Intel's Sonoma Platform: Dual Channel Performance
September 23, 2005 at 08:02:00 AM, by
In this article we take a look at what dual channel memory really does for system and GMA900 performance.
Introduction
With Centrino having been around since the days of 2003, we have seen it go through many changes. As a quick recap, Centrino refers to the collection of the Pentium M processor, Intel chipset, and Intel WiFi card. You can have a Pentium M notebook with Intel chipset and it not be "Centrino," but still have the architectural features we are discussing here. From its lovable infancy with the Banias CPU to the more mature 90nm Dothan, the Centrino platform has grown in a lot of ways. As everyone here should now know, Intel's latest Centrino iterative has added a LOT of features to the platform and also added a few disadvantages. In this article we will be looking at what dual channel REALLY does for Sonoma and how dual channel memory impacts the Intel GMA900 integrated graphics. Before we jump into the meat and potatoes of this review, let's get a basic Centrino history laid out to make sure everyone is on the same page.
We won't be going into many architectural details, but let's just establish a general idea of where Centrino has been. For those who want to get into the nitty gritty of the platform architecture development, take a look at our extensive Dothan vs Turion article.
Centrino History | |||
Platform | Carmel | Carmel | Sonoma |
Processor | Banias | Dothan | Dothan |
Level 2 Cache | 1MB | 2MB | 2MB |
FSB | 400MHz | 400MHz | 533MHz |
Manufacturing Process | 130nm | 90nm | 90nm |
Chipset | Odem (855) | Odem (855) | Alviso (915) |
| RAM | DDR 266/333 | DDR 266/333 | Dual Channel DDR2 533 |
Wireless | 2100/2200BG | 2200BG/2915ABG | 2200BG/2915ABG |
| Graphics | Intel Extreme 2/AGP 4X | Intel Extreme 2/AGP 4X | Intel GMA900/PCI-E |
| I/O Support | PC Card | PC Card | ExpressCard |
| Hard Drive Support | Standard IDE | Standard IDE | SATA |
| Audio | AC' 97 | AC' 97 | Intel High Definition |
As you can see from above, not much changed in the Centrino platform before Sonoma. The Dothan CPU was indeed a godsend, bringing higher performance with lower power consumption, but the platform itself remained largely unchanged. The memory bottleneck was still the biggest limitation of the Centrino platform, along with support for aging technology like the AGP and PCI buses. Sonoma would change all of this, bringing the mobile community the latest technologies that desktops have taken advantage of for quite some time now.
Setup MethodThis time around we have several different test benches, and a number of different tests that were run. All machines were set on the power scheme "Home/Office Desk" with the AC plugged in, ensuring the systems will run at their highest performance. For the battery tests the machines were set to "Portable/Laptop" to activate Intel's Enhanced Speedstep technology, allowing the machines to downclock to preserve battery life. Screen brightness and audio were both set to 50%, and WiFi was turned on. Each test was repeated 3 times to ensure accuracy. Before each test was run, the laptop was rebooted and its hard drive defragmented.
As you will see in the configurations below, we exclusively used IBM/Lenovo Thinkpads. The similar configurations allowed for consistent benchmarking that would skew results as little as possible. You will also note in the configuration that all Thinkpads were run at 1.6GHz, despite the higher clock of the CPU's (1.73GHz for the R52 and 1.86GHz for the T43). This was to create a standard benchmarking platform to allow us to truly extrapolate the results of our testing. Identical CPU clock speeds will let the testing show what benefits the Sonoma platform truly offers. The clock speeds were modified using the program RMClock.
Business Winstone 2004 Batterymark measures a laptop computer's battery life by simulating real-world usage. The program measures the time it takes to drain the battery by running applications such as Microsoft Office XP, Norton AntiVirus, and WinZip through a series of scripted activities to drain the battery in a realistic way.
3DMark 2001 SE PRO build 3.3.0 measures graphics performance by benchmarking the CPU, memory, and graphics through a series of 21 tests, including simulated games, theoretical tests, DX8 feature tests, and image quality tests. Resolution was set to 1024x768 with all default settings.
3DMark 2003 build 3.6.0 measures DX9 performance through a series of 3D game based sound, graphics, and CPU tests. Resolution was set to 1024x768 with all default settings.
