You are here

Sandy Bridge, Part 1: Intel GMA HD 3000/2000 Graphics

Sandy Bridge Pt.1: Intel GMA HD 3000/2000 Graphics [Updated: 05 January]

January 2, 2011 by Lawrence Lee

POSTSCRIPT 2: GMA HD 2000 Test Results

05 January 2011 - see page 7

In the Fall of 2008, Intel launched the Nehalem processor architecture, the first major change to their chip designs in a couple of years. Both the CPUs and motherboards for the new LGA1366 socket were (and still are) relatively expensive, so only power users and enthusiasts were willing to plunk down the cash for the new hardware. A year later, they released more cost-effective processors on a new socket, LGA1156. Like most budget CPUs, they were cut down, this time with lower clock speeds, a slower interface, and a dual channel memory controller instead of triple. They also split up some of the functions of the northbridge chip (eliminating it), putting some on the processor and bundling the rest with the southbridge in a new chip called the PCH. This made the platform more energy efficient, enabling broad appeal for both home and enterprise markets.

In Q1 of 2009, LGA1156 finally got dual core processors manufactured using the new 32nm fabrication process and an integrated graphics chip (albeit a 45nm one) placed on the same package to further improve efficiency. Dubbed Intel GMA HD, this GPU turned out to be just about perfect for decoding high definition content. Its 3D performance was significantly faster than its predecessors, although it still didn't have enough mustard to take on AMD's best IGP.

The next salvo from Intel takes place today, the release of yet another processor microarchitecture, dubbed Sandy Bridge. As you can imagine, a large scale tech product launch is both nerve-wracking and arduous for the company involved, their partners, the retailers, and the press covering it. While most of our readers are end-users who reap personal benefit from new products, they too are equally afflicted. As the NDAs lift, the web explodes with information from dozens of publications, each with their own take on the latest and greatest. Then the masses get to pore over this mountain of data to form their own opinions, before deciding whether to rush out and buy these new gizmos. (Editor's Note: That Intel set the embargo lift date for Sandy Bridge to Jan 2 deserves long rants by/for tech journalists. Has Intel not heard that Christmas to New Year is a huge holiday break for most of the western world?)

It's an especially daunting task for a launch like this one, as there are new sockets, chipsets, and motherboards as well as the CPU. Rather than throw everything into a single article and overwhelming both you and us, we're going to try to break down Sandy Bridge into its constituent parts and discuss the merits and drawbacks of each over the next week. Obviously there will be some overlap as you can't really discuss one without the other as everything is brand new. Today's topic is Intel's new high-end integrated graphics chip (if you think there can be such a thing), the GMA HD 3000 in the new Sandy Bridge CPUs.

The Chips

A Sandy Bridge processor broken down.

The first thing you need to know about desktop Sandy Bridge chips is they require a new socket, LGA1155 which will eventually replace LGA1156 — even though the latter is barely a year old. Similarly, LGA1366 will be superseded later this year by LGA2011. Assuming LGA775 is still clinging on at that time, Intel will have a ridiculous total of five desktop sockets. So what is different, besides the loss of one pin? Well, for starters the CPU shares its die with an integrated GPU, allowing it to share the L3 cache and providing quick to the integrated memory controller. It all fitted on the same die, even with quad cores, by manufacturing both the GPU and CPU using a 32nm fabrication process. As usual, many minor technical optimizations and improvements have been made, with the end result being an estimated 20~30% performance jump compared to the previous generation.

2nd generation Core i3/i5/i7 lineup.

The new CPUs come in three flavors, following similar conventions with regard to nomenclature. The i7's are fully equipped quad cores with all the amenities, while the i5's lack Hyper-Threading and have 2MB less L3 cache. The i3's are dual core processors with only 3MB of cache and lack TurboBoost, advanced virtualization features, and the AES-NI decryption/encryption instruction set. Power consumption has been kept down, with the quad and dual cores having TDPs of 95W and 65W respectively. Low power 35W/45W/65W models are also on the menu, but we have no information on pricing or availability at launch. The latest missive from Intel tells us that all desktop/mobile SNB SKUs will be launched at CES, and quad cores will be available in the same week. Dual core parts – both desktop and mobile – will ramp in availability 4-6 weeks later. This goes for ULV mobile parts too.

