Would You Like Some GPU With That?

Today Intel is officially unveiling its worst-kept secret in a long time – a pair of new CPU families based on a new version of Nehalem microarchitecture: Clarkdale (not to be confused with Clarksfield, which is the mobile variant of Core i7) and Arrandale. We’ll only be looking at Clarkdale today, as Arrandale only pertains to mobile processors. Both fall under the Westmere fab process (I just love all these code names, don’t you?).

Today’s Clarkdale launch marks the first die shrink to 32nm. Perhaps more interesting though, is the fact that the processor also houses a GPU, right on the chip. That’s right, after finally moving the memory controller from the North Bridge to the CPU, Intel has decided to move the GPU there as well.

If anything, this move simply sets things up for integrated Larrabee, because I can’t think of too many reasons why this would be useful for the end user. You will still need to buy a certain type of motherboard to make use of this GPU, and it’s doubtful that those will be much cheaper than a standard “G Series” equivalent motherboard. I believe this does completely (and legally) cut off 3rd parties from making IGP chipsets for these processors.

The new GPU (which has no discrete name by the way – it’s simply titled “GMA HD” for now – the model number is taken from the CPU) has a few quality enhancements to speak of, compared to the previous top chip from Intel (GMA X4500HD, found on G45 and G43 chipsets):

• Dual video HW acceleration
• Sharpness  post-processing  for HD content (prior could only do SD)
• x.v.Color output
• Dual simultaneous HDMI output
• 12bpc (bits per channel) output via HDMI and DisplayPort
• DisplayPort Audio
• Dolby TrueHD and DTS HD Master Audio

As far as performance goes, the new core features 12 execution units (up from 10), and some more new features that do little more than add checkmarks to a product specs page. They will come clocked at 733 MHz for most models (some X4500HD chips ran as high as 800 MHz), except the 661, which is clocked at 900 Mhz. Naturally, that is the one Intel sent out for testing. Which is unfortunate, because there is no telling whether it will ever become available in great numbers. Curiously, while the “6×1″ model has a much higher GPU clock speed, it loses support for VT-d , V-Pro, and Intel TXT technologies (not that those matter to the vast majority of users, anyway).

It probably goes without saying that the new GPU is pretty much useless in today’s top games. In fact, Intel no longer even pretends that people will ever want to use it to play games like Modern Warfare 2 or Crysis. Instead, they are going after the millions of people who play games like World of Warcraft, Lego Indiana Jones, and Sims 3. And of course there are the billions of people in China, Korea, and Taiwan who play MMO games (the MMO market is ridiculously huge in Asia compared to North America, where you only really have “WoW” and “This Month’s Failed Attempt” to choose from).

On to the desktop CPU, of which there are five new models being launched today (well six, if you count the 661). They are all dual-core chips, a first for Nehalem. Interestingly, the 2-die solution doesn’t simply add a GPU to a CPU that is identical to Lynnfield. Instead, the 45nm die houses the GPU as well as the PCI-E controller and memory controller that was normally found with the CPU cores (see the Lynnfield diagram for reference). The 32nm die is where the cache and two cores reside. This will certainly put a damper on memory latency, and one wonders whether Intel would be able to make a 32nm die today that contains 2 cores, a PCI-E controller, and a memory controller. Also, can you really call this a “32nm CPU” when half of it resides on a 45nm die?

All chips support HyperThreading, so can process up to four simultaneous threads. Turbo Boost is supported by all the Core i5′s, with multiplier increases of 1x or 2x, depending on whatever. The Core i3′s will only ever run at their base clock speed.

These new Clarkdales also see a shrink in L3 cache size – down from 8MB on Lynnfield and Bloomfield, to 4MB. Of course, there are only 2 cores now, so that makes perfect sense.

TDP is rated at 73W across the board, except the 661, which draws up to 87W thanks to its overclocked GPU.

The only new feature being added is a new instruction set which enhances AES encryption performance. If these instructions are used, AES performance can be increased by a significant amount (we’ll be looking at this in our performance results).

Here’s a quick overview of how the entire Nehalem Desktop family looks:

Model	i3 530	i3 540	i5 650	i5 660	i5 670	i5 750	i7 860	i7 870	i7 920	i7 950	i7 975
Family	Clarkdale					Lynnfield			Bloomfield		Bloomfield XE
Cores Threads	2/4					4/4	4/8
Base Clock (GHz)	2.93	3.06	3.2	3.33	3.46	2.66	2.80	2.93	2.66	3.06	3.33
Max Turbo	N/A		3.46	3.60	3.73	3.20	3.46	3.60	2.80	3.33	3.60
Turbo Power			1/2			1/1/4/4	1/1/4/5	2/2/4/5	1/1/1/2	1/1/1/2	1/1/1/2
GPU Clock (GHz)	733 (i5 661 = 900)
Uncore (GHz)						2.13	2.40		2.13		2.66
Official Memory Support	2 x DDR3-1333								3 x DDR3-1066
DMI/QPI Speed	DMI 2.5 GT/s								QPI 4.8 GT/s		QPI 6.4 GT/s
Socket Type	Socket H (LGA 1156)								Socket B (LGA 1366)
L2 Cache	256 KB per core
L3 Cache	4 MB Shared					8 MB Shared
Thermal Design Power	73 W (i5 661 = 87 W)					95 W			130 W
Die Size	81 mm² CPU, 114 mm² GPU					296 mm²			263 mm²
Transistors	383M CPU, 177M GPU					774 Million			731 Million
Hyper-Threading	Yes					No	Yes
VT-d	No		Yes (Except 661)			No	Yes
Price (Jan 2010)	$125	$145	$195	$208 (661=$210)	$300	$195	$280	$540	$289	$570	$970

A New Chipset

To be able to use the integrated GPU, you will need to use a new chipset from Intel. The entry-level version, which Intel sent for this launch, is the H55. If you wanted, you could use Clarkdale on a P55 motherboard; your Core i5 or Core i3 will work just fine, but the GPU will be idle. I tested the 661 on P55 boards from Intel and Asus, and they worked for the most part. No BIOS versions will support proper timings though, so I could not change multiplier settings at all, and memory timings could not be changed either.

The H57 chipset is what matches up to the P55 better – there are still 16 lanes of PCI-E going directly to the CPU, there is still a 2.5 GT/s DMI interface to the Southbridge (or “PCH”, if you prefer) which still has 8 more lanes of PCI-E, 6 SATA 3.0 ports, 14 USB 2.0 ports, etc. The same RAID controller is supported by H57, but the H55 lacks it completely. It is also worth noting that none of the new chipsets support more than a single 1×16 PCI-E configuration, so Crossfire will have to be done using some PCI-E lanes from the PCH. You’ll need a P55 to get proper 2×8 PCI-E support.

What the H57 does add to the P55, is support for “Remote PC Assist” technology, “IdentityProtect” technology (Sentry Peak), and Quiet System Technology (QST, which allows cooling to be controlled via software). There is also a Q57, which is identical to P57 but adds support for Active Management Technology 6.0 and TDT Anti-Theft technology. Yeah, that.