help me undervolt E8400 (need motherboard recommendations)

[onlinespending]

I have an E8400 that I would like to undervolt and underclock dramatically, since I am using the CPU in server that will be on quite often. I'd like to push it to its limits, something like 0.8V @ 1GHz. However I am struggling getting anything below 1.1V using the BIOS and CrystalCPUID.

The motherboard is an ASUS P5K-VM, which allows me to manually set the CPU VCORE to 1.1V. CrystalCPUID on the other hand only allows 1.15V, even worse. I even tried the ASUS AI Suite hoping that it would allow me to control the CPU Voltage through software, but I could not find anything.

Am I out of luck? Is this a matter of a motherboard limitation? Are there workarounds?

I also have a MSI G33M-FI motherboard and can swap that one that I am using in an HTPC with the Asus one if need be. Unfortunately at the time I bought these microATX boards they were the only ones that I could definitely say had on-board Gigabit Ethernet controllers (Marvell 88E8056) that supported Jumbo Frames and was hanging off of the PCIe bus and not standard PCI. My Gigatbit Ethernet switch supports Jumbo Frames and I wanted the maximum network performance possible.

[ces]

To my recollection there is perhaps 10 watts difference between the G33 and the G31 chip sets. If the goal is reducing energy draw, maybe you should focus your efforts on moving to a g31 board.

[onlinespending]

True, if that's the case perhaps I need to explore that. But as I mentioned it was nearly impossible to find a microATX board at the time that had an embedded Gigabit Ethernet controller that supported Jumbo Frames and was on the PCIe bus and not just PCI. Getting optimum Gigabit Ethernet was top priority to me.

But it's been 2 years since then, so many more microATX boards have come out and I have many more options. So I guess a good question would be, what are considered the ultra-low power LGA775 microATX boards?

[Jay_S]

Getting optimum Gigabit Ethernet was top priority to me.

There are many variables contributing to network performance. Jumbo support is the last thing I look for anymore. Back when CPU's couldn't handle the I/O to saturate gigabit, jumbo's made more sense. But even then, they only made sense for large, sequential transfers. With anything latency-sensitive (gaming, VOIP, video conf), jumbo framing can actually hurt performance.

Network performance is ultimately limited by the physical transfer speeds of your HDD's and the file copy engine & protocol in your OSes.

Jumbos only work with gigabit. Some switches get funky when gigabit+jumbo gear is connected with non-gigabit+non-jumbo gear (printers, network audio streamers, VOIP phones, wireless AP's, etc). You can mix jumbo gear with non-jumbo gear if you segregate jumbo & non-jumbo traffic w/subnets or VLANs. To confuse the situation more, not every manufacturer conceives of jumbo framing in the same way, so mixing vendors may toss you.

So I wonder: is it worth it to design around jumbo support?

Here's a couple of forum posts from Small Net Builder, about a guy's DIY linux NAS. He's hitting 108MB/s network copies with linux software raid 5 array, and it idles around 45W w/drives spun down.

http://forums.smallnetbuilder.com/showthread.php?t=3144
http://forums.smallnetbuilder.com/showthread.php?t=2429

No jumbos needed. Just a fast disk subsystem.

[onlinespending]

I have no doubt that you can still achieve high bandwidth without Jumbo Frames. From my personal experience I have seen a difference. A part of it may also be because I've already invested in making sure all of my devices and my switch are Jumbo Frame compatible.

But the bigger thing may actually be the controller itself and the bus limitations. 2 years ago, many of the microATX offerings had the GBe hanging off of the PCI bus. Not good. Benchmarking will show you there is a noticeable difference between a PCIe solution and a PCI solution. On a related note, a couple years back Intel included a "CSA" bus (Communications Streaming Architecture) for onboard gigabit ethernet on some of its motherboards. Because it was a dedicated bus, benchmarks showed that it had superior throughput to PCI and even PCIe gigabit ethernet solutions. Sadly it looks like this has been abandoned.

Along those lines, I really don't want a Realtek controller. From the benchmarking I remember reading, they are inferior to Marvell's solutions. Whether that's still true today, I am a bit influenced by the past results.

But at any rate, assuming I can find a motherboard with a PCIe Gigabit Ethernet controller (ideally Marvell and ideally with Jumbo Frame support), I do want that board to be as low power as possible and one that allows extreme undervolting and underclocking of the CPU. Anyone have any suggestions for such an LGA775 board?

[Jay_S]

Well, there's always Intel - in my experience, you can't go wrong with Intel NICs. But you (mainly) only find them on Intel boards.

The Intel DB43LD ($77) is the least expensive board Newegg sells with an Intel NIC. More info at Intel's product page. It uses Intel's 82567LM NIC, which supports up to 9k jumbos.

