AMD OFFICIALLY BROKE SILENCE on the Fiorana and Kroner platforms, right on top of IDF. Most of the specifications are already known, but there are a few interesting things left to tell you about.
We brought you the first pictures of the platform at Computex, including the single socket and dual socket variants. The idea is simple, take the tried and true AMD 1207 platform and replace the bug-ridden Nvidia (guys, fix your SSD bugs, it is long past embarrassing now) or the almost-all-there Broadcom chipsets with an ATI variant. This brings AMD up into the realm of stable server platforms at long last.
There are three chipsets available, the SR5650, SR5670 and SR5690, all variants of the RS890 chipset, known on desktops as the 785G. While the name is halfway between the 780 and 790 chipsets, this one is based off the next generation 8xx silicon, not the 18-month old 7xx versions. The main difference between the three new SKUs is the PCIe2.0 lanes, with 22 on the -50, 30 on the -70, and 42 on the -90. Only the -90 supports hotplug though, something that seems rather odd not to include on the lower variants.
For power, the northbridges pull TDP of 13W, 17W and 18W for the -50, -70 and -90, respectively, and all tie in to the SP5100 southbridge with a 4x PCIe lane. That southbridge takes only 4W TDP, and has 12 USB2.0 plus 2 USB1.1 ports and 6 SATA plugs with RAID capabilities. If you squint a bit, it looks suspiciously like the SB710 in the 785G, imagine the coincidence.
The new platform is basically a way to validate the chipsets before the G34 and C32 platforms hit the channel late in Q4 or early in Q1 2010, depending more on the elves in the large OEM server validation labs than AMD. Since it is effectively shipping in desktop boards now, there are unlikely to be any showstopper bugs, but HP, IBM and Dell tend not to take chipmakers at their word on this.
When it is given the thumbs up, it should be shipping in a lot of socket 1207 boards, but that is not much to get excited about. When Socket G34/C32 ships in Q1 2010 or so, once again, OEM dependent, this will be the future of AMD.
Those new platforms bring DDR3 into the mix. Socket 1207 didn’t have the pinouts, but the CPUs themselves are perfectly happy to run with that memory. The next big feature is a return of the 8-way systems with a new name, MCM. Socket G34 is a very odd configuration, an 8-socket almost fully connected system, folded and twisted into a 4-socket MCM. The good news is that you effectively get two sockets for the price of one, and a bit more performance out of the deal as well. The bad news is that AMD gets paid for one socket instead of two.
When you have this many sockets and cores, you need to have more I/O than usual. 42 PCIe lanes doesn’t cut it for 48 cores, so AMD did what it pioneered in the x86 space, it allowed multiple northbridges per system. Even better, you can mix and match 5850, -70 and -90 parts based on needs, cost and power. Nehalem does support similar features, but board makers are telling us that the implementations so far leave a bit to be desired, especially if you desire stability.
Last up is the most important and also least understood technology in these new chipsets, the IOMMU, or AMD-Vi in marketingspeak. When AMD was slapping Intel silly on virtualization tests a year or two ago, that was because it could fully virtualize the memory controller while Intel had to eat cycles to make up for it. This gave AMD a healthy lead in one of the hottest growing segments of the market until Penryn based CPUs won out through sheer horsepower.
Nehalem brought the same memory virtualization capabilities to Intel, negating any technical advantage AMD had. AMD didn’t stand still, however, and the IOMMU breaks down one of the last barriers to virtualization, I/O virtualization in hardware. It isn’t a panacea though, you still need the I/O devices to support the technology, and of course the dreaded driver support. We expect NICs to get there first, followed by GPUs in relatively short order.
Why is this important? The holy grail (part III, subsection 7) of virtualization is graphics, and its absence is painfully obvious to anyone following the technology. With IOMMUs and supporting devices, you can have fully virtualized 3D apps running with GPU acceleration across multiple VMs. This will change the game quite a bit, pun intended.
A Kroner for your thoughts?
The other name that AMD talked about, Kroner, is not as much a new platform as it is a new form factor. It is a 1/2U version of a Fiorano server, not quite a blade, not quite a 1U. Both Supermicro and Tyan have been pushing these for a while, and they make a lot of sense for dense compute environments.
Narrow but long boards allow OEMs to stuff a bunch of things in that they couldn’t fit on most blades, from drives to cards, without giving up a full 1U slot. AMD was conspicuously absent from this space until Kroner, but now I expect a lot of 2-socket 1/2U white boxes to be popping up. When G34 hits in a few months, this will make for a dandy little 24-core per half-U solution.S|A