Article: Power Delivery in a Modern Processor
By: anon³ (ilove.delete@this.thelt3042.fake), May 14, 2020 5:22 pm
Room: Moderated Discussions
Michael S (already5chosen.delete@this.yahoo.com) on May 14, 2020 12:29 pm wrote:
> linear regulators with very low drop certainly exist, but they have one big disadvantage
> - they poorly reject medium frequency noise (200 KHz to few MHz) when it comes from input voltage
> side. Only those with drop of 300 mV or more are filtering incoming ripples really well.
>
> But, then again, he probably don't care all that much about regulators with small output current.
>
It's worse than this. In general, dedicated linear regulator ICs of all types reject very little above about 1MHz. One may be able to do a bit better with a discrete fast error amp + pass transistor architecture, but it's rarely worth bothering. Rejection of high frequency noise is achieved with more capacitors, possibly in combination with inductors or ferrites. This is a problem if you want to slew the voltage quickly. Control bandwidth on the timescales needed for a modern CPU VRM is not a trivial task.
The overall highest performance linear regulator I know of is the LT3042 (and friends, it's a family): https://www.analog.com/media/en/technical-documentation/data-sheets/3042fb.pdf
And there's a fascinating app note about how one actually _measures_ performance that good, LTC AN159: https://www.analog.com/media/en/technical-documentation/application-notes/an159fa.pdf
> linear regulators with very low drop certainly exist, but they have one big disadvantage
> - they poorly reject medium frequency noise (200 KHz to few MHz) when it comes from input voltage
> side. Only those with drop of 300 mV or more are filtering incoming ripples really well.
>
> But, then again, he probably don't care all that much about regulators with small output current.
>
It's worse than this. In general, dedicated linear regulator ICs of all types reject very little above about 1MHz. One may be able to do a bit better with a discrete fast error amp + pass transistor architecture, but it's rarely worth bothering. Rejection of high frequency noise is achieved with more capacitors, possibly in combination with inductors or ferrites. This is a problem if you want to slew the voltage quickly. Control bandwidth on the timescales needed for a modern CPU VRM is not a trivial task.
The overall highest performance linear regulator I know of is the LT3042 (and friends, it's a family): https://www.analog.com/media/en/technical-documentation/data-sheets/3042fb.pdf
And there's a fascinating app note about how one actually _measures_ performance that good, LTC AN159: https://www.analog.com/media/en/technical-documentation/application-notes/an159fa.pdf