Of particular interest is the "K" variant i5/i7 which trade fewer virtualization features for unlocked multipliers. Without a wide multiplier range, Sandy Bridge CPUs are severely limited in overclocking capability because the base clock frequency cannot be increased much higher than the stock 100 MHz. The "K" models are also the only chips with HD 3000, while the rest get by with the slower HD 2000. Today, we are using i5-2500K for testing, a quad core running at 3.3 GHz, higher clock speed than any quad on LGA1156. With a 1k unit price of US$216 it won't be much more expensive than the Core i5-760 which operates at only 2.8 GHz.

GMA HD 2000 / 3000 Graphics

Like GMA HD, there are two versions of the new Intel HD graphics. They are both clocked the same but; HD 2000 has 6 execution units (similar to shader/stream processors) while HD 3000 has 12. Oddly enough HD 3000 is only available on the "K" series, though the i7's allow for a higher maximum dynamic graphics frequency. The GPU core has a variable speed that works much in the same way as TurboBoost. TurboBoost overclocks the CPU cores depending on how many are in use and what Intel calls Power Sharing overclocks the GPU in a similar fashion. The two chips reside on the same die and the GPU will speed up to the maximum frequency unless the whole package exceeds the designated thermal limit.

New video features.

Along with a bump in clock speed, the new GPUs support the HDMI 1.4 standard with 3D stereoscopic playback, and hardware encoding for H.264 and MPEG-2 video (for applications that support it) and minor image enhancement features. The features-set has also been upgraded to DirectX 10.1 and Shader model 4.1. With all these improvements and changes to the architecture, Intel claims their new GPU is comparable to entry level discrete cards from its competitors.


Chipset comparison.

LGA1155 kicks off with two consumer chipsets, H67 and P67. The only difference is that H67 has support for the integrated graphics and its associated features, while the more expensive P67 has "Performance Tuning" which is code for overclocking. MSRP for Intel's first boards, the DH67BL and DP67BG are US$107 and US$184, respectively, quite a large difference. It seems that purchasing a "K" series processor for a few dollars more is just the beginning. Thrifty enthusiast overclockers are more than justified to feel disappointed.

Motherboard Details: Intel DH67BL

Our first H67 motherboard is the DH67BL courtesy of Intel, a fully featured microATX model with eSATA and USB 3.0. Despite developing a new socket, Intel took pity on end-users by not changing the CPU heatsink mounting hole configuration.

The H67 chipset is very similar to H55 with the only apparent differences being a pair of SATA 6 Gbps ports, a minor increase in PCI Express bandwidth, and the shedding of native PCI support. A bridge chip must be employed to provide PCI slots.

Intel DH67BL specifications.

As a budget board the DH67BL is devoid of fancy cooling and ugly old school brown capacitors litter the PCB. Also notice how portions of the PCB are taped up — this is not a production retail sample.

From another angle.


BIOS: performance screen.

BIOS: processor overrides.

BIOS Summary: Intel DH67BL
Max Non-Turbo Ratio
0 to 50
Memory Frequency
DDR3-1066, DDR3-1333
Memory Voltage
1.20V to 1.80V in 0.05V increments
Memory Timing Control
Graphics Max Multiplier 0 to 255
GPU Voltage
1.00V to 1.52V in 0.02V increments

Given the limitations of the chipset and those imposed by Intel, the DH67BL's BIOS has little to offer an enthusiast. As we used an i5-2500K processor, the maximum non-turbo multiplier was a generous 50 but the chipset doesn't support anything that high. In fact, no ratio we entered had any effect on the clock speed. The only thing that is really tweakable is the onboard graphics. The maximum dynamic graphics frequency is defined as the host clock frequency (100 MHz) x 0.5 x graphics max multiplier. The default multiplier is 22 (for 1100 MHz) while the maximum is an astonishing 255 (12.75 GHz).