PCMark 2005 Advanced build 1.0.1 is the latest update to Futuremark's popular overall system benchmarking program. The 2005 version adds multithreading, DirectX 9, Windows Media Player 10, virus scanning, High Defintion video playback (WMVHD), and a vast number of other tests to its suite. Testing your computer's CPU, RAM, hard drive and graphics card, PCMark05 drives your computer to the max to determine its strengths and weaknesses.
SiSoft Sandra is mainly a diagnostic/informational tool, but also provides useful synthetic benchmarks. Centering on memory and CPU performance, Sanda's benchmarks provide a consistent testing method of a computer's arithmetic and memory performance.
Doom 3 (Version 1.1) using the built-in time-demo, initiated with the console command timedemo demo1 precache'.
Half Life 2 (Source Engine 7) timedemos Canals 08 from Anandtech.com
| Specifications | Lenovo Thinkpad T43 | Lenovo Thinkpad R52 |
| Processor | Intel Pentium M Processor 750 | Intel Pentium M Processor 740 |
| Front Side Bus | 533 MHz | 533 MHz |
| Chipset | Intel 915 PM Express | Intel 915 GM Express |
| Wireless LAN | IBM a/b/g II mini-PCI (802.11a/b/g) | Intel PRO/Wireless 2200BG (802.11a/b/g) Bluetooth Wireless IrDA |
| LCD | 14.1" SXGA+ TFT LCD (1400x1050) | 14.1" XGA TFT LCD (1024x768) |
| Hard Drive | 60GB Hitachi 7K60 7200RPM 8MB Cache | 60GB Hitachi 7K60 7200RPM 8MB Cache |
| Memory | DDR2-533 SDRAM (Varying configurations) CL=4 | DDR2-533 SDRAM (Varying configurations) CL=4 |
| Graphics | ATI Mobility Radeon X300 | Intel Graphics Media Accelerator 900 128MB Dynamically Shared Video RAM |
| Graphics Interface | S-Video/TV-out | S-Video/TV-out |
| Optical Drive | DVD+/-RW MultiBurner | DVD+/-RW MultiBurner |
| Modem | Integrated V.90 Modem | Integrated V.90 Modem |
| Ethernet | 10BASE-T/1000BASE-TX Ethernet | 10BASE-T/1000BASE-TX Ethernet |
| Audio | Realtek AC' 97 | Realtek AC' 97 |
| Audio Interface | Microphone, two stereo speakers, headphone/line-out with SPDIF support | Microphone, two stereo speakers, headphone/line-out with SPDIF support |
| Ports | 2xUSB 2.0 Express Card Parallel Port | 2xUSB 2.0 |
| Weight | 5.2 lbs. with 6-cell battery | 6.1 lbs. with 6-cell battery |
| Size (W x D x H) | 12.2" x 10" x 1.0-1.2" | 12.4' x 10.2' x 1.4'-1.5' |
| Operating System | Windows XP Professional w/SP2 | Windows XP Professional w/SP2 |
| Battery | 4,600 mAh | 4,600 mAh |
Now let's get on with the testing!
Dual Channel
Dual channel memory is not a new innovation: AMD's Socket 939 Athlon 64 is dual channel capable and most of Intel's Pentium 4 chips have been dual channel capable for quite some time. However, aside from a few toaster-oven implementations of the dual channel Pentium 4 in a laptop, the Sonoma platform is the first to bring dual channel memory to a laptop.
Now while most people have certainly heard of dual channel memory, what does it really mean? Dual channel memory refers to doubling the data path width between the memory and memory controller. In Intel's case, they implemented two DDR2 channels capable of running at 533MHz. When a pair of modules are used, the DDR2 data path effectively increases from 64-bits to 128-bits. This doubles the theoretical memory bandwidth from 4.2 GBytes/sec to 8.6 GBytes/sec. While this sounds all well and good, one needs to take into account the bandwidth of the Front Side Bus. A 133MHz FSB that is 64-bits wide and does 4 transfers per clock cycle yields a maximum theoretical bandwidth of 4.2 GBytes/sec.
As you can see from the diagram above, the data from the memory travels to the Northbridge (labeled Alviso) with a total bandwidth of 8.6 GBytes/sec and then hops on the Front Side Bus to CPU at a bandwidth of 4.2 GBytes/sec. This is like building a 4-lane highway into a busy city to alleviate traffic congestion, only to keep the existing 2-lane highway once inside the city. It should help some, but in real world practice chances are we won't see much improvement. That is what we aim to investigate with our dual channel testing.
