The story of desktop graphics chips from Intel began on February 12, 1998, when i740 chip was announced, the first graphics chip from Intel. Intel's engineering team worked hard on this solution together with the team from Real3D, Lokheed Martin Company division dealing with the graphics solutions for military needs.

The final product developed by the two companies features a lot of interesting characteristics. For instance, the graphics cards based on i740 didn't store any textures in the local graphics memory (now it sounds really weird, don't you think so? Just imagine the today's fast graphics chips addressing slow RAM all the time…). Instead, they used the DiME mode (Direct in Memory Execution) for textures located in the RAM. Nevertheless, it used to be one of the most advanced solutions at that time: the performance of i740 appeared competitive against the rivals' background and the 3D image quality was simply excellent especially compared with S3 Sanage3D, NVIDIA Riva128 and 3dfx Voodoo. No wonder, that the industry turned very positive towards the new chip from Intel and almost every more or less significant graphics card maker announced a product on i740, while the latter appeared another weighty argument in favour of the AGP bus.

But the time passed and after new generations of NVIDIA, 3dfx, Matrox, S3 and ATI chips appeared, the popularity of i740 started going down little by little. It got outdated very rapidly having moved to the mainstream graphics market and then to value one.

The next graphics chip from Intel was i752, improved version of i740 solution with the implemented multitexturing and single-pass tri-linear filtering. However, we have never seen any i752 based graphics cards, as the chip didn't boast high performance at launch already, so the company decided not to produce any i752 chips considering them an unprofitable business.

Here the story of discrete graphics chips from Intel ends and another epoch of integrated chipsets starts. When Intel faced the necessity to develop an integrated solution for value systems, they suddenly remembered about i752 graphics chip. It was adapted for i810 designed to support PC100 SDRAM and Intel Celeron CPUs. The graphics core was integrated into the GMCH (Graphics and Memory Controller Hub) and worked at 100MHz. Besides, there was an enhanced GMCH version aka GMCH-100, where the core acquired 4MB of local SDRAM graphics memory working at 100MHz. Despite the fact that i752 integrated into GMCH "learned" to communicate directly with the memory controller (at 800MB/sec) and not via the AGP bus (at 533MB/sec), the use of local graphics memory improved the performance of i752 core quite noticeably. And when i810E supporting PC133 SDRAM came out, the working frequency of the local graphics memory grew up to 133MHz, having speeded up the graphics core by another bit.

The next reincarnation of the notorious i752 was i815 chipset where the integrated graphics core worked at 133MHz. In terms of graphics i815 didn't differ that much from i810 so we will not dwell on it here.

i810 and i815 appeared extremely popular solutions among all integrated chipsets. All the mainboard manufacturers offered their products based on i810 and i815 chipsets. And even now they are listed among the best Socket370 chipsets in the industry.

Here the glorious story of the integrated graphics from Intel was interrupted for a long period of time. The new Pentium 4 processors came out, new chipsets for them were announced too. VIA and SiS started conquering a new market having introduced the first integrated chipsets for the new platform, and Intel didn't bother to say its word.

And certainly, some other company could have appeared our today's hero, if Intel hadn't been active enough to start struggling for this market niche finally. On May 20 the company announced its new i845G and i845GL integrated chipsets for Pentium 4 and Celeron platforms.

We have already taken a closer look at these chipsets in general in our article called "RDRAM vs. DDR SDRAM. New Rivalry: i850E against i845E and i845G". Today we will pay special attention to a new integrated graphics core from Intel and will compare its performance with the already existing solutions from Intel's competitors.

i845G and i845GL Features

Here is a list of the major features implemented in the Intel i845G/GL integrated graphics core:

3D:

• 200MHz core frequency;
• 1 pixel pipeline capable of laying 4 textures via a single pass (*see below);
• Supports 1x1 textures up to 2048x2048, 12 MIP-levels;
• S3TC/DXTC, FXT1;
• 32bit colour, 32bit textures;
• Supports bilinear, tri-linear and anisotropic filtering;
• Full support of Cube Environment mapping;
• Bump Mapping support: Emboss and Dot3 methods;
• Lines anti-aliasing, FSAA via supersampling;
• Supports different modes of alpha-blending and fog;
• Flexible description format for DirectX6.0 vertexes, indexed vertex arrays, Triangle strips & fans;
• Supports 16bit and 24bit Z-buffer, 16bit and 24bit W-buffer, 8bit Stencil buffer;
• Max supported resolution in 3D equals to 1600x1200x32 for 85Hz.

2D:

• 256bit BLT engine;
• GDI+ functions support;
• Scalability and drawing of semi-transparent objects;
• 32bit transparent cursor support;
• Colour space transformations;
• 8bit, 16bit and 32bit graphics modes.

