Benchmarks: Intel's 32nm Clarkdale

At this week's Consumer Electronics Show (CES) in Las Vegas Intel is introducing five Core i7 mobile processors, eight Core i5 CPUs and four Core i3 chips. Accompanying the new desktop and mobile platforms are seven new chipsets. For the desktop Clarkdale (Core i5 and Core i3) chips, Intel is using the 32nm Westmere manufacturing process, which is a die-shrink of the 45nm Nehalem architecture. The special feature of both the Clarkdale desktop and Arrandale mobile chips is the integration of the graphics processing unit (GPU) into the processor package — although not the same piece of silicon. Dual-core Nehalem-based processors are therefore only available with integrated graphics.

The new integrated GPU, called Intel HD Graphics, supports DirectX 10.1. The new GPUs strengths lie in better video acceleration rather than 3D performance (although it also offers better 3D, the improvement is minimal for more demanding 3D games). AMD's existing 790GX chipset with integrated graphics, announced back in the summer of 2008, offers slightly better 3D performance than Intel's HD Graphics. The 800-series AMD chipsets, planned for April 2010, are likely to outperform Intel's technology by some margin and will exploit DirectX 11. In addition, Intel's HD Graphics does not support DirectCompute or other techniques, such as Nvidia's CUDA or ATI's Stream, designed to harness the GPU's computing power to accelerate 2D and non-graphical applications.

The new Intel platform offers benefits over previous Intel parts and current AMD competition when it comes to application performance versus power consumption. For example, with a standard clock speed of 3.33GHz, an Intel Core i5 661 can match AMD's fastest quad-core Phenom X4 965 chip running at 3.4GHz in most benchmarks. And so it should, because the new Intel chip will cost $196 (£122) in thousand-off prices, just a dollar more than the AMD processor. The Clarkdale chip is particularly powerful when running programs that use the CPU's six new commands for AES encryption. In such circumstances it can offer about 10 times the performance of comparable processors.

Although the AMD Phenom X4 965 can compete with the dual-core Clarkdale in terms of application performance, it's soundly beaten when it comes to power consumption. Intel's 3.33GHz dual-core Core i5 661 consumes only 45W at idle, which is considerably less than the Phenom X4 965's 74W. There are several reasons for that low figure. One is that Intel's Clarkdale chip uses a 32nm manufacturing process, while the AMD processor is still produced in 45nm. The remaining components in the new Intel platform (GPU, chipset) will be produced in 45nm, which still represents a vast improvement over the existing Intel G45 chipset, which uses 130nm and 65nm components.

The Clarkdale processor with its 32nm CPU and 45nm integrated GPU, plus the H57, H55 and Q57 chipsets, together form Intel's new desktop platform.

Chipset technology, power consumption

Because the memory controller and GPU are integrated in the Clarkdale processors, the new chipsets have only two components. By integrating the GPU, Intel has dispensed with the G45's Northbridge chip. Whereas the CPU is manufactured using the 32nm (Westmere) process, Intel is employing the older 45nm technology for the manufacture of the other chipset components. This approach nevertheless represents a step forward from the G45 chipset, whose components are still produced in part by 130nm and 65nm manufacturing.

By integrating the GPU and memory controller into the Clarkdale processor, the new desktop platform consists of just two components.

Intel is aiming the new H57 RAID and H55 non-RAID chipsets at home users, with the Q57 chipset marked out as the enterprise part due to its vPro out-of-band management capabilities.

The H57, H55 and Q57 chipsets expand the 5 Series family. The P55, which is already available, does not support Clarkdale's integrated graphics.

The reduction of the component count saves costs, while the smaller manufacturing process results in chips that draw less power. In our tests, the complete system, comprising Intel's DH55TC Tom Cove motherboard, 3.33GHz Core i5 661 with two 2GB DDR3 memory modules, a DVD drive and an OCZ Vertex Turbo SSD, consumed only 44.5W at idle. The previous-generation system with a 3.13GHz Core 2 Duo 8500 and G45 chipset draws almost 14W more. If you add a DirectX 11 graphics card, such as an ATI HD 5750, power consumption increases by 22W.