However, Intel boards have a reputation for little/no voltage adjustment. And based on the manual, the DB43LD has none.

What about finding a great under-clocking/volting board and adding a dedicated PCIe NIC?

[onlinespending]

I looked at the datasheet for the E8400 and it would appear I could do a relatively straightforward mod to the VID bus to undervolt the CPU. Right now the motherboard only allows values from 1.1V to 1.6V I believe.

But if I keep that range from 1.225V and 1.6V and tie VID[6] high I will effectively change the range from 0.825V to 1.2V. This would be perfect because I could run an extreme underclock at the low end of the range (0.825V) and the stock frequency at the high end of the range (1.2V, which is the default Vcore).

Anyone have any experience doing this?

[Jay_S]

No personal experience, but I believe you use conductive ink and tape off around the surrounding pins. Not something I would attempt with a $170 CPU.

[ces]

I looked at the datasheet for the E8400 and it would appear I could do a relatively straightforward mod to the VID bus to undervolt the CPU. Right now the motherboard only allows values from 1.1V to 1.6V I believe.

Doing that would scare the heck out of me.

[onlinespending]

No personal experience, but I believe you use conductive ink and tape off around the surrounding pins. Not something I would attempt with a $170 CPU.

Yeah, I see that is what most people do. Unfortunately there's no VCC pin adjacent to the VID[6] pin, so I'm sort of SOL there. But in some ways I'd almost rather find the trace on the motherboard and do the hack there. I'll do some more investigation.

[onlinespending]

No personal experience, but I believe you use conductive ink and tape off around the surrounding pins. Not something I would attempt with a $170 CPU.

Yeah, I see that is what most people do. Unfortunately there's no VCC pin adjacent to the VID[6] pin, so I'm sort of SOL there. But in some ways I'd almost rather find the trace on the motherboard and do the hack there. I'll do some more investigation.

Well I finally tried modding the CPU to run at a lower VCore. Because of the fact that VCC and VID[6] are not adjacent, I couldn't get away with using conductive ink. So I carefully soldered the two pins using a really thin wire (37 AWG, which is the same size as the human hair). It appeared to work in a sense, since the AC power reading was about 20W lower. However, as you can imagine it didn't post. I suppose the thing is trying to boot at its stock frequency of 3GHz but at a Vcore of only 0.8V. Obviously that's not gonna fly. Outside of modding the BIOS file itself to somehow change the 8-bit VID code the BIOS uses, I believe I'm out of luck.

So I'm back to my original question. What sort of board can I get (preferably G31) that does allow undervolting? Thanks!

EDIT: Wait a second. I take that back. If the thing is simply booting at 1.2V that's a VID of 01000010. So VID[6] being tied high shouldn't have any impact. Only if it's setting it to 1.225V or above would it essentially flip to the lower voltages (1.225V would become 0.825V), so I'm actually confused now why this didn't work. I suppose it's possible the BIOS uses a higher VCore than default for the boot sequence?

[onlinespending]

OK I was able to get it to work. I realized after looking at the E8400 datasheet a bit more, that creating an open on the VID[6] pin would pull it high at the VRM. So I first removed any mod and went into the BIOS to set the voltage manually to 1.2V, the FSB to just 200MHz, and the multiplier to 6x (so a 1.2 GHz CPU frequency).

Interestingly enough, manually setting the voltage even to 1.35V, which should have flipped it to 0.95V (because VID[6] was high), actually set it to 1.35V. CPU-Z reported the same thing. However, if I set the VCore to AUTO in the BIOS, it fell to about 0.8V (perhaps 1.225V flipping to 0.825V). My only explanation for this is that setting it to AUTO allows the CPU to control the VID bus, where as setting it manually somehow controls it directly to the VRM. The one annoying thing about this however is that everytime I start the system the BIOS complains of an Overvoltage Warning and I need to hit F1 to proceed (prevents the system from restarting without human intervention).

At any rate I had my system running at 1.2GHz and a Vcore of 0.8V. Idle AC power consumption dropped 5 - 10W, while full-tilt the power dropped 15W (this was by just raising the Vcore to 1.2V but keeping the but keeping the clock running at 1.2GHz).

While that's better than nothing, my total idle power consumption is more than I would like. I am running three 2TB Caviar Green drives in a RAID5 configuration on a Highpoint RocketRAID 2310. So that certainly limits just how low I can go, but I'm really opting towards getting an Atom mini-ITX system. The D510 looks good, but of all the boards I've seen, either it has a PCIe slot (something I need for the RAID adapter) but no GBe, or it has GBe but only a PCI slot. I can't win!