Video and Overlays:

• Deinterlacing with Bob and Weave methods;
• Capable of synchronizing the display and a TV-Out with the graphics stream source;
• Video source resolution up to 720x480;
• Overlay scalability;
• Brightness, contrast, saturation and hue;
• Overlay source resolution up to 720x576.

We are, first of all, interested in the 3D part implemented in the new Intel graphics core. Let us compare 3D parts of the integrated i845G/GL graphics core and the cores of the competing chipsets: ProSavageDDR P4M266 from VIA Technologies and SiS650 from Silicon Integrated Systems:

A few comments to the table above:

• Texture units of i845G/GL GMCH can operate maximum 4 texture coordinates and hence can process max. 4 textures within a single pass (in case of one-dimensional textures). However, the one-dimensional textures are hardly ever used in the today's games: all the textures you can find there are two-dimensional ones with two texture coordinates each. The maximum number of two-dimensional "standard" textures processed by the texturing units of i845G/i845GL is 2. Therefore, the spec-table above mentioned that i845G/i845GL feature 2 "standard" texturing units per pipeline.
• The datasheets claim that the graphics core of i845G/i845GL GMCH supports Full-Scene Anti-Aliasing via 4x supersampling. Unfortunately, the current drivers do not support this function.
• Despite the fact that the specs for SiS650 from Silicon Integrated Systems do not list the graphics core model, we know for sure that there is a slightly modified SiS315 discrete graphics chip integrated into SiS650 chipset (see our SiS315 Review).
• Two pixel pipelines of the integrated SiS650 graphics core, like SiS315, can merge during texturing this way forming a single pixel pipeline with 2 TMUs.
• The current driver version for the integrated SiS650 graphics core does not see hardware T&L and FSAA support, which are typical of SiS315. It is quite possible that these functions are simply closed by the driver, while it is also possible that they have been cut off physically to reduce the chipset cost.

Some interesting features of the integrated i845G/GL graphics core deserve our particular attention.

Firstly, i845G/GL chipsets are known to use a part of the system memory as graphics memory. However, they do not use a fixed amount of memory every time but change the occupied memory dynamically depending on the graphics core needs. At present the graphics core can count on the maximum of 64MB RAM, however, the upcoming driver versions this number should grow up to 72MB. This RAM organization allows using it in a most efficient way, as no memory is wasted when the graphics core doesn't need much of it and at the same time the sufficient amount of memory is granted to the graphics core every time it really needs it.

The second interesting feature of the i845G/GL graphics core is the Zone Rendering mode. When this mode is enabled, the frame to be rendered is split into rectangular zones. Inside these zones only those polygons are built, which touch the rectangle. The system CPU determines if the polygon belongs to a particular zone and sorts out all polygons for the zone. As a result, the frame is built of rectangles like that as if they were bricks. This approach has nothing to do with the "real" tile architecture of the well-known PowerVR/STM KYRO and KYRO II chips (tile frame buffer and Z-buffer are not implemented as an on-die cache, there is no deferred texturing, etc., see our STM Kyro II Review: Hercules 3D Prophet 4500 64MB Graphics Card). However, it can increase the graphics core performance quite a lot, due to more efficient use of the internal frame buffer caches, Z-buffer, textures and vertexes: when the scene is built within a small marquee, smaller data packs are used and the probability of getting into a cache increases.

i845G/GL chipsets support S3TC/DXTC and even FXT1 texture compression format developed by 3dfx. Automatic compression and the use of preliminarily compressed textures allow reducing the memory bus workload with the minimal quality losses.

Besides a standard analogue VGA-port, i845G/i845GL chipsets feature 2 digital DVO-ports, which can serve to implement TV- and DVI-Outs. DVO-ports are multiplexed via the AGP bus that is why you can use a special expansion card (ADD, AGP Digital Display) installed into the AGP slot to implement TV-Out and DVI-Out on your mainboard. Unfortunately, i845G/i845GL chipset do not support multi-monitor configurations even though they allow sending the signal to several receivers simultaneously.

Moreover, there appeared one more pleasing feature: 32bit colour support. As you remember, the good old i752 could boast only 16bit modes.

These were the highs of the integrated i845G/i845GL graphics core. Now we have to say a few words about the drawbacks.

First, one pixel pipeline with 2 TMUs even working at 200MHz is like a baby-talk against the background of the contemporary solutions. The solution featuring graphics core architecture of the kind is absolutely unable to compete even with the lowest-end discrete graphics chips. We will prove this point later on in the review.

Second, i845G/i845GL do not support tri-linear filtering working simultaneously with multitexturing, sticking to the "best" traditions of the last century. It means that if you decide to enable tri-linear filtering, you will have to disable multitexturing, which will result into the necessity to build a scene in several passes. As you can imaging, this is a deadly situation for integrated graphics with its low memory bandwidth.

