Before discussing advantages and disadvantages of the graphics core integrated by Intel into the 845G chipset let's make clear the requirements we can set for integrated graphics.

It is obvious that the basic function of integrated solutions is to cope with standard (2D) graphics in the GDI Windows implementation, under popular office applications. Moreover, economical PCs based on mainboards bundled with integrated graphics systems should be equipped with respective monitors, i.e. 15" or inexpensive 17" models. With this taken into account it makes sense to provide passable (without considerable performance drop) support of 1024x768 at 32bit color depth, adequate signal sharpness and standard (complying with the VESA standard) vertical-repetition frequency - 85 Hz. If you want more today (before price cuts for graphics and professional monitors of 17" and 19"), go with external graphics solutions based on popular chips from Matrox, ATI, NVIDIA and others, successful models of which even of the lower price range (about $50) are usually able to work correctly in 2D at 1280x1024@85 Hz and higher.

Besides, in the 3D graphics field you shouldn't expect from integrated solutions any extraordinary performance level. Business graphics, rendering and games after working hours will be satisfied with the API DirectX 7 and OpenGL 1.2 and a performance level different from 0. Today such performance is typical of aging solutions of the TNT2 level and modern GeForce2 MX200 based models.

As to compatibility and flawless functioning, the requirements become stricter: integrated video can't be disabled, and problems of compatibility with some applications can just make you buy an external graphics solution.

In short, the graphics core integrated into the Intel 845G meets all these requirements. This can be the end of the article :-).

However, we have some more data for those who are interested in details not very important for integrated solutions and in a comparison analyses of the performance.

Contestants

First of all, take a look at the test stand and software used in the tests.

• Processors
• Intel Pentium 4 2.2 GHz, Socket 478
• AMD Athlon XP 2100+ (1733 MHz), Socket 462
• Mainboard:
• DFI NB76-EA (BIOS of 17/05/2002) on i845G
• Jetway P4MFA (BIOS A03) on VIA P4M266
• MSI MS6533 (BIOS 1.1) on SiS 650
• Gigabyte 7VKML (BIOS F3F) on VIA KM266
• Leadtek K7N420D (BIOS of 12/03/2002) on NVIDIA nForce 420-D
• Abit NV7-133R (BIOS 84) on NVIDIA nForce 415-D
• Memory: 2x256 MBytes PC2700(DDR333) DDR SDRAM DIMM Kingmax, CL 2
• Video cards:
• SUMA GeForce2 MX200 (32 MBytes)
• Leadtek GeForce2 MX MAX (MX400, 64 MBytes)
• ASUS 8200T5 GeForce3 Ti500 (64 MBytes DDR)
• Hard drive: IBM IC35L040AVER07-0, 7200 rpm

• Windows 2000 Professional SP2
• DirectX 8.1
• Intel Inf 4.00.1011
• Intel Application Accelerator 2.2
• Intel Extreme Graphics Driver 6.13.01.3086
• NVIDIA nForce UDP 1.0
• SiS AGP 1.09f
• SiS Video Driver 2.05a
• VIA 4-in-1 4.38
• VIA ProSavageDDR Driver 13.93.24 w/Util
• NVIDIA Detonator 28.32

• MadOnion 3DMark 2001 SE
• idSoftware Quake III Arena v1.30
• Gray Matter Studios & Nerve Software Return To Castle Wolfenstein v1.1
• SPECviewperf 6.1.2
• VirtualDub 1.4.9 + DivX codec 5.0 Pro
• BAPCo & MadOnion SYSmark 2002 Office Productivity
• BAPCo & MadOnion SYSmark 2002 Internet Content Creation
• WinAce 2.11

Here is a summary table of the mainboards used today:

Now let's turn to the configurations used in the tests of performance of the graphics chips:

As you can see, the maximum memory size allocated for the graphics core of the 845G is not enough. I don't understand why it's not allowed for this chipset to use up to 32 MBytes of memory for unforeseen 2D needs. The BIOS sets, in fact, only the initial memory size for needs of the main frame buffer and additional buffers. When 3D applications are run the memory allocated for the graphics core can expand dynamically up to 48 MBytes. By the way, textures are not copied into the local buffer of the card, like in case of an external graphics solution, - they are sent by the 845G core directly from the memory of the current application.

