.0 | 29.3 |
25 | 12.9 | 14.9 | 14.9 | 17.9 | 18.7 | 29.8 |
, , . , CPU . . FAIR , , :
- -;
- , , , ;
- FAIR , ;
, - .
- .
:
- 1 â 7 , "" 4 â 8 â 16 ;
- ;
- -, . 8 latency 500 , , -, , 8 ;
- , . . -;
GPU
SSD, 256+ GB | NVME, 256+ GB | |
RAM | 32 GB | 32 GB |
8+ | 12+ | |
3 GHz+ | 3.5 GHz+ | |
2 | ||
AVX2 | ||
GPU | 1 | 1 |
8 "" | 16 "" | |
, | 280 | 320 |
95- , | 430 | 476 |
99- , | 520 | 592 |
1000 | 25.0 | 43.4 |
500 | 12.5 | 21.7 |
(1 / RTF) | 85.6 | 145.0 |
10.7 | 12.1 |
3 GPU:
- GPU Nvidia 1070 8+GB RAM ;
- GPU Nvidia Quadro 8+GB RAM (TDP 100 â 150W) ;
- Nvidia Tesla T4, , TDP 75W;
, GPU ( " "). , :
- " (300 ) ". . TDP . TDP 75 150 , - 50-75% ;
- " ". , 2 (+ );
- " ". "" Tesla SLA Nvidia. SLA , .. Tesla 2-3 . "" â ;
- " 2+ ". . Quadro T4 1 ;
- " ". Tesla A100 US$12,500 . Quadro T4 ( ) ;
- " ". "" â 3-4 , . â . , Nvidia 24/7 "" , . 100 Nvidia AMD 3 ;
â . â GPU 2-3 .
CPU
SSD, 256+ GB | SSD, 256+ GB | |
RAM | 32 GB | 32 GB |
8+ | 12+ | |
3.5 GHz+ | 3.5 GHz+ | |
2 | ||
AVX2 | ||
4 "" | 8 "" | |
, | 320 | 470 |
95- , | 580 | 760 |
99- , | 720 | 890 |
1000 | 11.1 | 15.9 |
500 | 5.6 | 8.0 |
(1 / RTF) | 37.0 | 53.0 |
4.6 | 4.4 |
C
- GPU 10 â 15 RTS ( RTS GPU - 500 â 1,000). CPU 1 GPU ( ), , . - ;
- CPU- 5 RTS, , latency throughput;
- . â , - ;
- 50 , 2 GPU (+ ) ;
- GPU - 2-3 , CPU;
FAIR , 50% . â . 20-30 RTS , - 40-50% . :
- GPU , ;
- FAIR;
- , GPU , ;
- - , ;
? MIT.