Intel's Haswell Moves Voltage Regulator On-Die

MojoKid writes "For the past decade, AMD and Intel have been racing each other to incorporate more components into the CPU die. Memory controllers, integrated GPUs, northbridges, and southbridges have all moved closer to a single package, known as SoCs (system-on-a-chip). Now, with Haswell, Intel is set to integrate another important piece of circuitry. When it launches next month, Haswell will be the first x86 CPU to include an on-die voltage regulator module, or VRM. Haswell incorporates a refined VRM on-die that allows for multiple voltage rails and controls voltage for the CPU, on-die GPU, system I/O, integrated memory controller, as well as several other functions. Intel refers to this as a FIVR (Fully Integrated Voltage Regulator), and it apparently eliminates voltage ripple and is significantly more efficient than your traditional motherboard VRM. Added bonus? It's 1/50th the size." Update: 05/14 01:22 GMT by U L : Reader AdamHaun comments: "They already have a test chip that they used to power a ~90W Xeon E7330 for four hours while it ran Linpack. ... Voltage ripple is less than 2mV. Peak efficiency per cell looks like ~76% at 8A. They claim hitting 82% would be easy..." and links to a presentation on the integrated VRM (PDF).

Comments

From a former power supply designer - Neat!

[jimmyswimmy]

That's some amazing work. The current state of the art in CPU power supply designs hasn't changed in 15 years. 12V in, low voltage out, and the output voltage has been moving lower and lower for years, with designs below 1 V. If you figure you had a few percent of tolerance in the early years when everything ran off 2.5V and that few percent remains constant, then at 1 V you have almost no room for slop. So there are a lot of output capacitors there, both those electrolytics (you always hear people complaining about them but they're CHEAP) and ceramics. The ceramics cost a fortune and you need a lot of them to get your tolerance down - the first half microsecond of a load step is entirely the ceramic capacitor's response, not the controller or anything else. Moving part of the VR onboard allows them to reduce the parasitics significantly and they can probably tolerate a little higher tolerance as a result, but moreover they can get rid of some of those ceramics in the whole system - ultimately many of those on the motherboard.

So this is taking a lot of cake out of company mouths. Analog, Intersil, IRF, ON, who else - manufacturers of controllers, MOSFETs. Inductors, ceramic and 'lytic vendors are all going to lose out a bit here. Potentially Intel can reduce the platform cost vs. AMD as well, which is interesting. There is still an onboard VR but it will be 12 - 2.4 V, wherever they think the sweet spot is for efficiency and size. And the first real change in this industry for a long time. Cool work.

Re:Heat

[viperidaenz]

If you core requires 1V and 90 watts you need to transfer 90A through your PCB traces, up in to the chip, across the bond wires (if there are any) and on to the die.
If your die has a regulator on board and accepts 12V instead, and is 80% efficient you only need to transfer 9.4A. You've just lowered your resistive losses by about 100x. If the connection between the external VRM and die is 0.001ohms, at 90A you waste 8.1W. at 9.4A you waste 0.088W.