SeaMicro fabric could be the glue for future AMD chips

The reasons for AMD’s decision to purchase SeaMicro are coming into focus quickly. AMD’s John Fruehe has written a blog post explaining what AMD gets in this acquisition, and it’s worth reading. Most notably, Fruehe says the key technology involved is SeaMicro’s interconnect fabric:

The basis of the SeaMicro technology is the interconnect fabric. A custom ASIC allows multiple computers to be linked together via their PCI Express bus. This fabric allows a large number of servers, to be tied together in a fabric within a dense chassis, all sharing a high-performance connection to networking and I/O peripherals. Servers go from being a large box to a small card as they are aggregated into this environment.

Fruehe then makes it very clear how this technology ties into the "SoC approach" to chip design that AMD’s execs communicated at the firm’s Financial Analyst Day. SeaMicro’s fabric may be the glue that binds together future AMD chips:

At the most recent AMD Financial Analyst Day, Mark Papermaster and Lisa Su both spoke of the new AMD direction towards SOCs (system on a chip) that will allow us to take a wide range of IP and integrate it into purpose-built SOCs that meet particular computing needs. With SeaMicro we now have the fabric to pull all of these together as well. The SeaMicro IP that helps connect all of these computers together could one day be built directly into these SOCs, providing an easy way to tie systems together with the smallest silicon footprint possible. But not just AMD SOCs, the technology will be available for others as well.

More immediately, perhaps, AMD can begin offering high-density server solutions with much tighter integration at the system and rack level. Writes Fruehe: "Clearly, we are moving from a being merely a silicon provider to delivering systems-level integration and capability." That is, of course, the lion’s share of SeaMicro’s current business. However, AMD doesn’t plan to compete with its customers in any great measure, as Fruehe explains:

Does this mean that AMD is heading into the systems business? Not at all. We have partners like HP, Dell, IBM, Cray and a host of others that design and deliver systems better than anyone, so we’ll really leave that business to them. We’ll maintain the SeaMicro business and current customers to ensure that we continue to deliver the products that they need to run their business. But the long term the goal is to broaden the offering by also integrating the technology into platforms from our OEMs, creating a larger ecosystem of products for customers to choose from.

So it seems AMD’s current partners will get access to the SeaMicro fabric technology, potentially enabling them to create high-density server solutions, presumably based on AMD microprocessors. Meanwhile, AMD remains largely a technology firm, not a server vendor.

Comments closed
    • ghjtdge
    • 8 years ago
    • smilingcrow
    • 8 years ago

    It has sadly come to this; AMD are so desperately lost that they are now resorting to using glue to fix their CPUs. I look forward to the FX 8756 UHU version.

    • dpaus
    • 8 years ago

    I think this is a major investment in Bulldozer/Piledriver’s future. With their new GPGPU architecture and this interconnect, Bulldozer cores could become ‘process aggregator’ nodes in extremely scalable compute-clusters.

    ‘could..’

      • Flatland_Spider
      • 8 years ago

      It will be interesting to see how people incorporate the GPU into their code via OpenCL. I’m not sure how it can help, but I’m sure other people have better ideas.

    • bcronce
    • 8 years ago

    The new enthusiast computer. Motherboard is just a backplane with a few slots. Then you purchase Compute “modules”, each with some CPU sockets and memory.

    Like a mini-cluster.

      • dpaus
      • 8 years ago

      1982 called; they want their S-100 bus systems back.

        • khands
        • 8 years ago

        Everything old is new again I swear.

          • Grigory
          • 8 years ago

          Calling your mom as we speak.

            • Arclight
            • 8 years ago

            Well played

            • khands
            • 8 years ago

            Too much plastic in her now, wouldn’t be the same.

      • BaronMatrix
      • 8 years ago

      Don’t steal my ideas…

    • willyolio
    • 8 years ago

    going from system on a chip to a server on a card… neato!

      • TheMonkeyKing
      • 8 years ago

      Does this mean instead of BTO (Built to order – like Dell), we will have the option of FTO (fabricate to order)? By this I mean AMD will FTO SoCs (ooops way too many acronyms) for other companies? Because I do not see AMD just building or designing one-offs unless that one-off is for licensing to another company to manufacture.

