Supported GPU Overview

Neural Network technology used by Voicegain requires hardware capable of executing computation fast enough to support real-time speech recognition for multiple simultaneous ports/sessions.

Here are the cards that should work – however, not all of them fit a specific type of server. Some of them may have cooling, power, or other problems if put in a specific server, so it is good to verify with this page (Qualified System Catalog | NVIDIA) for fitment. Or you can check on a server manufacturer page to see if they offer certain combination. :

NOTE: speed comparisons are based on Nvidia benchmarks and may not reflect Voicegain inference performance.

• https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units#Tesla
  • Below is selection of Tesla cards that Nvidia says are particularly meant for deep learning
    these cards should work fine in rack servers – they do not have built-in fans and rely on server fans for cooling
    • Tesla P100 Data Center Accelerator | NVIDIA (I would use this one only if you can get a deal for a used one as this card is a bit old)
    • NVIDIA V100 | NVIDIA (same note as above) (about 20% of the speed of A100-40GB)
    • NVIDIA T4 Tensor Core GPU for AI Inference | NVIDIA Data Center (single slot, half height card, about 12% of the speed of A100-40GB)
    • A10 Tensor Core GPU | NVIDIA (single slot card, about 30% of the speed of A100-40GB)
    • A30 Tensor Core GPU for AI Inference | NVIDIA (about half the speed of A100-40GB)
    • NVIDIA A100 | NVIDIA (40GB version of A100 is at least 8x faster than T4)
• https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units#RTX_Ax000_series
  • These cards may not work in some rack servers because of the cooling (they will work in Dell R740)
    • RTX A4000 Graphics Card | NVIDIA (single slot card)
    • Introducing RTX A5000 Graphics Card | NVIDIA
    • NVIDIA RTX A6000 For Powerful Visual Computing | NVIDIA
• https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units#Quadro_RTX_x000_series
  • These cards may not work in some rack servers because of the cooling issues. Plus they require more bulky power connectors that may not fit in every server case
    Also these are older generation than RTX_Ax000 series.
    • Quadro RTX 5000 Graphics Card | NVIDIA Quadro
    • Quadro RTX 6000 Graphics Card | NVIDIA Quadro
    • Quadro RTX 8000 Graphics Card | NVIDIA Quadro

Gaming cards are a good deal (significantly cheaper than the cards listed above) but they have the following problems:

• Nvidia license does not allow their use in data centers
• Cooling flow is not designed for tight spaces in servers
• The power connectors are from the top (not back) and may be impossible to connect in some servers.

Gaming Cards work fine in a lab, in a workstation.