Slashdot Mirror
Will "Group Hug" Commoditize the Hardware Market?
Will the Open Compute Project’s Common Slot specification and Facebook’s Group Hug board commoditize the data center hardware market even further? Analyst opinions vary widely, indicating that time and additional development work may be necessary before any sort of consensus is reached. At the Open Compute Summit last week, Frank Frankovsky, director of hardware design and supply chain operations at Facebook, announced both the Open Slot specification and Facebook’s prototype Open Slot board, known as “Group Hug.”
Group Hug’s premise is simple: disaggregate the CPU in a way that allows virtually any processor to be linked to the motherboard. This has never been done before with a CPU, which has traditionally required its own socket, its own chipset, and thus its own motherboard. Group Hug is designed to accommodate CPUs from AMD, Intel, and even ARM vendors such as Applied Micro and Calxeda.
if only it came true. Even if politics weren't involed, it still wouldn't be easy at all I imagine. It'd require processors to either rely on a standard memory controller, or to implement their own, along with all sorts of other similar challenges with performance vs. compatibility.
All I see are links to other slashdot articles. Are we going for a new record here? First the ridiculous post about Microsoft welling their entertainment division, now this. And the same style of headline too, which of course is answered with, "No."
Mr Editor, can you at least post a link to some information, like maybe the site where this specification is detailed? Maybe the project web site itself?
One architecture that supported "variable CPUs" was S100 where it is was typical to have a CPU card, one or more memory cards, and multiple I/O cards all plugged into a backplane. There were CPU cards for the Apple ][, but these were complete computers on a card that simply allowed use of the Apple ][ I/O.
Given today's multi-gigahertz processors with gigahertz memory access, I would think it would be difficult, if not impossible to effectively separate the CPU and the memory by very much. Similarly, it gets pretty complicated with high speed DMA I/O when you move it away from the memory it is accessing. I'm sure it could be done, but the performance is going to suffer just from the physical distances. Add in connector resistance and noise and you have ample justification for putting the CPU, chipset and RAM in a very small module that then plugs into the rest of the computer for I/O.
Also see wire-wrapped and bit-sliced.
Great minds think alike; fools seldom differ.
Perhaps I just don't get the use case for anyone other than someone like Google or Facebook. In the grand scheme of things for companies running enterprise applications, compute just isn't that large a portion of overall costs and certainly doesn't need to be dynamically swappable. The operational and change management costs of tracking and provisioning CPUs dynamically would greatly outweigh the benefits.
but i remember a motherboard vendor semi did this
yes, it was your typical motherboard
but it had a pci-e card with cpu + ram + chipset you could slot-in
pretty sure the board was intel while the slot-in was amd
no reason it couldn't work the other way
The CPU is a small part of the cost of the server
What is the point in doing this? Where is the return on investment?
I do no think it means what you think it means. Something that is a commodity product is fungible, meaning any indiviual product from any of various vendors is effectively interchangable with any product of the same kind from any other vendor. Computer hardware has been commoditized for a long time. While processors are not wholely interchangable (AMD vs Intel), the motherboard/CPU combo generally is. Everything else in a computer can more or less be swapped out with a different brand with the same or similar features. All pricing is based on cost and perceived value. The only way it could be more of a commodity is if someone came up with a way to plug any processor into any motherboard socket. Oh, and bonus points to anyone who can guess why the retail companies are moving away from separate box systems to all-in-ones. HINT: Look at the upgrade path for laptops.
There is no "-1 offended" or "-1 you don't agree with me" mod options for a reason.
If you want it to be, server hardware is already commoditized. All the priciest components are interchangeable (you *can* buy DIMMs from wherever and cram them in your server from a technology standpoint). Apart from Intel and AMD playing this game where the IOH is generally affine to some particular CPU generation, things are already there (and the IOH is a pretty inexpensive part however you slice it, and could already be subbed out for a PCI-e constructed device if they saw fit). Now the catch is that for *most* traditional IT shops, there continues to be value in well-integrated systems. This will not change that picture.
I see this as another step toward two goals:
-Getting ARM into the datacenter in some reputable fashion (which may or may not make sense, depending on whether a compelling performance per watt case can be made that offsets the energy/manufacturing gap that might be incurred from requiring more packages to get to performance desired).
-Rebranding 'whitebox'. White box vendors are viewed as the low cost alternative to HP/Dell/IBM, but image wise they are viewed anywhere between 'unacceptably bad' to, at best, 'just as good' from a select portion of the market. Putting cost aside, no one thinks of white box as 'better' than the expensive names. A lot of open compute at the system level is the same thing that has been the reality for the last decade with a shiny new name. The same standards that everyone already followed are getting highlighted more explicitly. This is the opportunity, through marketing, to change minds to say 'better' in some cases or at least make the 'unacceptable' segment of the market take another look.
XML is like violence. If it doesn't solve the problem, use more.
You could probably have an ARM, low load, low energy comsumption processor and a nice High-Performace processor on the same board. You'd just then manage when the high-performance activates, and you could probably switch any (assuming hot-plug) without taking it offline.... It's nice to dream, isn't it?
I don't care if I'm wrong. I only care about everyone obtaining something from the discussion.
do the cards have room for 4-8 / 6-12 ram slots each? and yes that's full size ram.
pci-e X8 is limited IO why not at least X16?
X8 can be used up by 1 video card on it's own.