On the whole, the integrated graphics of i845G/GL chipsets don't have anything outstanding to boast being just classical integrated chipsets, which graphics cores compete with the system CPU on accessing the memory.

The differences between i845G and i845GL are the following:

• i845GL doesn't support 533MHz bus, unlike i845G.
• i845GL doesn't support AGP slot that is why you will never be able to use an external AGP graphics card with an i845GL based mainboard.

All other features of the two chipsets are absolutely the same.

We managed to test D845GBV mainboard from Intel based on i845G:

…and D845GLAD mainboard on i845GL:

Well, now that've lost a bit of the former optimism, let's pass over to 3D performance of the i845G/i845GL compared with the results shown by VIA ProSavageDDR P4M266 and SiS650.

Testbed

We tested D845GBV and D845GLAD mainboards on i845G and i845GL chipsets respectively, ASUS P4S333-VM mainboard on SiS650 and a sample of VIA P4MA mainboard on VIA ProSavageDDR P4M266.

The boards based on chipsets supporting 533MHz FSB, namely on i845G and SiS650, were tested in the following system:

• Intel Pentium4 2.4GHz CPU with 533MHz FSB;
• 2 x 256MB PC2700 CL2.5 DDR SDRAM by Crucial;
• Fujitsu MPF3153AH HDD.

We used the following software:

• Detonator 29.20 driver for Windows XP for NVIDIA GeForce2 MX400 graphics card;
• Driver version 6.13.01.3091 for the graphics core of i845G/i845GL;
• Driver version 2.05a for the graphics core of SiS650;
• Driver version 13.93.34 for the graphics core of VIA P4M266;
• Windows XP;
• DirectX 8.1;
• 3DMark 2001 SE;
• VillageMark;
• Serious Sam: The Second Encounter.

Chipsets supporting FSB 400MHz, namely i845GL and VIA P4M266, were tested in a system differing from the previous one only by the CPU. We used Intel Pentium 4 2.4GHz with 400MHz bus.

For a better comparison we have also tested i845G based system with an external graphics card based on NVIDIA GeForce2 MX400 chip working at 200MHz and equipped with 32MB 128bit graphics SDRAM working at 166MHz.

Performance

At first, we ran 3DMark2001 benchmarks, as usual. Keeping in mind what kind of graphics solutions we tested, we decided to use only those modes where the level of detail was set to "Low Detail". The results are split into two major groups: the chipsets supporting 533MHz bus (yellow) and those supporting 400MHz bus (blue).

Well, i845G and i845GL failed to perform even comparably to NVIDIA GeForce2 MX400, which we actually had expected to happen: GeForce2 MX400 is twice as fast as any of the testing participants, boasts hardware T&L, not to mention the fact that it doesn't have to share the memory bandwidth with the system CPU. Nevertheless, Intel newcomers managed to defeat all other integrated competitors with a significant advantage. It was the higher working frequency of the graphics core together with the larger integrated caches of the i845G/GL solutions. The drivers optimization is also very important (remember the world's history, SiS315 and Savage4 graphics cards and drivers from SiS and S3…).

It is interesting to watch how the performance drops as the resolution increases. In lower resolutions NVIDIA GeForce2 MX400 was almost twice as fast as i845G/i845GL, while in 1024x768 Intel's chipsets didn't fall behind that greatly. It looks as if Zone Rendering technology really worked as it was supposed to. Unfortunately, the drivers do not allow enabling/disabling Zone Rendering manually, so we didn't manage to find out how great is its positive effect on the performance.

The test in Serious Sam: The Second Encounter were run in two modes: "speed" (16bit color) and "quality" (32bit color):

Again Intel i845G and i845GL appeared hopelessly behind NVIDIA GeForce4 MX400, outpacing chipsets from VIA and SiS.

Large caches and higher clock frequency of the graphics core integrated into i845G/GL show their best in the synthetic fillrate test:

The results of i845G/i845GL in all testing modes appeared closer to theoretical values (200 megatexels per second in case of 1 texture and 400 megatexels per second in case of multitexturing), than by other integrated chipsets. Intel engineers did a great job on the graphics controller optimization, having squeezed the maximum out of the approach towards RAM utilization.

It is especially interesting to see what Zone Rendering technology is worth here. To see how it works, I resorted to VillageMark test, a synthetic benchmark displaying a scene with large overdraw rates:

I wonder why Intel chipsets got beaten by everyone else except VIA P4M266 and the leader appeared SiS650? The answer is very simple. Ft, Zone Rendering technology is not tile architecture and hence, it cannot make the overdraw rate equal to 1. Second, VillageMark uses three textures and tri-linear filtering for each scene object. SiS650 graphics core didn't use tri-linear filtering and built the scene in 2 passes, like NVIDIA GeForce2 MX400. i845G/i845GL honestly used tri-linear filtering and therefore appeared unable to involve multitexturing. The results are shown on the diagram.