Common Test Components:

LGA1155 Platform:

  • Intel Core i5-2500K processor
    - quad core, 3.3 GHz, 32nm, 95W
  • Intel DH67BL
    motherboard - H67 chipset, microATX

LGA1156 Platform:

Testbed device listing.

Measurement and Analysis Tools

Video Test Suite

1080p | 24fps | ~22mbps

x264: Crash is a 1080p x264 clip encoded from the
Blu-ray version of an science fiction film. It features the aftermath
of a helicopter crash. It has an unusually high bitrate for video of this type.

1080p | 24fps | ~3.4mbps

Flash: Avatar is a 1080p Flash trailer of the motion picture "Avatar" from YouTube.

1080p | 24fps | ~33mbps
Blu-ray: Disturbia is a short section (chapter
4) of the Blu-ray version of Disturbia, the motion picture, played
directly off the Blu-ray disc. It is encoded with H.264/AVC.

1080p | 24fps | ~36mbps
Blu-ray: Becoming Jane is a short section
(chapter 7) of the Blu-ray version of Becoming Jane, the motion
picture, played directly off the Blu-ray disc. It is encoded with

Estimating DC Power

The following power efficiency figures were obtained for the
Seasonic SS-400ET used in our test system:

Seasonic SS-400ET Test Results
DC Output (W)
AC Input (W)

This data is enough to give us a very good estimate of DC demand in our test
system. We extrapolate the DC power output from the measured AC power input
based on this data. We won't go through the math; it's easy enough to figure
out for yourself if you really want to.

Testing Procedures

If available, the latest motherboard BIOS is installed prior to testing. Certain services/features
like Indexing, Superfetch, System Restore, and Windows Defender are disabled
to prevent them from causing spikes in CPU/HDD usage. We also make note if energy
saving features like Cool'n'Quiet/SpeedStep or S3 suspend-to-RAM do not function

Our main test procedure is designed to determine the overall system power consumption
at various states (measured using a Seasonic Power Angel). To stress CPUs we
use either Prime95 (large FFTs setting) or CPUBurn depending on which produces
higher system power consumption. To stress the IGP, we use FurMark, an OpenGL
benchmarking and stability testing utility. Power consumption during playback
of high definition video is also recorded.

Our video test suite features a variety of high definition video clips. If
the video (and/or audio) skips or freezes, we conclude the board's IGP (in conjunction
with the processor) is inadequate to decompress the clip properly. High CPU
usage is indicative of poor video decoding ability on the part of the integrated
graphics subsystem.

CPU Clock Speeds

Before we start, we should note that our i5-2500K processor sample (or the DH67BL board) was a little off in that the bus speed operated at lower than normal levels. It should be 100 MHz, but we caught it running at 89 MHz when idle, and 97 MHz with all four cores on load with TurboBoost enabled.

CPU-Z screenshot: i5-2500K at idle.

CPU-Z screenshot: i5-2500K at full load.

Power Consumption

As LGA1156 quad cores lack integrated graphics, we pit the i5-2500K and its HD 3000 graphics up against the Core i5-661. It's an 87W dual core processor with a slightly higher clock speed, Hyper-Threading, and the previously top-end GMA HD graphics clocked at 900 MHz. To make things a little more fair we tested the i5-2500K at stock settings and with two cores disabled. A Radeon HD 5450 was thrown into the mix to compare the onboard graphics to a proper entry level discrete card.

As a platform, the i5-2500K paired with the DH67BL was very efficient, boasting a small improvement over the i5-661 in power consumption when idle and an almost 10W advantage when playing x264 video. With two cores disabled, only load power was truly effected, but this probably isn't indicative of how much juice a proper dual core Sandy Bridge will actually use.

At stock settings, adding FurMark to a Prime95 load on the i5-2500K induced an 11W jump, but when two of the cores were disabled, this difference increased to 18W. This is evidence that the "power sharing" GPU feature is actually working. With two cores inactive, the extra thermal headroom appears to allow the GPU to safely clock itself faster. The HD 5450 added about 6W to the total energy usage.

