Intel has always been the first to master newer, finer CPU process technologies. That wasn't always easy, especially when they introduced architectural changes at the same time. Prescott is a fine example. For this reason, for the last few years, Intel has been strictly alternating microarchitectural and newer process technologies. The scheme isn't new. For example, company's first desktop dual-core processors were basd on the Smithfield die, which was a doubled Prescott made using the polished 90nm process technology. The transition to 65nm was marked by the rollout of Presler that had the same architecture as Smithfield. It was followed by the fundamentally different Core 2 series. However, having major architectural differences, Conroe was made using the same 65nm process technology. When they needed to move to 45nm, they made Wolfdale, which was the same Conroe (to a first approximation), just 45nm. This very process technology was used to make the first Nehalem CPUs rolled out late in 2008. In other words, they either moved a mastered architecture to a newer process technology, or based a newer architecture on a mastered process technology.
These days, it's time to move from 45nm Nehalem to 32nm Westmere. Obviously, there's no sense in searching for some inner differences between those two. The purpose of Westmere is to provide transition to the 32nm process technology. Why do they bother to master new process technologies at all? Well, other things being equal, that has the following advantages:
It's not possible to get all of the above at once, because the advantages are mutually exclusive. So they have to choose some they consider more important and combine them to a certain extent. Here's an example. Clarkdale mainstream CPUs rolled out in January are the first Westmere-based products. Since mainstream products should be inexpensive, Intel both used a finer process technology and halved cores/cache compared with Bloomfield and Lynnfield. As a result, while the latter were aimed at the over $200 segment, the former (while also having graphics cores) could even be cheaper than $100. Power consumption was also reduced, in spite of increased clock rates -- to make up for performance lost when halving the number of cores.
Now we welcome another product of the Westmere family -- Gulftown -- aimed at the top, primarily server, market segment.
Why did they jumped from the low-end to high-end, bypassing mid-end? The matter is that Lynnfield doesn't need to be replaced yet. CPUs based on that core have no rivals and are selling great anyway. Of course, moving them to 32nm could reduce cost prices, but that's unimportant right now. In turn, the rollout of Clarkdale was required, because inexpensive Core 2 CPUs need replacement, while competing products needed an alternative.
Gulftown is also required, because there can't be too much performance where servers are involved. Besides, the platform has had its rivals, from both Intel and AMD. So, of the aforementioned advantages the 2nd and the 4th seem the most interesting. Of these two the 2nd is more important, since these newer CPUs are primarily intended for servers, a field where multithreading is not a novelty. Moreover, Intel has already released Xeon 7400 (Dunnington) hexacore processors before, so decreasing the number of cores down to 4 in Xeon 5500 can be considered a step back.
AMD has already announced 12-core Opterons. Those have low clock rates, but can provide even a dual-socket machine with fair 24 cores capable of processing 24 threads simultaneously. The same amount of threads can also be handled by a machine with two Xeon 5600 CPUs. The latter have only 12 actual cores in total, but offer higher specific performance. Though it's clear that such serious AMD Opterons, as well as Intel Dunningtons, are primarily intended for machines with 4 and more sockets. So they'll have to compete with systems based on Xeon 6500/7500 solutions to be rolled out this spring. Still, increasing the performance of dual-socket servers and workstations may also come in handy.
That's why Gulftown is what it is -- about 1.5 x Bloomfield: 6 cores instead of 4 and 12MB L3 cache instead of 8MB. Correspondingly, it has 1170M transistors instead of 731M. Despite that its die size is even a bit smaller 248mm² instead of 263mm² -- thanks to the 32nm process technology. Otherwise this monster would've been too large and expensive. They might have had to reduce its clock rate or redesign cooling system. But none of that is needed the way things are now. The new Xeon 5600 CPUs fit into the same thermal envelopes and have comparable clock rates as the old ones. They have similar prices, too. And can also work in pairs, offering 12 high-frequency cores, each supporting 2 threads. In other words, a machine based on two Xeon 5600 CPUs will rarely lose to the one based on four Xeon 7400 CPUs. Or it might not lose at all, because even Xeon 5500 did great.
However, you can't say Intel has only been interested in performance. The new series offers solutions to fit any taste. It's crowned by the top-end hexacore Xeon X5680 3.33 GHz and quadcore Xeon X5677 3.47 GHz. Despite the prefix "X", both CPUs have 130W TDP and the same release price of $1663. These processors replaced the former leader W5590 3.33 GHz. In other words, today, you can get 1.5x cache and either 6 cores at the same clock rate or 4 cores at higher clock rate for the same price and in the same thermal envelope.
One step below ($1440) are hexacore X5670 2.93 GHz and quadcore X5667 3.06 GHz (both have 95W TDP). These two obviously replace X5570. In turn, X5560 and X5550 are replaced by X5660 and X5650, respectively. The novelties have the same clock rates and TDP, but feature six cores.