I'll wait for Happy Ending.
Do you even lift?
These aren't the 'roids you're looking for.
Looking at the photos of the backplane, it looks like S100 era technology. Where is the trendy stuff? I want to see hermetically sealed illuminated glass-like blocks, changing color and sliding out automatically on detection of failure, a high-bandwidth optical interface on each edge, power inductively coupled to avoid metal connectors, an eerie surround sound voice saying "Dave ... Stop" ..
I compute
He computes
The computer is busy computing.
Open Computing or Open Computation is correct. "Open Compute" is not.
I suppose these people are the same set who would say "I have a drive license" or "The pilot passed his fly test".
The commodity interconnect that allows scalability is at the Ethernet layer, not the CPU socket. More and more of the motherboard is going inside the CPU's silicon. The integration there is less interesting and making it modular at that layer will only increase cost.
Faster networks, cheaper switch fabrics, and more compact low-power CPUs will have a bigger impact on the density of a cabinet.
http://pbfcomics.com/115/
It gripped her hand gently. 'Regret is for humans,' it said.
almost everything we see in consumer devices has been done before in some market, or at the NSA (the latter of which will not talk about it, but we know because of James Bamford)
i remember the psych department at the university had an 'old computer' historical display set up in one of their windows. the 'motherboard' was just a bunch of slots you would fit wire-wrapped boards into. one was the cpu board, one was memory, whatever.
not to mention all of the "upgrade your PC" cards from the 1980s - put a 286 CPU-on-a-card into your 8088 "IBM XT", heck you could even put a PC card in your Mac.
im pretty sure 'industrial' users like Airplanes etc have had similar setups.
I came along just a bit later, but this part especially is true:
"users could buy complete service manuals for their computers (and have a good chance of fixing them!), there was a ton of info about the OS's"
The virtual complete absence of true user manuals to this day baffles/angers me.
When I took 'computer class' in the mid-90s we still learned mostly in versions of DOS and we used 5 1/4 and 3 1/2 floppys (mostly the latter).
We could afford 2 computers that could run the current version of Windows.
Thank you Dave Raggett
And add in some optical links so we can finally scale motherboards to something awesome.
Being limited to certain designs / lengths because of electrical circuitry...madness.
I am John Hurt.
the digital group -
http://www.pc-history.org/digital.htm
http://www.bytecollector.com/the_digital_group.htm
This has never been done before with a CPU, which has traditionally required its own socket, its own chipset, and thus its own motherboard.
Unibus in 1969. S-100 in 1974. MBus in 1991. Even consumer-level PowerMacs had replaceable/upgradable CPU cards in the 1990s. HyperTransport in 2001. Yawn.
I think the is the basic idea, which is why the whole idea won't work. Basically, they are sawing the motherboard in 2, where the CPU and memory are on the daugterboard, and the rest of the components (SATA,USB3, PCIe slots, sound, video outputs) are all that remain on the motherboard
Why would it work any less than a graphics card? Isn't that the same? GPU and memory on a daughterboard with a fast interface to the motherboard.
Science is all about firing a drunk pig out of a cannon just to see what happens.
What I don't see in TFA is something that describes how the one big hurdle of this type of design will be overcome- the IMMENSE costs! The speeds that processors and RAM runs is so high you can't just drop it down on a board and expect it to work- you're in a long loop of simulate, build, test, repeat, and each iteration is extremely costly- we're not talking Arduino here- in reality, these boards (populated) in mass production will cost hundreds in just BOM costs, not counting the assembly. If the biggies in the industry are truly willing to foot the bill, great, but no matter what, these boards will still remain expensive, and likely still in the NDA wasteland as the individual parts that make it up are unlikely open up their documentation (or distribution chains) any time soon (yes, Broadcom, I'm talking about you).
Here's the link to the article from opencompute and my statement below that. http://www.opencompute.org/2013/01/16/ocp-summit-iv-breaking-up-the-monolith/
The interesting part about this is that we are just re-using "old" technology. Telecom did and still uses this method of workloads. Standardize the backplane and let the Intel's, the AMD's of the world to play on it vs trying to build your own chip......
"We had announced the same system that we built over a year ago, here is my interview on the product at the arm techcon. http://www.engineeringtv.com/video/ArmBlock-16-Packs-16-ARM-Cores. We also did a joint effort with Oracle Java team https://blogs.oracle.com/oslab/entry/16_processor_arm_system at the Oracle Open World and Java One. The base idea is that the backplane is the standard and any manufacture (intel, amd, arm, tilera, power pc, etc...) can be used on the system. We felt that by using Arm in this architecture would give the community the best representation of the power to performace capabilities. if anyone has any questions I can be contacted at Stephenm@miwdesign.com."
the digital group -
http://www.pc-history.org/digital.htm
http://www.bytecollector.com/the_digital_group.htm
Exactly. There was a run of DEC XL servers that could run Pentium or Alpha chips, just swap the CPU daughtercard. Not as old as your links, but I see nothing new here....
That's a lot of words to describe a simple cluster. If their clustering technology is better than the rest, then fine... but don't pretend like they fscking invented it, or we'll all have to pay licensing fees for decades-old tech, a la Apple.
A telephone support person: "Hey, I have a group hug user over the phone again."
A manager: "I can't take this anymore. Give him a virtual hug and send him to www.linux.org."