Characteristics and specification

Now let's compare the specs with a modern budget solution because comparison with old external models makes sense only for performance estimation, and we must understand clearly what features we lose when choosing a more advantageous and undemanding integrated graphics core:

Card	Intel 845G	GeForce 4 MX 420
Chip, revision, driver version
Chip	845G	NV17
Revision		A3
Driver version	6.13.01.3086	29.40
Main parameters
Pipelines	1 (?)	2
Texture units per pipeline	2 (?)	2
Textures at a pass	4	2
Core frequency, MHz	200, 266 (?)	250
Fillrate (million pixels)	200, 266 (?)	500
Fillrate (million texels)	400, 533 (?)	1000
RAMDAC, MHz	350	350*2
Local memory parameters
Memory frequency, MHz	100, 133,166 (UMA)	166
Memory bus, bits	64 (DDR)	128 SDR
Memory size, MB	8	64
Memory speed, ns	5-8	5
OpenGL version	1.3	1.3
DirectX version	8.1	8.1
GDI+ acceleration	Yes	Yes
Pixel pipeline
Pixel shaders	No	No
Texture stages	4	2
Combination stages	4	8
Multisampling	No	2,3,4 samples
Bump mapping	DOT3	EMBM, DOT3
Clipping surfaces	0	0
Vertex pipeline
Vertex shaders	No	1.1
Vertex streams	1	16
Vertex shader constants	No	96
Matrices for blending (max.)	No	4
Indexed blending	No	No
Light sources	No (no T&L)	8
N-Patches	No	No
RT-Patches	No	No
Primitives	65536	65536
Vertices	65536	65536
Other parameters
Pure Device	Yes	Yes
Sprite size up to	256	64
3D textures	No	No
Environment mapping	Yes (with anisotropy)	Yes (with anisotropy)
Anisotropic filtering	Yes	Yes
Anisotropy degree up to	2 bi-/trilinear samplings	2 bi-/trilinear samplings
Fog	FOGVERTEX FOGRANGE FOGTABLE	FOGVERTEX FOGRANGE FOGTABLE
Texture compression formats	DXTC1..5	DXTC1..5

Well, everything necessary is provided, except a hardware T&L. Taking into account a successful software emulation of TCL and vertex shaders, especially on Pentium4 processors equipped with the SSE2 we must admit that this solution is justified and even fashionable to some degree (remember the SiS Xabre). Dual-monitor support and full-screen anti-aliasing, in the light of integrated solutions, shouldn't be even remembered of. On the other hand, the EMBM is not supported.

Well, we've got a TDMS transmitter for DVI and a TV-Out interface which, if necessary, are multiplexed to outputs of the AGP but only if a special daughter interface card is inserted. But there is only one CRTC in the graphics core, which means that it can display only one image. It's pleasant to see that the core integrated into the 845G supports GDI+, line anti-aliasing, deinterliving and video overlays, color conversion, and all popular texture compression methods. Hardware support of the fond edge anti-aliasing is provided. The core ia able to execute 2D and 3D operations in parallel and independently, which accelerates rendering for non-fullscreen 3D applications (as a rule, not games).

Here is the version and a list of extensions of the OpenGL ICD:

Vendor: Intel
Renderer: Intel Brookdale-G
Version: 1.3.0 - Build 4.13.01.3086

Extensions:
GL_ARB_multitexture
GL_ARB_texture_border_clamp
GL_ARB_texture_compression
GL_ARB_texture_cube_map
GL_ARB_texture_env_add
GL_ARB_texture_env_combine
GL_ARB_texture_env_dot3
GL_ARB_texture_env_crossbar
GL_ARB_transpose_matrix
GL_EXT_abgr
GL_EXT_bgra
GL_EXT_blend_color
GL_EXT_blend_func_separate
GL_EXT_blend_minmax
GL_EXT_blend_subtract
GL_EXT_clip_volume_hint
GL_EXT_compiled_vertex_array
GL_EXT_cull_vertex
GL_EXT_fog_coord
GL_EXT_packed_pixels
GL_EXT_packed_pixels_12
GL_EXT_rescale_normal
GL_EXT_secondary_color
GL_EXT_separate_specular_color
GL_EXT_stencil_wrap
GL_EXT_texture_compression_s3tc
GL_EXT_texture_env_add
GL_EXT_texture_env_combine
GL_EXT_texture_filter_anisotropic
GL_3DFX_texture_compression_FXT1
GL_IBM_texture_mirrored_repeat
GL_NV_texgen_reflection
GL_WIN_swap_hint

The situation looks excellent: the card possesses almost all popular capabilities of the times of TNT2 and even more. I wonder if the OpenGL driver of Intel was developed by them or it was bought (licensed) from one of the famous players of the graphics market? The following extension makes me think so: GL_3DFX_texture_compression_FXT1.

Tiles, Zones, Parts, Flaps, Bricks...