We used three different memory configurations for our testing purposes, two of which were in dual channel mode. On the T43, 1x512MB DDR2 533 CL4 manufactured by Samsung is the standard complement. We added to that another stick with the same specifications from IBM, also manufactured by Samsung. To measure the performance between 1x512MB and 2x256MB, we needed some extra modules. We ended up getting two high quality Crucial 256MB PC2-4200 DDR2 SO-DIMM modules (CT435137), which of course run in dual channel mode.
Memory Test Configurations | |||
Total Capacity | 512MB | 512MB | 1GB |
Modules | 1x512MB | 2x256MB | 2x512MB |
Latency | CL=4 | CL=4 | CL=4 |
| Dual Channel | No | Yes | Yes |
Brand | IBM | Crucial | IBM |
Manufacturer | Samsung | Micron | Samsung |
Enough chit-chat, let's get down to business!
SiSoft Sandra Memory Bandwidth
As you can see from the test above, the memory bandwidth results are all within a few percent of eachother. Going from 1x512MB to 2x256MB brings the bandwidth from 2961 MB/s to 3007 MB/s, a meager 1.55% increase in bandwidth. Of course that is within our expected margin of error, so it really isn't proper to say there is a definite bandwidth advantage between the two. However our knowledge of how dual channel memory operates tells us that while it is more efficient, we will not see a sizeable increase in actual memory bandwidth. As stated before, the system is already very efficient due to the synchronous memory clock to FSB ratio, with the high latencies being the weakest link.
SiSoft Sandra Cache/Memory Benchmark
This test is a little unique in that it tests the entire memory and CPU cache subsystems. It combines a little bit of everything with floating point and integer tests, as well as varying data block sizes. This test should give us a little more insight into the effect of dual channel on the memory subsystem as a whole.
The score generated by this test is a "combined index," with the final results interpreted into a Megabytes per second data rate. Going from 1x512MB to 2x256MB yielded a more tangible 4.14% increase. The overall score actually went down slightly with the 2x512MB test, but it is within range to either write it off as expected variation or the higher memory density in the 512MB modules.
PCMark05
This is another synthetic benchmark which essentially confirms Sandra's results. The memory test specifically tests read, write, copy, and latency performance with the following block sizes: 16MB, 8MB, 192kB, and 4kB.
Going to 512MB dual channel yields about a 1% increase in memory performance, as measured by PCMark05. The total scores only varied by 0.5% with the different configurations, which isn't even worth putting up a bar graph.
Dual Channel Cont'd
SuperPI
We used this popular benchmarking program to evaluate the performance of the CPU and memory subsystems. For this test we calculated Pi to 1 million digits.
Again, nothing revolutionary in terms of dual channel performance. The dual channel configuration, in 512MB or 1GB, dropped the SuperPI calculation a full second; a 2.1% increase in performance.
3DMark 2001
Due to the age of this once-GPU oriented benchmark, it now often serves as a good CPU/memory benchmark.
Despite being more CPU/memory limited, 3DMark 2001 shows no improvement in dual channel configurations. It seems that memory capacity if anything improves performance, although the performance differences are within our expected margin of error. Regardless, 3DMark 2001 shows no dual channel advantage, possibly due to the relatively low end GPU's.
Adobe Photoshop CS
Adobe's Photoshop CS program is the industry standard in digital image editing. When processing small pictures, filter times are low, and any recent CPU will suffice. However, when processing large images, a powerful CPU is needed to keep filter times low.