Windows 7 64-bit, 1280x1024. Watts; shorter bars are better

Windows 7 64-bit, 1280x1024. Watts; shorter bars are better

Although the AMD Phenom X4 965 system is uncompetitive at idle with the Intel-based competition, it falls even further behind under full processing load. Flat out, the AMD system, which is based on an MSI DKA790GX Platinum motherboard, consumes nearly three times as much power as its Intel counterparts.

Intel HD Graphics: technology, 3D performance, GPGPU computing
Intel's 45nm GPU offers significantly better 3D performance than its predecessor in the G45 chipset. At least that's true for the most powerful Clarkdale processor, the Core i5 661, which Intel made available for these tests. It's worth noting that the GPU in the Core i5 661 is clocked at 900MHz, whereas other versions of Intel HD Graphics run at 733MHz.

The integrated HD Graphics chip is not suitable for demanding 3D games. In the end, competition for the 3D performance crown comes down to a straight fight between ATI and Nvidia, whose GPUs use up to three billion transistors, with Intel left on the sidelines.

The 3D tests using Futuremark benchmarks make sobering reading. Intel's HD Graphics chip is not even as fast as the AMD 790GX chipset, which is based on a Radeon HD 3300 and has been available since mid-2008. The 3DMark Vantage benchmark shows that the ATI Radeon 5750 and 5870 chips are up to 37 times faster than Intel's latest GPU when running modern games.

That 3D performance is likely to be even worse with the other Clarkdale graphics chips, which are clocked at 733MHz rather than the Core i5 661's 900MHz.

Windows 7 64-bit, 1280x1024. Longer bars are better

Windows 7 64-bit, 1280x1024. Longer bars are better. (No result obtained for the Core 2 Duo 8500 + G45.)

Of course, graphics chips do not only run 3D and video, but can also provide enormous computing power to traditional applications in general-purpose GPU computing, or GPGPU. In addition to the interfaces offered by graphics specialists, such as CUDA from Nvidia and Stream from ATI, there is Windows 7's DirectCompute, while Mac OS X 10.6 Snow Leopard has OpenCL.

But the number of applications that can tap into these interfaces remains small. One of the pieces of software that does just that is CyberLink's MediaShow Espresso, a video compression tool that supports ATI and Nvidia GPUs for hardware-accelerated video compression under the MPEG-4 standard.

The results from MediaShow Espresso show two things: the use of the GPU confers a huge advantage, and without it the throughput is heavily influenced by the capabilities of the CPU itself. Users who want to compress video into MPEG-4 format regularly should take a closer look at this software. From taking 79 seconds on the Core i5 661 system with Intel HD Graphics, substituting the Radeon HD 5750 brings the time down to 34 seconds. That's only a second slower than the Phenom II X4 with the same GPU, whereas without external GPUs the AMD part is comfortably faster at 64 seconds over Intel's 79. Even if Intel releases an updated driver that supports DirectCompute, it remains to be seen whether an integrated GPU would make much difference applications such as MediaShow Espresso. For example, it's worth noting that the CUDA-supporting Nvidia-ION platform provides no acceleration for the CyberLink tool. It remains true that when doing work like this, money spent on an appropriate graphics card is frequently a sound investment.

Windows 7 64-bit, 1280x1024. Seconds: shorter bars are better

Intel would rather its graphics chip were seen as powerful tools for video playback — and given its achievements in compression and 3D, that view is understandable. Intel's chip does not do badly with video. The graphic below shows the main differences compared to Intel's previous GPU. In particular, the possibility of simultaneously decoding two video streams is likely to appeal to Blu-ray fans. With its support for Dolby TrueHD and DTS HD Master-Audio, the Intel product has no obvious weaknesses in audio.

Flash Player 10.1 must be installed if you want hardware-accelerated playback of YouTube videos, a requirement that also applies to the ATI and Nvidia competition.

