![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
NPUs?
- Thread starter philb2
- Start date
jacuzz1
Supreme [H]ardness
- Joined
- Feb 14, 2004
- Messages
- 7,315
What is an NPU?
- Joined
- May 18, 1997
- Messages
- 55,634
Neural Processing Unit.
jacuzz1
Supreme [H]ardness
- Joined
- Feb 14, 2004
- Messages
- 7,315
OIC, AI cards. We are there already?
dogDAbone
Gawd
- Joined
- Nov 24, 2022
- Messages
- 666
Well, my Omega-particle-powered, quantum-phased positronic NPU has been working just fine for about 10 yrs now, and the only "motherboard" I needed was that grey stuff thats floatin around in me head 
/s
But seriously, current CPU's & GPU's are really just becoming useful for anything significant, so I think "NPU"s are a ways off...
/sBut seriously, current CPU's & GPU's are really just becoming useful for anything significant, so I think "NPU"s are a ways off...
Id love to see more creative socs then the monolithic x86 cpu, as their current fab processes seem to work well interconnecting chipplet/memory.
What makes an optimum NPU configuration architecture wise?
Ive been waiting for AMD chips with a decent sized FPGA and cache/memory pool. That could be used for accelerating a wide variety of things including neural networks (or any ridgedly defined algorithm). The "software" support doesnt exist to easily use something like that at a consumer level.
I wonder if AI acceleration would run into a similar issue with most of the software dominated by the nVidia suite. Additionally, there isn't nearly as much value to many of these workloads on a local machine. Like it or not, these models work best on a very large centralized scale with enormous amounts of data.
What makes an optimum NPU configuration architecture wise?
Ive been waiting for AMD chips with a decent sized FPGA and cache/memory pool. That could be used for accelerating a wide variety of things including neural networks (or any ridgedly defined algorithm). The "software" support doesnt exist to easily use something like that at a consumer level.
I wonder if AI acceleration would run into a similar issue with most of the software dominated by the nVidia suite. Additionally, there isn't nearly as much value to many of these workloads on a local machine. Like it or not, these models work best on a very large centralized scale with enormous amounts of data.