Server market will be a tough fight this year

It is nothing like desktops and mobile other than AMD's introduction of Barcelona.

Let's look at 2P servers first. The big news for Intel is the Stoakley chipset, and it brings, well, not a huge amount of big advances to the table. The nothable thing is upping the snoop filter to support the 6MB of cache per chip on the 2S Penryns, and making room for a large cache variant of Woodcrest and others, should it be needed.

It will also have up to 44 PCIe1 lanes or 21 PCIe2 lanes, more memory capacity, more memory RAS, and a lemony fresh scent. It is an improvement in most areas over Blackford, but doesn't break much new ground. We hear the on again/off again 1600FSB is on again, but I don't expect to see that before Woodcrest is supplanted by Wolfdale (Wolfdale-DP to be exact).

With a better chipset, the bounding boxes of Woodcrest and Wolfdale are enlarged a bit, but don't expect miracles. 3.33GHz Woodys are waiting for an AMD response before they are let loose, and you can bet the 1600FSB with an attendant 3.4/3.6GHz part will tip up if Intel feels threatened, it has the TDP headroom.

On the AMD side, we come back to BB and AB, before and after Barcelona. Before it is a yawner, 3.0GHz in February, 3.2 on a 90nm process in Q2, and then Barcelona. As that part will be backwards compatible to one degree or another with current platforms, we don't expect to see many more K8 CPUs after that, just backfilling with 65nm parts.

Mid-07 will also bring the first new platform, AM2+. The main thing this brings is split power planes to run the Barcelona core and uncore parts at different voltages. This lowers TDP and ups the frequency headroom. Both have been thorns in AMD's side for a bit because of the need for single voltages across the chip. AM2 parts will work in AM2+ sockets and the other way around, so don't look to this transition as a major stumbling block, just lack of realised gains if the partners screw up the chipsets.

The new core will bring AMD back into the hunt in most respects - core for core. I think it will have a commanding lead on FP workloads but lose on most Int based benchmarks. Because most server loads are Int, this will hurt AMD disproportionately.

Here is where things get confusing. We said core for core, clock for clock, and we think AMD will be right there in the clock and power race for the second half of 2007. It won't be the lopsided pummeling of the past couple of quarters, but it won't be AMD vaulting into the lead, core for core. We defy you to find a single core server based on either of these chips, though.

The first way AMD is going to pull ahead, sometimes, is the chip architecture. AMD will have a quad core part, Intel will not. This is not really a big deal. On one core, AMD will lose most benchmarks, on two cores, it will pull a little closer in, and at four cores, will probably win many more benchmarks. This is for 1S systems.

On 2S, the same holds true, but AMD starts out in a better position. 2S 2C will be an Intel win, 2S 4C will probably be very close, and 2S 4C will be a win for AMD in most cases. The platform and the bus come into play here, and Intel has made admirable gains to take the creaky old P4 bus and twist it into something that can hold it's own.

On 4S systems, well, it is still an AMD clean kill, be it now or after Barcelona. Four way systems are AMD's playground, and until Tigerton/Clarksboro come on line in Q3, it will stay that way. This all assumes one big "if" in the year, if Barcelona comes out on time, on target, and in volume. We can't understate how important this is. If AMD fails here, it is time to pack up and go home.

That brings us to the second big server intro of the year, the first peek at what the Merom cores can do in a 4S server. That is Tigerton, and having played with it first hand, I can say it is fast, stable and runs all 16 cores at speed. Compare and contrast that to AMDs Barcelona demo and you can see why I am worried about AMD's execution, especially with the stakes as high as they are.

I expect that Tigerton will scale pretty well, Clarksboro is basically two Blackfords spot welded together with the relevant bits extended from two FSBs to four. It will be fast, power hungry and require a lot of pins, but did we mention it will be fast? If cost and heat are your big problems, you might want to look elsewhere, but it will be fast.

The Tigerton vs Barcelona will be the second big match of the year, and it will really be a close war at 4S, this one is really too close to call. Again, AMD's unquestioned supremacy in 4S is over, the question now is will it keep the lead? Probably it will, but again, it all comes down to Barcelona execution.

The last thing big thing in 2007 is Socket AM3. This one enforces the split voltage planes that were optional on AM2+, and AM2+ CPUs will work with AM3 mobos. AM3 chips will work on AM2+ mobos also, but with some features lost. It again should be an easy move, but will only start to appear in the waning days of 2007. Don't look for any real impact here, it will be low volume in low socket counts.

The other trickles are the Intel 45nm server products. Once again, they will be out in low volumes in Q4, look for the real ramp in 2008. The Wolfcrest-DP and Yorktown parts will bring lower power, much better FP, and most likely a 1600FSB. They will have larger caches, new instruction sets like SSE4 and Rockton, and generally be better in every way. The biggest kick there will be in raw clock speed.

One last bit to keep in the back of your mind is memory. AMD uses registered ECC DDR2 while Intel uses FB-DIMMs. AMD can clock up the memory if it wants, currently it is at 667MHz for servers, but 1066MHz is looming for desktops, so speed bumps can come if AMD wants to go there.

Intel has 533 and 667MHz FBDs with 800, the next logical step, not on the horizon. From what we have been hearing, it is not going to happen either, most likely due to power. The 533MHz to 667MHz move needs a lot of juice, and 800MHz most likely takes more. It will get ugly at 16 sticks of memory.

This handcuffs Intel performance a bit, or at least puts a ceiling on it. It is far from the end of the world, but every little bit adds up, good or bad. To fix the power problems, Intel will introduce several technologies on FBD to cut power use, shutting down unused lanes and other things. Goldbridge 2 will be a big step in the right direction, but it won't be enough. The big changes in memory come in 2008 with the advent of micro-buffers.

2007 will be split in two, the first half will see AMD getting slapped around by Intel with no real recourse. It will be closer than it is on desktops, but still not pretty. With Barcelona, assuming it is on time and target, things get interesting again. Who is on top depends on a complex matrix of cores, sockets and workloads more than any single metric.

Later on, we see a renewed Intel in the 4S space, and from that point on, it is a fight from top to bottom. No clear winners, no clean kills, just all out war. Servers is by far the most interesting element for microprocessors in 2007, no question of that. µ