And now the most interesting thing - to improve effectiveness of operation with memory (which in case of UMA architectures is very sensitive to the "culture" and frequency of requests) they use a tile approach for image rendering. Intel calls it Zone Rendering, becasue the concept of tile architecture introduced so unsuccessfully by creators of the PoverVR and Kyro families (yet in the times of Microsoft Talisman) meant nothing else but regular buffer partitioning into equal nonoverlapping zones with independent successive rendering in each. At that time Tile meant arbitrary zones with possible overlapping. Probably, to avoid missmash in the cocepts or not to associate its graphics core with not very successful products (Kyro), Intel decided to name it "Zone Rendering". As a result, most analysts and consumers do not have any idea about this peculiarity of the architecture of the graphics core integrated into the 845G.

So, what do we get with the tile architecture (let's call it this way as we got used to it) in case of an integrated graphics core with shared memory? A lot of advantages:

1. Considerable ecomony of the memory bandwidth (only necessary texture values are accessed)
2. Significant optimization of the memory access character - minimum of read-write reswitching
3. Low requirements for a graphics core clock speed

The main problem of tile architectures is transfer and binning of triangles. But in case of an integrated solution which has the whole system memory at its disposal, including geometry, this problem can be solved successfully, especially with a powerful central processor. Besides, the bottleneck in the form of an AGP bus is lacking this time - the graphics core in the 845G addresses the system memory controller directly, without any virtual AGP bus and bridge. It reduces access delays much. In the extreme case the graphics core can use the whole system memory bandwidth.

For further optimization of the memory access the graphics core of the 845G has capacious separate caches for textures and geometry, as well as a special index cache for quicker binning of polygons to tiles. Later, having compared results of the 845G with classical integrated graphics cores, we will see how these declared optimizations tell upon a real core performance.

The documentation on the 845G provides no official information on a configuration of the chip's shading pipelines, but we must take into account that in case of a tile architecture this issue can be quite complicated. For example, in some conditions the chip can shade N visible pixels per clock, and in other conditions - M pixels, and this figure can vary. On the other hand, the number of texture values sampled per clock is usually fixed even in case of tile architectures, and we can catch on this parameter. According to the test results, (if we assume that the graphics core of the 845G works at a frequency synchronous with the memory - 266 MHz) the number of texture units are two and the number of shaded pixels per clock is one. But remember that this pixel is always visible, i.e. when comparing with classical architectures it's necessary to multiply this figure by an overdraw factor.

3D performance, synthetic tests

Fillrate:

Here and further we will give results of the GeForce2 MX400 for comparison. In the one-texture mode the graphics core integrated into the Intel 845G is edged by the MX400 in 800x600x16bpp and beats it in 1024x768x32bpp, despite a smaller bandwidth! The fact that it's unnecessary to transfer depth values (Z) through the bus has an effect. In case of multitexturing the 845G is thriving - above all, it is able to apply 4 textures at a pass, which is much better than two. Note that even with the tile architecture all pixels are rendered, that is why in real applications we can get additional advantage in the effective fillrate. But this assumption will be checked later.

Now the geometry performance:

A powerful central processor can help much in such tasks. It falls behind the hardware T&L by a great margin in case of a small number of light sources - because it's necessary to bin a great number of polygons to tiles and this mustn't depend on the number of light sources. However, the high-frequency processor calculates lighting much more effectively as compared with the geometrical unit of the GeForce2 MX, which allows the 845G to go on a par with the latter in case of the maximum number of sources. But remember that in real applications the processor will have some more tasks to deal with.

Point Sprite Speed:

The MX400 is above all. It seems that the graphics core of the 845G doesn't have any special hardware means for fast sprite output (note that sprites themselves are not the best form of optimization of output of such tile architecture). As a result, each sprite turns into two normal triangles. After that all triangles must be binned which takes the most part of time of the test.

Vertex Shader Speed:

None of the competitors support vertex shaders on the hardware level, they are emulated by the central processor. But in this case the MX400 doesn't require polygon binning from the CPU, that is why it takes the lead.

DOT3 bump mapping:

In the 16-bit color the MX400 fulfills the task much faster but in the 32-bit color the tile architecture of the 845G catches up with it as it is less demanding as far as the memory bandwidth is concerned.

In general, in the synthetic tests the 845G solution from Intel shows a performance level comparable to or just a little lower than the MX400's one. Let's see what we can get in complex (real) tasks. It seems that the high load of the CPU will worsen results of the 845G with the tile architecture.