Our CPU/VRM measurements were taken from the
AUX12V connector with the help of a pair of digital multimeters and an in-line
0.01 ohm shunt resistor and represents the approximate draw of the +12V line by the processor including inefficiencies lost to the VRMs. Again we see the i5-2500K's GPU using more power when two of the CPU's cores were disabled, even more than the Radeon HD 5450.

Video Playback

Running on only two cores, the i5-2500K uses slightly more CPU cycles when rendering most videos than the i5-661, though the i5-661 does have Hyper-Threading that may help in this area. Flash playback was much more demanding, taking up 40% of the processor's resources, but this is likely a driver/software issue with the new hardware as GMA HD handles the same content with only 15% CPU usage.

While Adobe updates their Flash Player often, upgrades that affect hardware acceleration are few and far between. We typically use the latest version of Media Player Classic - Home Cinema to play x264 video using DXVA, but that wouldn't work properly either. After trying different players we settled on PowerDVD 10, which by no coincidence, was recommended by Intel.

Strangely, despite the high CPU usage in high definition Flash clips, power consumption during Blu-ray playback was improved compared to the i5-661. And while the HD 5450 helped GPU accelerate video a bit better, its idle energy use kept the overall draw higher.

3D Performance

While 3DMark is a completely synthetic benchmark, it does provide a useful "ballpark" assessment
of 3D performance. We also threw in a standalone benchmark of Lost Planet 2, a first person shooter, running at 1360 x768 with all details set to low.

BIOS: Performance screen.

The claim that Intel's new graphics chip is comparable to an entry level discrete card is supported by our benchmarks. In our tests, the performance jump over GMA HD was more than 100%, and GMA HD 3000 also just managed to slide past the HD 5450. There wasn't much of a performance penalty when two of the i5-2500K's cores were disabled, but the difference was less noticeable with Intel's integrated chip than the Radeon. Presumably the GPU clocked itself higher with two cores, enough to offset the difference.


Intel's HD 3000 graphics processor is a huge leap over its predecessors, packing a level of 3D performance similar to that of an entry level discrete card like the AMD Radeon HD 5450. More impressively, the chip is on the same die as the CPU, occupying a small section of space and sharing cache and a memory controller with four processing cores. This arrangement gives the embedded GPU inherent advantages and allows it to dynamically alter the clock speed to an appropriate level depending on how many CPU cores are working and how much power the combined package is using. Power consumption has been kept in check even though the quad core i5-2500K we used for testing runs at pretty peppy 3.3 GHz (more with TurboBoost).

When it comes to rendering high definition content, HD 3000 doesn't seem to use fewer CPU cycles than GMA HD, though this may change with time as software develops. We got poor results playing Flash video, and some of the software players we tried did not work with GPU acceleration, so obviously there is room for improvement. With software compatibility and drivers, it is usually a matter of waiting for developers to optimize codes to take advantage of the new GPU. HD 3000 does play video with noticeably less power consumption. We were already pretty happy with energy efficiency of Clarkdale, so this is a very pleasant surprise.

While there is much to like about Intel's latest graphics chip, it will not make a dent in the desktop graphics market. For video playback, integrated solutions have been good enough for some time, although it is possible the boost from HD 3000 could make particularly demanding or poorly encoded HD video more easily playable. The 3D performance has been given a huge boost, but it is still not enough to produce a comfortable desktop gaming experience unless you're playing older titles or using an antiquated low resolution monitor. HD 3000 is also only available on the "K" series processors, so a large chunk of them will be purchased by those with intentions of overclocking. The P67 motherboards are better for this, but they require discrete graphics. It is as good as an entry level graphics card, but we aren't a big fans of those either. In our view, a US$40 card like the HD 5450 should only be purchased if the system lacks integrated graphics or has an older integrated GPU that doesn't accelerate high definition video. For gaming, just doubling the budget to a paltry US$80 will net you a significantly faster card that may actually play the latest PC games at modestly high screen resolutions without epileptic stuttering.

