The NPU has become a component very important thanks to the advance of AIand AMD has been one of the companies that has best played its cards in this regard with its XDNA 2 line, which offers a good level of performance maintaining high efficiency and low impact at the silicon level, which facilitates its integration into increasingly complex socs, such as the Ryzen AI 300 series.
When we think of an NPU, it is most likely to come to mind that idea, that of an integrated component next to a CPU and a GPU, A reality that AMD is thinking of changing. Rahul Tikoo, senior vice president and general manager of the PCs unit for AMD consumption, commented that they are considering the possibility of launching dedicated NPUS.
It would be a very intelligent, and totally logical movement. Oems as Dell have already confirmed a month ago they could use A Qualcomm AI 100 inference card for PCwhich contains a high performance NPU, in its laptop Dell Pro Max Plus. However, they are still not clear if this project will be ahead, so AMD cannot talk about dates.
What power would a dedicated NPU of AMD have?
It is the key question, and the answer is that it would have to be much more powerful than an integrated NPU, that is clear. Right now, the most powerful AMD in its integrated NPUS catalog Reach 50 tops, Figure that exceeds 40 tops required by Microsoft Copilot+ to function at home without any performance problem.
Combining two units could climb the yield at 100 tops, and continue climbing units to increase the power. However, the objective of this dedicated NPU It would not be to offer a huge amount of power, but to get a good balance between performance, consumption and working temperature, and overcome any graphics card dedicated in this regard.
How could you connect and use on a PC?
The easiest way would be to launch them as expansion cards compatible with the PCIe interface, in such a way that its installation would be as easy as opening the PC and inserting it into a PCIe slot that we have free.
They could also Integrate into M.2 units with a similar or identical format to that of SSDs. This has been precisely the choice of a Startup of Santa Clara, which launched a dedicated NPU last May with a power of 200 tops in M.2 format that has a consumption of only 8.25 watts
For comparative purposes I remind you that PS5 Pro has hardware dedicated to accelerating the power of 300 tops. This same startup designed a PCIE expansion card with four NPUS that can reach 1,000 power tops. This are already major words, since we have great performance with a contained consumption.
Do NPus integrated in expansion cards make sense?
The truth is that yes, but with nuances. These types of products are an interesting idea Because it allows updating equipment that still offer good performance, but that do not have AI acceleration capabilities, and that do not need or do not want to set up a graphics card with this type of acceleration.
It will also improve the performance of equipment that has dedicated hardware for AI, but that no longer offers the power they need. For example, Imagine that a point comes where Microsoft Copilot+ requirements increaseand that a 40 tops NPU is no longer enough to enjoy all its jobs.
If we have an integrated NPU next to the CPU We will not be able to expand your capabilities unless we Buy a new team. Thanks to these dedicated NPUS expansion cards, we could improve the performance of that team and continue using it, instead of being forced to buy another.