Intel releases Tigerton

TODAY, INTEL launches Tigerton on Clarksboro, the first Merom derivative for four sockets. AMD had essentially a free ride in this segment until now, but starting today it faces really stiff competition.

There are two parts launching, the Tigerton chip, essentially a Clovertown with the ability to do four sockets not turned off, and the Clarksboro chipset. Together, they make the Caneland platform. A good way to think about them is take two Woodcrest on Blackford systems and spot weld them at the chipset. That is more or less Caneland.

We told you about the chips earlier, they come in six QC variants and eight DC ones. The DC parts are called the 7200 series, E7220 is the 2.93GHz variant, E7210 is the 2.40, both have 8MB of cache, consume 80W and are on a 1066FSB.

The 7300 series are the quad cores, and all of them are also on a 1066FSB, but things get interesting from there. The E7310 runs at 1.60GHz and the E7320 hits 2.13GHz, both are 4MB cache and consume 80W. The E7330 is a 2.40GHz part with 6M of cache and it also takes 80W, pay attention to this one. The E7340 has 8M of cache but is exactly the same otherwise.

Below those, they have the L7345, a 1.86GHz 8M 50W part, and a 2.93GHz 8M 130W X7350 at the top. For those of you that need XTreme Powa, your X has arrived for quads. Other than that, the usual opaque Intel naming schemes are now that much more confusing.

There are several interesting new twists to these CPUs, the cores themselves are exactly the same as the current Clovertowns, so why is that odd? The 2C 8M parts for one, there is no variance in cache size across the dies other than 2MB or 4MB versions. These two parts fit neither mold.

The 6MB part is probably a fused off 8MB version, it is not a new die. This one is almost assuredly done simply to allow for more SKUs, an artificially made up part because the Merom core can only do full clock dividers, IE 266MHz steps on a 1066FSB. If marketing demands 4 SKUs, there is little else you can do. It also points to 2.66GHz parts being over the 80W cap, so that one would be out, there can only be one Xtreme, or it is eXXXtreme edition, and if 2.93 fits, use it.

The other interesting ones are the 2C 8MB parts, and again, Intel did not double the cache on a Woodcrest, they turned off cores 1 and 3 on a Clovertown. This means each core does not share the cache, they each have the full 4M for themselves. It is a dual die quadish core with a bunch of fuses blown.

The chipset is also quite interesting. It has four FSBs so each core is on it's own dedicated link, this should be very fast but suck a lot of power. It also has only four FBD channels, but each can do eight FBDs, raw DIMM count will not be a problem. RAM bandwidth per core is half what Clovertown + Blackford is, so this might be an area to watch, especially without larger caches.

Blackford had something called the snoop filter, essentially a cache that allows you to not send out cache snoop traffic from one FSB to the other if the data was not there. It was a good idea in theory but didn't work out all that well in practice. Some variants of Blackford/Greencreek actually had it turned of.

Clarksboro has it revamped, turned on and enlarged. Aside from the unstated fixes to whatever needed fixing, it is able to see four times the cache. Blackford could only see 16M of cache, enough for two quad cores with a total of 8M cache per socket. If you went any higher, the chipset would have a fit, so that is why you never saw large cache Woodcrests or Clovertowns, the chipset would not handle it.

This was rectified in Clarksboro, this chipset can see 64M of cache. What this means is Intel is not limited to 8M cache per socket, they can support 16M now. Does this mean we will see 'large cache' Tigertons? Probably not, but I wouldn't put it past Intel to do so on the Penryn 4S variant. Either way, this chipset should take 45nm 4S parts when they are available, they can see the full 12M per socket.

So, how much power does this chipset consume? My original math said it was a lot, if you take a 35W Blackford and double it, you get 70W, lose a bit from a process shrink, call it 60. Bump up the speed a bit to allow for routing 4 FSBs instead of two and you are easily above 60W.

My internal math was wrong, it only takes 47W, still a lot, but nowhere near as bad as I thought. Part of the savings comes from the drop in FSB speeds from 1333 to 1066, and losing 4 FBD lanes doesn't hurt either. In any case, the chipset take a lot of power, but you do get high performance for your electrons.

One other little bit that may end up as a sticking point is PCIe, the chipset only has 28 lanes. This is probably enough for most applications, I can see it becoming limiting, you can only put in 3 x8 cards, two raid and a 10GigE perhaps. It comes back to my long standing gripe about Intel gutting the chipset market, there is no plan B anymore, so if 28 isn't enough, tough.

Back to the more pedestrian parts, pricing. The chipset itself, Clarksboro, will sell for $224 in 1K quantities. The dual core Tigertons will cost $856 and $1177, the low voltage 7345 will be $2301, as will the XxxTreem 7350. The other mid range quads go from $856 to $1980.

In the end, I think Intel has a winner here. The same things they did with Clovertown in the 2S could happen again in the 4S space. Chipset power, FBD and PCIe channels, and potentially FBD power are all potential sore points, but in the end, performance is what counts, and on that front, I am pretty sure Caneland will deliver.

Because of the nature of the beast, you won't see much more than canned high end benchmarks, but that is OK, these are not gaming boxes. The next few weeks will be quite interesting, a 4S Caneland vs a 4S Barcelona should be quite a fun shootout. µ