3D performance, game tests

First of all let's take a look at the 3DMark 2001 SE scores in the software T&L mode (because of the lacking hardware realization of it in the 845G):

The Intel's solution doesn't look so brilliant in the absolute values (it's right on the level of the MX200), but it has quite good scores as compared with other integrated solutions. The results of the SiS 650 are weaker, and those of both VIA's chipsets equipped with the ProSavage are much weaker. The nForce is a leader among the integrated solutions, but remember that solutions based on it are not cheaper than the 845G+MX200 system or even a more efficient one with the MX400. Besides, the nForce can be used only with the Socket A, that is why its comparison with the 845G is just an academic issue.

Here are individual tests for those who want some details:

The positions of the cards are the same except a small advantage of the 845G over the MX200 in the low details mode - a low load of the central processor allows the tile architecture to show its power. Certainly, a more powerful CPU will make the gap separating it from the MX400 smaller. Each test is described deeper in the following reviews: MadOnion 3DMark2001 and NVIDIA GeForce3 and NVIDIA GeForce4 Ti 4400 and 4600.

Now let's turn to the OpenGL, i.e. its usage as API for popular Quake3 engine based games:

From the standpoint of accelerators of the GeForce2 generation the Quake3 engine is well balanced. That is why positions of the cards remain the same as in the overall 3DMark 2001 test and they don't change in case of different color depths. The MX400 is above all, it is followed by the nForce, then comes the integrated Intel 845G, then we have the SiS 650 and at last two ProSavage based outsiders. But playability of the FPS up to 800x600x32bpp is more important here. It makes 3D capabilities of integrated solutions look not so hopeless.

And now a modern test - RtCW:

The same situation, only the 845G gets weaker as compared with the MX200 in the 32-bit color mode. More effects?

The conclusion is that it's possible to play popular games in moderate resolutions even in an office equipped with only integrated video solutions on the 845G.

Professional graphics

For professionals we offer the SPECviewperf results (such rendering tasks may take place even at office, for example, economists may happen to face them):

In some subtests the 845G has some advantage over the MX400. It's becasue of the bottleneck in transferring geometry and attributes (AGP) and the high-quality OpenGL driver. However, you should remember that this test uses quite old methods in image displaying, and in case of less conservative applications that use new OpenGL extensions (non-gaming as well) the GeForce2 may turn out to be weightier.

How integrated graphics works in usual applications

Attention! Remember that the results of the platforms based on the nForce and the second integrated chipset VIA KM266 are obtained with the other processor, and they are given only for comparison of the overall potential of the tested platforms.

First comes the MPEG4 encoding:

With the integrated graphics the 845G falls behind by 3%, and as a result it goes on a par with the SiS 650. The VIA's solution is 1 minute late.

Now the SYSmark which is very close to intensive work of a real user:

The layout is the same, but the differences are much smaller now - both between the chipsets of different companies and between the integrated and non-integrated versions of the 845G (in the Office Productivity test the difference is not greater than the measuring inaccuracy).

Finally we have the WinAce test:

Now the gap between the integrated video and external one on the 845G is about 8%, the situation is very similar to the above, but the SiS is much slower this time, and together with both VIA's chipsets it lags behind the main group by 15–20%.

I must say that in such tasks the 5% loss of performance doesn't give cause for regretting for a considerable sum of money saved on the bulk purchase of office PCs :-).

2D speed and quality

The 2D speed is sufficient. At least, in 1024x768x32bpp@85 Hz we noticed no flaws or defects. The 2D signal quality in this resolution is on the level of an average GeForce3, which is quite enough even for external 2D cards of the low price range (about $50). The maximum acceptable rated resolution is higher, but it makes no practical sense to work with the 845G based cards (at least with those we had in our lab) in the resolution over 1280x1024x32bpp@85 Hz becasue the image will be blurry.

Conclusion

The graphics integrated into the 845G easily outscores its direct competitor - SiS 650 and of course the VIA P4M266.

We have no more candidates able to play the role of a popular integrated solution for the Pentium 4. Let's congratulate Intel with its successful move which will help Pentium 4 based systems penetrate into all niches, even into the office one.

It is interesting (maybe it's just accidental concatenation of circumstances), that the company which has the possibility to create a much more efficient integrated solution for the Pentium 4 (NVIDIA) hasn't got the license yet. And weak SiS got it (however, after the release of the Xabre the situation may change, but taking into account the results of the Xabre 400 tested in our lab - the changes won't be considerable)... For a more complete picture we should wait for an integrated solution on the Xabre and an ATI's chipset with the Radeon graphics core.

At present Intel is the king for the Pentium 4 platform, but nobody doubts in it...

Write a comment below. No registration needed!