      I don’t see this being an option for the Enthusiast crowd, unless it means that you could be several pre-built SoCs that can connect together. (No more buying processor+motherboard+ram at Newegg but more like buying SoC “A” + SoC “C”?)

    • Anomymous Gerbil
    • 8 years ago

    Possibly a dumb question, but how does this relate to HyperTransport?

    For example, is HT purely an on-chip / between-chip solution, whereas SeaMicro can be used between boards/systems/racks/etc?

      • The Dark One
      • 8 years ago

      In a completely uninformed guess, I’ll say Hypertransport might be too expensive in terms of die area or power draw or something.

        • mesyn191
        • 8 years ago

        IIRC the problem with HT vs PCIe/Infiniband/whatever is that it used too many pins to move the same amount of data. Given that CPUs often have dual or quad memory controllers on die along with other stuff maybe they’re worried about becoming pad limited.

        Could be too AMD just got tired of trying to improve HT. I think they’re still about the only guys who use it anymore even though years ago it looked like for a while they were going to get some serious 3rd party support. Some companies even made socket compatible add on FPGA’s and crazy dedicated task co-processors that used HT to talk to an Opteron that was used to boot the OS and such while all the heavy lifting was done by those chips. I think this was around 2002-2004. Cool idea but I guess it was a flash in the pan.

      • My Johnson
      • 8 years ago

      Yeah, HyperTransport popped into my mind also. I guess it was never as scalable as PCIe?

      • mesyn191
      • 8 years ago

      There were actually a few systems made with HTX slots several years ago. HTX = HyperTransport Slot. IIRC it was just a PCIe/X slot turned around backwards and it connected directly to the CPU. Haven’t heard much about it in a long time, they made it really easy to do cool stuff with custom processing modules and certain types of scalable main frames. I think Cray did some huge 8000+ CPU Opteron systems by using them along with their own secret sauce bridge chip.

      Wiki has some info. and pics of the slot, yea its a flipped PCIe x16 connector with another PCIe x1 connector in line:
      [url<]http://en.wikipedia.org/wiki/HyperTransport#Add-on_card_connector_.28HTX_and_HTX3.29[/url<] Anyways HT 3.whatever is pretty good at scaling up to 16 CPUs, I think 32 CPUs is doable too but only makes sense for a few specialized computing tasks so they're rare to see. For systems greater than that you needed some sort of bridge chip otherwise you got hit with a big performance penalty.

        • Flying Fox
        • 8 years ago

        Yup, I think this pretty much spells the end of the HTX effort. I see pretty much the same rhetoric when they unveiled the HTX architecture and talking about co-processors and specialized applications gluing stuff together.

          • shank15217
          • 8 years ago

          Or maybe not..

          [url<]http://www.hypertransport.org/default.cfm?page=ProductsProductsByCompany[/url<]

    • sschaem
    • 8 years ago

    This deal kind of make sense, beside becoming an Intel system provider… (Thats twisted, specially if that become a decent source of income for AMD)

    What make this highly questionable is the price. I would have been taken back with a 50+ million deal, but close to over 300 million dollar for a pcie interconnect asic seem so wrong for a company that design complete chipsets (including pcie3) and server processors.

    And thats 334 million for IP that currently only work with Intel CPUs.
    So how much more R&D, for how long, to make the seamicro IP work with AMD chips?
    Anyone know how many people this will add to AMD payroll ?

      • Alchemist07
      • 8 years ago

      It shouldnt matter how many people this adds. Assuming SeaMicro were profitable, they should give AMD enough additional revenue to cover their own employees…

      • mesyn191
      • 8 years ago

      If they’re just using a modified PCIe bus/logic it should be easy to get that working with their own chips. There is likely nothing exotic going on there per se.

      The real problem and time consuming task is validating the fabric with their server chips. If it was just down to getting something that would work AMD could probably have something out based on this tech late next year. Due to validation it’ll probably take at least nearly 2 years to get something out for servers with this stuff I bet.