To test Photoshop CS performance, we used the DriverHeaven Photoshop V2 benchmark script, which performs various operations on a huge 5610x3740 image. Below are the individual test results in seconds, thus lower is better:
| Photoshop CS | 512MB (Single) | 512MB (Dual) | 1GB (Dual) |
| Texturizer 1 | 23.5 | 21.4 | 13.0 |
| CYMK Color Convert | 6.8 | 6.8 | 6.7 |
| RGB Color Convert | 12.0 | 11.9 | 9.4 |
| Dust & Scratches | 15.8 | 14.3 | 9.5 |
| Watercolour | 71.3 | 62.9 | 61.8 |
| Texturizer 2 | 16.8 | 12.2 | 10.9 |
| Stained Glass | 16.7 | 15.7 | 14.2 |
| Lighting Effects | 12.6 | 12.4 | 12.0 |
| Mosaic Tiles | 34.6 | 32.0 | 32.0 |
| Extrude | 91.8 | 91.8 | 91.4 |
| Smart Blur | 66.4 | 65.7 | 63.9 |
| Underpainting | 64.8 | 62.1 | 65.1 |
| Total | 433.1 | 409.7 | 390.1 |
As you can see from the results, 1GB RAM rules. Despite the obvious landslide in favor of 1GB system RAM, dual channel or not, there is also a very tangible performance increase when going to dual channel with only 512MB RAM. Photoshoppers should shoot for 2x512MB or even 2x1GB memory configurations, as this is one area where Sonoma's dual channel capability really shines.
GMA900
Most of you are familiar with the dirty word that most people call GMA900. Intel's Graphics Media Accelerator 900 is by far an excellent integrated graphics solution, with true competition only coming from ATI's RadeonXpress200M IGP. Of course for those who want acceptable 3D performance, they do not look at the GMA900.
As we have shown numerous times, performance in modern DX9 games is non-existant. Since GMA900 uses a unified memory architecture, i.e. shared memory, its performance is highly dependant on the type and capacity of main system RAM. The Sonoma platform brought higher bandwidth (although not really higher performance) RAM and dual channel support, both of which mean potentially better integrated graphics performance. 533MHz is the fastest supported DDR2 RAM on the Sonoma platform, unfortunately hindered with a higher latency rating of CL=4. A synchronous FSB/MEM clock ratio and dual channel 128-bit memory controller should help speed up shared memory performance, thus increasing 3D performance of the GMA900.
3DMark 2001
Although 3DMark01 is generally used as a CPU-limited test, in the case of GMA900 it is definitely still GPU limited. We see a surprising 12.2% gain in performance when going from 1x512MB to 2x256MB memory configuration. A similar increase of 11.8% is seen with the transition from 1x512MB to 2x512MB.
3DMark 2003
This slightly newer benchmark from Futuremark still gives a lot of modern GPU's a run for their money, so it definitely beats the GMA900 into an unrecognizeable pulp. However what we do see here is an 8.1% performance increase going to 512MB dual channel, and 7.7% running two 512MB sticks versus one. The performance gain is less than that of 3DMark01 because of the more advanced and intensive graphics processing required in 3DMark03; higher memory bandwidth can only improve performance so much if the GPU core is plugging everything up.
3DMark 2005
Before we present the total 3DMark05 score, let's take a look at the CPU subset.
Futuremark's CPU testing consists of vertex shader calculation and floating point calculations. We see a very surprising 15.3% performance increase with 512MB dual channel, and again a slightly less 9.8% increase with 1GB dual channel. These results tell us that increasing the amount of bandwidth available to the GMA900 decreases the performance impact on the CPU. With more data in a given period of time being transferred to the GPU for it to process, there is less "slack" for the CPU to take up in its own 3D performance. The overall negative impact on the CPU is lessened, but keep in mind that the CPU scores do reference CPU limited tests like vertex shader and floating point calculations.
The performance increase brought by 1GB and 512MB dual channel modes is identical, an expected 12.3%. Due to the more complex graphics calculations, 1GB RAM cannot manage to increase performance with such lackluster core performance. While a 12.3% increase from dual channel may sound like a great performance increase, keep in mind that at a score of 255 you are still watching a literal slide show; and it isn't even that pretty.
GMA900 Cont'd
PCMark 2005
Let's first take a look at the graphics subsystem score, as measured by the following tests that PCMark05 runs:
- 2D - Transparent Windows
- 2D - Graphics memory 64 lines
- 2D - Graphics memory 128 lines
- 2D - Video playback
- 3D - Fill rate
- 3D - Polygon throughput
- 3D - Pixel shader
- 3D - Vertex shader
Here we see a 6.1% increase with 512MB dual channel, and 5.6% with 1GB dual channel. The relatively simple tests of PCMark05 yield expected results: moderate improvement yet nothing to write home about.
PCMark05 shows a 1.5% increase in general system performance, which was caused entirely by the graphics subsystem as the rest of our test components were within 1-2% in each memory configuration. The 1GB configuration brings the increase to 2.1%, showing the benefit of higher memory capacity. In the end, our last synthetic benchmark shows what we expect to see in our gaming tests: a moderate increase yet nothing to make GMA900 even remotely acceptable as a 3D solution.