The next chapter of our review is devoted to the quality of texture filtering provided by i845G/i845GL and their operation in 16bit colour mode.

3D Image Quality

We tested the quality of texture filtering in Serious Sam: The Second Encounter in two modes: "speed" (bi-linear filtering and 16bit colour) and "quality" (tri-linear filtering and 32bit colour).

i845G/i845GL

Speed		Quality

We don't have any comments about the texture filtering quality. Serious Sam: The Second Encounter in quality mode allows enabling not only tri-linear filtering, but also anisotropic filtering. Unfortunately, i845G/i845GL do not support simultaneously working tri-linear and anisotropic filtering that is why we can see the borders between different MIP-levels, even though the textures are much clearer:

SiS650

Speed		Quality

Like SiS315, the integrated graphics core of SiS650 doesn't perform correctly even bi-linear filtering turning the picture into something horrible (see our SiS315 Graphics Chip Review). Even if we change the mode from "speed" to "quality", no improvement is detected.

VIA P4M266

Speed		Quality

ProSavage8 graphics core integrated into VIA P4M266 failed to perform multitexturing correctly (base texture + lightmap). As a result, the whole thing looks unnaturally bright. When we disable multitexturing, the performance drops down, of course, however, the picture gets more realistic:

Speed		Quality

The Level of Detail on ProSavage8 is higher than it should be, which causes some rippling, glimmering and other "pleasant" effects. Other than that everything is fine :)

NVIDIA GeForce2 MX400

Speed		Quality

As a reference for our comparison, we would like to offer you some screenshots taken on NVIDIA GeForce2 MX400. In "quality" mode we enabled both: tri-linear and anisotropic filtering. As a result, the texture filtering quality by GeForce2 MX400 is indisputably better than by the integrated competitors.

The dithering quality in 16bit modes if a very important parameter for integrated chipsets, because their integrated graphics is pretty slow against the background of the today's graphics solutions, so the use of 16bit modes increases their gaming performance significantly.

We tested the dithering quality in a scene from Serious Sam: The Second Encounter:

i845G/i845GL
SiS650
VIA P4M266
NVIDIA GeForce2 MX400

The dithering quality by SiS650 is undoubtedly the best of all. All the other integrated chipsets, like NVIDIA GeForce2 MX400 managed to cope with the dithering almost equally well, but nevertheless, worse than SiS650.

2D Image Quality

When we tested D845GBV and D845GLAD with ViewSonic P775 and Samsung SyncMaster 900 IFT monitors we noticed a slight blurring effect in 1024x768@85Hz. So, we wouldn't claim that the 2D image quality has got any better, unfortunately, and only the most undemanding users will be happy with what they will get. However, RAMDAC frequency of the i845G/i845GL graphics core makes 350MHz that is why it is quite possible that other mainboards on these chipsets will be showing better results, once the PCB layout is well-done.

Conclusion

Well, it looks as if Intel didn't try to create a Value graphics card killer when they were working on i845G/i845GL solutions: NVIDIA GeForce2 MX400 easily outperforms new Intel chipsets in 3D graphics. No wonder happened with the launch of the new chipsets from Intel. As usual, Intel's integrated solutions are most fit for all sorts of office and value systems, where no high graphics performance is required.

i845G/i845GL evidently borrowed a lot from the previous generation i752 graphics core from Intel. However, due to certain up-to-date enhancements, its performance is improved so greatly that the alternative integrated solutions from VIA and SiS are already unable to compete worthily with new Intel's chipsets. We expect this advantage of Intel's to last for a long time. Very soon there is i845GE chipset to come. Its graphics core will work at 266MHz, so that the gap between Intel and VIA and SiS will grow even bigger.

To speed up the integrated graphics more tangibly, either local graphics memory with high bandwidth or total redesign of the memory controller (like in NVIDIA nForce420 chipsets for Athlon platform) is required. So far, there I not a single mention about solutions for the kind for Pentium 4 platform. And they are very unlikely to appear in the near future: Intel is very happy with the leadership in the integrated graphics market for Pentium 4, so they will hardly wish to let NVIDIA interfere. On the other hand, NVIDIA is not striving for this market so far, they are pretty happy with the leading positions in the graphics chips and integrated graphics for Socket A.

So, the launching of i845G/i845GL from Intel made the company a leader in the field of integrated graphics chipsets for Pentium 4, and offered us new solutions, which are not only very functional and fast (see our article called RDRAM vs. DDR SDRAM. New Rivalry: i850E against i845E and i845G), but also offer the best graphics performance in their group.