Processor prices, technology, models, performance
There are six new dual-core processors for the desktop, with clock speeds of between 2.93GHz and 3.46GHz. All the models offer hyperthreading and can thus execute four simultaneous threads. Cache size is 4MB and the processors can directly address DDR3 memory at a maximum rate of 1,333MHz in dual-channel mode.

As well as supporting hyperthreading, the four Core i5 models also offer Turbo Boost, familiar from the Nehalem processor. In this mode, the chips' built-in logic ensures that the processor can overclock a single core under load to a specified maximum frequency provided the whole die does not pass preset heat limits. So with single-threaded applications that do not place a burden on both cores, the chip can increase its clock rate. The following table shows the maximum official Turbo Boost frequencies for the Core i5 models. The Core i3 models (540 and 530) do not support Turbo Boost.

Prices start at $113 for the non-Turbo Boost chips. With automatic overclocking, the processors cost between $176 and $284.

In the Clarkdale (desktop) and Arrandale (mobile) processors, Intel has extended the instruction set with six commands that provide benefits for AES encryption. So long as a piece of software can exploit the new commands, Clarkdale chips should see a theoretical maximum of a  10-fold improvement in performance against comparable processors. Encryption is always just one component in more complex tasks so the overall benefit will be more modest; the PCMark Vantage communications benchmark demonstrates this with roughly a two-fold improvement.

The PC Mark Vantage application test as a whole demonstrates the general level of  productivity benefits from the new Intel technology. These tests also show that systems with integrated graphics are at a disadvantage compared to those with a powerful discrete graphics card. Although the Core 2 Duo 8500 system with the G45 chipset failed to run the full test suite, performance was significantly higher in component tests with the ATI Radeon 5750.

Windows 7 64-bit, 1280x1024. Longer bars are better. (No result obtained for the Core 2 Duo 8500 + G45.)

Windows 7 64-bit, 1280x1024. Longer bars are better. (AES acceleration for the Core i5.)

Windows 7 64-bit, 1280x1024. Longer bars are better.

Windows 7 64-bit, 1280x1024. Longer bars are better.

Futuremark provides more information about PCMark Vantage in a white paper (PDF). The benchmark was conducted with the default settings and the 64-bit version. The values shown here represent the average results from three test runs.

Conclusion
With the arrival of the Clarkdale desktop and Arrandale mobile processors, Intel can now offer Nehalem technology for all applications, from the server to the notebook, and in dual- and quad-core packages. But the relatively high prices for the chips with Turbo Boost overclocking technology should ensure the continued popularity of the frontside bus-based Core 2 Duo and Core 2 Quad processors.

These models are available to OEMs from $113 (£70), while existing Nehalem quad-core chips cost at least $196 (£122). The new dual-core models with Turbo Boost are priced $176 (£109). For that figure you could buy a real quad-core processor from AMD that matches the Nehalem-based dual-cores for performance.

But those relatively highly-priced Clarkdale and Arrandale chips do include some benefits. For a start, they include a GPU that provides adequate performance for most users, even though it may not be suitable for demanding 3D games. On top of that, you get hardware acceleration for Blu-ray videos and good-quality audio.

The Intel chips' graphics performance also has potential benefits for non-graphical applications, which are not realised due to lack of drivers. General-purpose GPU computing, or GPGPU, enables applications to exploit the computing power of graphics chips for mainstream tasks. In addition to the interfaces offered by graphics companies, such as Nvidia's CUDA and ATI's Stream, there is Windows 7's DirectCompute, while Mac OS X 10.6 Snow Leopard has OpenCL.

The impressive benchmarks using CyberLink's MediaShow Espresso underline the capabilities of this technology, and Intel's lack of support for such interfaces is regrettable.

If you're not satisfied with the graphics performance of Intel's new processors, you could upgrade with a card such as the ATI Radeon HD 5750, which is available for about £100. A card of that power not only improves 3D performance but also that of 2D applications.

Setting aside the graphics issues, the strikingly positive aspect of the Clarkdale processors is their extreme energy efficiency. Total system power consumption of less than 45W (at idle) with a 3.33GHz dual-core processor is impressive.

Translation by Toby Wolpe. Additional material by Rupert Goodwins