THE DETAIL we presented here last year about the first Nehalems - in particular, the 2008 shipping flavours, the Bloomfield UP and Gainestown DP quad-core CPUs using the Tylersburg chipset - proved to be pretty much spot on during Intel's latest enterprise press talk earlier this week.

As we said before, these early birds come with unique three-channel DDR3 integrated memory controllers and one or two QPI links, each at up to 6.4 gig transfers/s - 25.6 GBytes/s bidirectional. We'll talk about the CPU core features soon - what are the limits of these processors' memory capabilities, and can we expect some true records?

Supported speeds
Officially, at this moment, Bloomfield and Gainestown will support up to DDR3-1333 memory speed in both standard and registered varieties (servers only for the latter), both ECC and non-ECC. Does this sound sluggish, knowing that this CeBIT saw OCZ, Corsair, Kingston and others show DDR3-2133 memories, running at CL9 no less, on Nforce 790i?

Not really - the first DDR3-1333 CL5 ultra low-latency parts are out, some of them in 2 GB density per DIMM module. On the desktop Bloomfield, with three channels and two DIMMs per channel, this gives you a comfortable - even for Vista bloatware - 12 GB RAM. The bandwidth? A whopping 32 GB/s total peak memory bandwidth across all channels, same as dual-channel DDR3-2000.

Furthermore, there is no FSB bottleneck that lets current Penryns access only half of that huge dual-channel DDR3 - the three channels in Nehalem are, through the on-board memory controller, accessed directly by the CPU's L3 cache. Even if Intel just bolted on a 3-channel version of X48 chipset's DDR3 controller, and removed the FSB blockage, you'd still be able to get at least 25 GB/s Sandra memory bandwidth scored with such DDR3-1333 memory configuration.

What about the dual socket part, the one to power the Skulltrail successor? Well, Intel's slides show it having slightly over four times the bandwidth of the best current dual-socket Harpertown with four FBD-800 channels. Knowing we easily get 9 GB/s there, we're talking about a 36 GB/s dual-socket workstation Sandra memory bandwidth score - a massive leap of nearly three times over the best AMD systems too.

Breaching the 100 GB/s barrier
Since this is a programmable memory controller, and Intel already supports DDR3-1600 and above via XMP SPD readings, I don't think Intel will miss a chance to enable even higher performance via support for higher speed memory.

Even with higher latency - if for some reason we used DDR3-2133 memory, the fastest for some time to come - we'd still get somewhere around 35 GB/s Sandra on the Bloomfield, and 50 GB/s on the dual Gainestowns. At that speed, we could only have one DIMM per channel, but well, 6 channels are there and you can still fit 12 GB in there on the 2-socket machine.

The peak results, while useless in real apps, are even nicer to read: a three-channel DDR3-2133 memory configuration would give you 51 GB/s peak memory bandwidth per socket, or a massive 102 GB/s on the dual-socket Gainestown - the very first for a desktop or workstation PC. On top of it, there is still the QPI interconnect bandwidth for remote memory access from that other socket.

OK, some may say, these are extreme cases, but still, even if you want to populate 2 DIMMs per channel and run twelve 2 GB DIMMs on your future Gainestowns for a very comfortable 24 GB RAM - presumably just enough for the next version of Windows to boot - there is still enough headroom left over.

The current X38 and X48 handle two DDR3-1600 CL7 DIMMs per channel (total four per mobo) very nicely, without ridiculous voltage pushing. The Nehalem memory controllers should be at least that good. That means, a full-capacity 24 GB six-channel, eight-core Gainestown based Skulltrail successor will still give you 76 GB/s peak memory bandwidth, or well over 40 GB/s real Sandra bandwidth - over quadruple the current Skulltrail with the best FB-DIMM memory I could find.

Based on the available info, I very much doubt that the Deneb 45 nm Phenom memory performance will come anywhere close to the Bloomfield - not to mention the Shanghai AMD server part vs the Gainestown, at least on DP, even on AM3 socket DDR3-enabled parts.

Now, the high-speed DDR3 just has to come down a little in price, and Intel's Nehalems will fly off the shelves. µ