The potential of GMA HD 3000 is much higher for notebooks as it can fill it for an entry level card without sacrificing battery life, employing switching technology, and/or presumably, adding to the cost. Many end-users bemoan the relatively low resolutions of modern laptops (1366x768 is fairly standard for 11.6~15.6" screens), but this unfortunate fact makes HD 3000 much more valuable for mobile computing. These days, 1080p monitors for desktop use are ridiculously cheap and abundant, but gaming at lower non-native resolutions on one can be depressing.

To put all this in context, GMA GMA HD 3000 cannot be isolated from Sandy Bridge of which it is an integral part. There is another competitive graphics-integrated CPU looming. AMD's low-power mobile platform (code-named “Brazos”) and the Embedded G-Series platform for embedded systems (code-named “eBrazos”), both based on the first Fusion Accelerated Processing Units (APUs), AMD integrated graphics are scheduled for deployment in Q1 2010. These are not high performance parts, but AMD has consistently outpaced Intel in integrated graphics, so we are looking forward to doing a side-by-side comparison of the budget i3-2100 with GMA HD 2000 and the coming new 18-watt E-Series “Zacate” AMD APU.

Our thanks to Intel for the i5-2500K and DH67BL sample.

* * *

Articles of Related Interest

Zotac IONITX-P-E: First CULV-embedded Motherboard

Gigabyte H55N-USB3: De Facto LGA1156 Mini-ITX Board?

Zotac H55-ITX-C-E: Stacked LGA1156 Mini-ITX Motherboard

AMD Athlon II X3: Affordable Compromise

Athlon II X4 610e & Phenom II X4 910e: 45W & 65W Quad Cores

Intel Core i5-661: A 32nm CPU with
Integrated Graphics

* * *

this article in the SPCR forums.

POSTSCRIPT: 05 January 2011

GMA HD 2000 Test Results

While GMA HD 3000 is certainly an impressive integrated graphics chip, most users who get on the Sandy Bridge train will either be using discrete graphics cards or HD 2000, the GPU included with the vast majority of LGA1155 processors. To test its capabilities, we used an US$120 i3-2100, a 3.1 GHz dual core Hyper-Threaded processor with HD 2000 on the same Intel DH67BL motherboard.

The i3-2100 is clocked 200 MHz lower than the i5-2500K, but as a true dual core, used less power than the i5-2500K with two cores manually disabled.

Measurements taken from the AUX12V connector showed similar energy efficiency. On load, HD 2000 seems to add an additional draw more or less equivalent to that of HD 3000 with the i5-2500K running all four cores

CPU usage during video playback was very low, almost as good as the i5-2500K paired with a Radeon HD 5450. Flash performance continued to suffer, though the i3-2100 fared better than the i5-2500K operating with two cores. It seems that Hyper-Threading does make a big difference in this regard.

During video playback, the i3-2300 used about 1W less than the rest of the field, nothing to write home about.

3DMark pegs the HD 2000's 3D performance about midway between GMA HD 3000, and GMA HD. Stand-alone benchmarks for the games Lost Planet 2, and H.A.W.X. 2 disagree with this assessment, with the HD 2000 scoring closer to Intel's first generation HD graphics.


HD 2000 is far less impressive than HD 3000, providing a more incremental 3D upgrade from GMA HD. It's the kind of unexciting baby-step improvement we're used to seeing from new integrated graphics platforms. The bright side is it brings to the table the same quality of video decoding as HD 3000 but in a slightly more energy efficient package. It has the same difficulty with Flash, but we expect that to get ironed eventually and in the meantime, even dual core Sandy Bridge processors can power through high definition Flash content with brute CPU strength alone.

Our HD 2000 test results reinforces our puzzlement at why HD 3000 was paired only with the "K" series processors. If Intel truly wants to a piece of the low-end discrete graphics market, putting HD 3000 on all their new processors would have been the way to go. It seems a waste that many HD 3000s will sit idle in overclocked systems with P67 motherboards and discrete graphics cards.

* * *

this article in the SPCR forums.


www SPCR