Another hexacore CPU, E5645 2.4 GHz (80W TDP), has been added to the E series. The older products -- E5540, E5530 and E5520 -- were succeeded by quadcore E5640, E5630 and E5620 with the same clock rates, but larger caches. The release prices are also the same, by the way.
The power-efficient family has also been updated, and considerably at that. Hexacore L5640 2.26 GHz and L5638 2.0 GHz both have 60W TDP, while quadcore L5618 and L5609 have the even lower TDP of 40W. The latter processors have the same clock rate of 1.86 GHz. The first of those supports Hyper-Threading and costs $530, while the second doesn't, but it costs $440.
So, as we can see, new Xeon 5600 processors replace older Xeon 5500. They work in the same motherboards, have the same TDP, but offer at least larger L3 cache. In many cases they also provide more cores or higher clock rates. In other words, you get more bang for the buck.
By the way, newer CPUs will perform better in certain tasks as well. The matter is that while Westmere doesn't fundamentally differ from Nehalem, it stil has something extra. Namely, the special AES-NI instruction set that speeds up encryption and decryption of the popular Advanced Encryption Standard (AES). As you might have already guessed, the instruction set made its debut in Clarkdale CPUs -- only to boost those by more than a factor of ten in synthetic tests. Of course, the boost is smaller in real applications, but even 5%-10% can come in handy. Moreover, even some "casual" programs (e.g., 7-zip and WinRar) support AES-NI already. As do modern operating systems: Cryptography API: Next Generation implemented in Windows Server 2008, as well as desktop Windows Vista and 7, also has AES-NI support. Which means any application using this API will work faster despite the intentions of its developer. Considering that the new expansion is supported by most Intel's server processors (excluding just Xeon L5609 and the ultra-affordable Xeon L3406 for Socket LGA1156) and many desktop CPUs, Core i5-6xx in particular, its popularity among software developers will only grow.
We've been mostly talking servers so far. Well, LGA1366 was initially positioned as a dual-socket server platform. But certain force majeure circumstances made Intel reposition it for single-socket servers and workstations and even for extreme desktops as well. The latter segment is now backed up by LGA1156, so things are on the way to proper order. Not letting multicore Westmere CPUs into the single-socket segment at all would've added even more reason. However, Intel had repeatedly promised to roll out an extreme desktop hexacore CPU, and they actually made one such processor. So now there are 8 quadcore and 7 hexacore 32nm CPUs for dual-socket machines, and 0 quadcore (!) and 1 hexacore CPU for all other segments. Well, you can say there are two: Xeon W3680 and Core i7-980X Extreme Edition are "off-cuts" of Xeon X5680. The former lacks one of the two QPI links, the latter also lacks support for registered memory. These models are also limited to 1066 MHz memory, traditional for the desktop LGA1366. However, the price is also $999 instead of $1663. Besides, in case of Core i7-980X, the missing features are largely made up for by users' freedom to adjust any multipliers -- one thing usual to the Extreme Edition series. Meaning you can easily overclock such a CPU while keeping bus's actual frequency or use high-frequency memory, e.g., DDR3-2200. All of the above makes Core i7-980X a very interesting product. Obviously, if you're willing to pay the price of a decent machine for a CPU alone. What exactly will you get for your money? Let's find out.
Note that we used a Xeon X5680 in these tests instead of a Core i7-980X. This shouldn't affect the results much, because we didn't mess with multipltiers or anything else. We just tested the processor in two modes: with Hyper-Threading enabled and disabled. Obviously, the latter mode isn't typical. Hyper-Threading is required for dual-core processors and is handy for quadcore ones in some tasks. But six cores are quite enough for a desktop machine, so why not see if you actually need Hyper-Threading in this case?
Note the question marks in the Turbo Boost column. They mean we still don't know for sure how this feature works in X5680. We only know that the maximum miltiplier increase is 2 steps (when 1 core is used). The pattern of Core i7-980X is 2-1-1-1-1-1, but we know that Xeon X5500 used to be more "aggressive" than Core i7: 3-2-2-2 vs. 2-1-1-1. It seems this advantage is no more, but we can't say for sure.
We compared four processors.
The selection of motherboards is easy to explain. We have already mentioned that both new Xeons and Core i7 can work with the existing infrastructure. Though DX58SO had no official support for any Xeon processor of 130W TDP, it worked with Xeon X5680 like a charm with the BIOS released on February 24, 2010. Just the available settings were greatly limited. We could only enable 1, 2 or all cores, but not 4 (fortunately, all 6 worked in the automatic mode). As for RAM, only 6 and 8 multipliers were available -- similar to those for the non-extreme Core i7 CPUs (older Xeon X5500 processors also support the multiplier of 10 in the manual mode). However, in the automatic mode, RAM worked at the desired 1333 MHz, so there were no problems with that. We believe this should allow one to replace an older Core i7-920 with a newer extreme CPU. Given anyone wants to, of course.
Write a comment below. No registration needed!