        • Washer
        • 8 years ago

        So, 300M for something you just dubbed non-exotic and the majority of the work to get a real shipping product comes from within AMD? Yeah, makes even more sense now…

          • mesyn191
          • 8 years ago

          The $300M is likely mostly for the IP, not so much for the fabric and engineering expertise they’re getting. AMD could’ve built something similar to what SeaMicro with their own resources and expertise but they likely never could’ve released it without paying SeaMicro anyways or getting sued. I guess they figured this was easier/cheaper.

            • dpaus
            • 8 years ago

            There’s also the speed-to-market factor to consider; even at AMD’s relatively small size vis-a-vis Intel, every month is likely worth $100 million to them, so this is small potatoes. Also, at least some part of the $300m would be for existing customers, who are providing at least some cash flow (which has [i<]some[/i<] NPV).

      • Flatland_Spider
      • 8 years ago

      Well, AMD has always been an Intel [u<]compatible[/u<] system provider. 🙂 The tech is pretty chip agnostic. Here's the article that mentions the tech is a CPU agnostic PCIe link to the magic ASIC. (http://semiaccurate.com/2011/07/18/seamicro-ups-core-count-to-768-per-server/)

      AMD also gets people who have an expertise in dense servers and the problems that come along with that. The biggest beneficiaries are probably going to be Tyan, Asus, SuperMicro, and the other whitebox manufacturers. AMD can create reference designs, and those guys won’t have to do as much engineering work.

    • bcronce
    • 8 years ago

    AMD has been making quite a few good long term decisions for the past few years, but they have this nasty bump because of Bulldozer. Once they get past it, I see a bright future.

    • tbone8ty
    • 8 years ago

    AMD just serv’d Intel in broad day light

      • mnecaise
      • 8 years ago

      Not so fast… Intel bought the Infiniband assets end of January. This might be reactionary, in order
      to stay competative.

        • Arclight
        • 8 years ago

        So…best case scenario AMD will still be mediocre?

      • maxxcool
      • 8 years ago

      Served mocha and a scone you mean… this is a life support move at best, meant to push them into a new niche that intel only has marginal interest in.

    • chuckula
    • 8 years ago

    What’s funny is that AMD has just bought a company whose entire product line is built around Intel CPUs.

    Even if SeaMicro completely cancels all its products tomorrow, AMD will be on the hook for the support contracts for Intel servers that are already out in the field 😛

    SeaMicro has some good IP and know-how, but it’s going to take time for that to translate into new AMD-based products that take advantage of what SeaMicro knows how to do well.

      • mnecaise
      • 8 years ago

      sorry for the meme reference…

      Imagine a Beowulf cluster of… Bobcat cores (Brazos cpus). Now they can build it.

    • Bensam123
    • 8 years ago

    So, they’re trying to do blade servers?

      • Corrado
      • 8 years ago

      Yes but much smaller. Think graphics card sized with 7 of them in a single ATX size back plane.

        • Palek
        • 8 years ago

        Yeah, [url=http://www.seamicro.com/sites/default/files/SM_DS04_v3.6_0.pdf<]this PDF[/url<] on their website contains a couple of pictures showing just how small a single SeaMicro "blade" is.

          • Bensam123
          • 8 years ago

          Interesting, it brings back the idea of coprocessors and daughter boards… It makes me wonder what would happen if you simply plugged one of these into a normal system, given it’s PCI-E complaint…

      • Bensam123
      • 8 years ago

      I got rated down for comparing them to blades, WTF XD

        • khands
        • 8 years ago

        Techreport voters are on the fritz today.

          • Bensam123
          • 8 years ago

          I gave you -1 to fit in with the rest of the folk. 😀

    • lilbuddhaman
    • 8 years ago

    And the first few revisions will only be able to do 8x, and once they finally get 16x working it’ll only be a small performance increase >_<

    edit: Geez no love for an admittedly bad pcie joke

      • khands
      • 8 years ago

      Yeah, tr readership needs to lighten up.

        • Grigory
        • 8 years ago

        Looks like we lit up lilbuddha. 😉

Pin It on Pinterest

Share This