Half Life 2
Pretty much everybody is familiar with this awesome first person shooter. One reason for its popularity is how well the game scales on older hardware. Desktops with ancient cards like the NVIDIA GeForce4MX are able to play under the DirectX7 codepath, rather well in fact. For our testing we chose the bare minimum quality options:
- 640x480
- Low Model Quality
- Low Texture Quality
- Simple Reflections
- Low Shadow Quality
- 0xAA (Anisotropic Filtering)
- Trilinear Filtering
- Low Shader Quality
We ran tests under both the DirectX 8 and DirectX 7 codepaths, to see what (if any) performance can be gained. There is of course no point in using the DX9 codepath. Half Life 2 actually detects the graphics card as having DX9 software support and DX7 hardware support. Let's take a look at the DX8 results first:
Ooohh, chill bumps. We see 0.04 and 0.05 frames per second increases with dual channel memory performance. Let's see if DX7 helps out any.
Well, there IS a whopping 32.4% increase from DX8 to DX7. However we are at a still incredibly unplayable 6.41 frames per second with the 1GB dual channel configuration. Will Doom 3 fare any better?
Doom 3
We used the build-in timedemo feature, again running with bare minimum image quality settings at 640x480 resolution. Doom 3 features several different rendering paths as well. We originally ran tests with every renderer, but only found that two of the options made a difference. Here is an excerpt from TechSpot about the rendering options:
r_renderer'x'. x specifies the rendering path to be used. Doom 3 features several hardware specific rendering paths, although it initially will detect the most appropriate path to use in the game. As such you shouldn't need to alter this setting. For informational purposes though I'll include them anyway - arb, arb2, nv10 (GeForce), nv20 (GeForce 3) and r200 (Radeon 8500). There is also an additional best option which indicates that the game detected and set the best rendering path to use.
As you can see the arb rendering path almost doubles performance, reducing Doom 3 from a slide show to a bad quality video. With all graphics options turned off Doom 3 looks a lot like Doom I, only with a true 3D implementation. Unfortunately we have shown here that GMA900 is unsuitable for modern 3D gaming under ANY circumstances.
Conclusion
The conclusions in our tests are pretty straight forward. Dual channel memory implementation on Sonoma yields very little additional performance. Synthetic benchmarks show discernable performance increases, but real world testing shows that one shouldn't worry much about investing in dual channel memory. Photoshop CS does show a measurable performance increase with dual channel 512MB, shaving 23.4 seconds off our total Photoshop processing time. However with most applications that will benefit from dual channel implementation, like Photoshop, more RAM usually brings more performance than dual channel. From this writer's standpoint, it is better to buy a 1x1GB stick of DDR2 RAM rather than 2x512MB DDR2. It will be interesting to see how DDR2 667 with a 667MHz FSB on the upcoming Napa platform fares, hopefully with lower memory latencies! Predictions for desktop DDR2 800 RAM are latencies of CL=5, so if DDR2 667 were to run at CL=4 we would see a tangible increase in performance over today's DDR2 533 CL=4.
Rounding up our look at Sonoma, we investigated Intel's Graphics Media Accelerator 900 IGP solution. It is well known to most that its performance is abysmal, but it has been questioned whether dual channel memory makes a difference. In this article we've shown that while dual channel memory helps GMA900's performance and reduces the burden on the system, it is still unsuitable for any modern 3D usage. We love that GMA900 uses a minimum of 8MB in Windows and dynamically allocates up to 128MB, it is just too sad that the GPU itself is too weak to do anything useful with the nifty Dynamic Video Memory Technology. Hopefully GMA950 in the Napa platform will improve on this, but what we're hearing is that a clock speed bump is all that is in order for 3D performance. Regardless, dual channel memory does not provide any real world performance increase. This is mainly due to the poor performance of the GPU core itself. Perhaps in the future a more powerful core coupled with a high bandwidth bus and dual channel memory will be able to provide budget gaming performance from a shared memory IGP.
Page:1/1
Stay up-to-date with the latest mobile technology news and laptop reviews with our free newsletter! Get free tips, news and information from Laptop Logic and find the best laptop deals online.
){kind=link}
Comments
There are no comments yet.