Original Link: https://www.anandtech.com/show/833
The always controversial Rambus DRAM is not as poor of a purchase today as it was a couple of years ago. This is partially due to the presence of a processor that can actually take advantage of the bandwidth offered by RDRAM and also because memory prices in general have been spiraling downwards. A quick visit to www.pricewatch.com will tell you that the lowest price on a 256MB PC800 RDRAM module is 3.28x as much as the lowest price on a 256MB PC2100 DDR SDRAM (DDR266) module. Although this is somewhat reminiscent of the price disparity between RDRAM and SDRAM from two years ago memory prices have come down so much that we're talking about a difference of $57; two years ago it would have been $600.
The dramatic reduction in memory prices over the past 24 months softened the launch of the world's first Pentium 4 DDR chipset, VIA's P4X266. Although the performance of the P4X266 clearly approached that of the RDRAM based Intel 850 chipset, cost of memory wasn't as big of a reason to go down the VIA path as it once was. Another problem that was stacked against VIA was the touchy issue of licensing for their P4X266. If you ask Intel, the P4X266 should not be sold because VIA does not have a license to use the Pentium 4 bus. If you ask VIA, they don't need to strike a licensing agreement because of a prior agreement Intel made with S3, a large part of which is now a part of VIA. However, if you ask the most important players in this game, the motherboard manufacturers, most of them won't touch this situation with a 10ft pole. There are a select few that have released P4X266 based motherboards, but none of the major manufacturers have elected to go to production with their boards. It isn't that they don't have them ready, since virtually all of the major motherboard manufacturers have a P4X266 design; it's mainly an issue of waiting for Intel and VIA to sort things out.
Until recently, most motherboard manufacturers have only been able to take a "wait and see" stance on the issue of 3rd party chipsets for the Pentium 4 platform. In a move very reminiscent of their earlier launch of the 735 chipset, SiS has stepped up to the plate and offered motherboard manufacturers a very attractive solution. The 645 chipset is SiS' answer to the P4X266; and after recently going to production many manufacturers are finding themselves instantly attracted to it as it is the first Pentium 4 chipset to receive an official license from Intel.
As if legality wasn't enough, the SiS 645 officially supports DDR333 SDRAM; giving the Pentium 4 another 25% of what it likes the most: memory bandwidth.
Two chips this time
One of the "features" of the SiS 735 chipset was its single chip design allowing for simplified motherboard layouts. The single chip was home to an integrated North and South Bridge connected by SiS' own Multi Threaded I/O Link (MuTIOL). The 735 chip was very large and the move to a Pentium 4 solution pushed the single chip design to the limits forcing SiS to revert back to a two chip design.
There are no real performance downsides to a two chip design and since motherboard manufacturers are used to producing these types of boards it's really no more difficult for them to adjust to a two chip SiS solution. The single chip design of the 735 was simply an added benefit to manufacturers since they didn't have to layout the traces connecting the North and South Bridges.
The SiS 645 North Bridge is made truly special by its memory controller. The memory controller is a slightly more advanced version than the one found in the SiS 735. Just like the 735's memory controller, this one supports both regular SDRAM and DDR SDRAM. Unlike the 735 however, the new memory controller officially supports the operation of the memory bus at 166MHz; more specifically, in the case of DDR SDRAM, DDR333 mode.
Chipset
Memory Bandwidth Comparison
|
||||
Intel
845
|
Intel
850
|
SiS
645
|
VIA
P4X266
|
|
Memory Bus Width |
64-bits
|
16-bits
|
64-bits
|
64-bits
|
Memory Clock Speed |
133MHz
|
300/400MHz
|
100/133/166MHz
|
100/133MHz
|
Number of Transfers per Clock |
1
|
2
|
2
|
2
|
Number of Channels |
1
|
2
|
1
|
1
|
Theoretical Memory Bandwidth |
1.06GB/s
|
2.4GB/s
- 3.2GB/s
|
1.6GB/s
- 2.7GB/s
|
1.6GB/s
- 2.1GB/s
|
With a 25% increase in clock speed over DDR266 modules, DDR333 offers a proportional increase in memory bandwidth; from 2.1GB/s of DDR266 to 2.7GB/s courtesy of DDR333. There are no differences in the modules themselves; rather the DDR SDRAM chips on the modules have to be certified for operation at 166MHz DDR. In spite of this requirement, you'll find that a lot of the higher quality DDR266 modules will work fine at 166MHz. In our tests, the Crucial DDR266 CL2 modules managed to run at 166MHz CL2.5 just as well as the officially rated DDR333 modules. For our tests we used a DDR333 module from Kingmax. This module makes use of Kingmax's BGA packaged chips.
Since the Pentium 4's FSB only runs at 100MHz (albeit quad-pumped) it's obvious that the 645 North Bridge has the ability to run the FSB and memory buses asynchronously. In particular, the 645 supports three different settings (FSB/memory): 100/100MHz, 100/133MHz, and 100/166MHz.
Taking a page from "How to Make a Successful Chipset 101", the SiS 645 North Bridge is also pin compatible with the forthcoming 650 North Bridge. The only difference being that the 650 North Bridge will feature an integrated SiS 315 video core making it very attractive for OEMs.
MuTIOL & The South Bridge
The 645 North Bridge uses SiS' MuTIOL interconnect to interface with the 961 South Bridge. This bus operates at 66MHz and is double-pumped meaning that it transfers data twice (and addresses once) per clock cycle. The bus is 16-bits wide and is bi-directional, allowing for 266MB/s of data to be transferred in each direction for a total of 533MB/s of bandwidth between the North and South bridges. This is twice what Intel's 850 and VIA's P4X266 offers between the two chips. Internally, the MuTIOL interconnect gets its name from the fact that there are multiple virtual channels, a total of 10 in all, that connect the individual parts of the South Bridge (for example, IDE controller, USB controllers, PCI slots, etc…) to the MuTIOL interface. These 10 virtual channels offer the same 1.2GB/s of bandwidth that the SiS 735 offered between its integrated North and South Bridges. All of this talk of peak bandwidths is mostly useless since it's very difficult to saturate the 266MB/s of bandwidth offered by Intel's Hub Architecture and VIA's V-Link interconnect much less the 533MB/s SiS' MuTIOL provides. In the future this will change, but that future is still quite a distance away.
SiS concludes that the need for more than 266MB/s of bandwidth between the North and South bridges is necessary based on the following bandwidth requirements:
SiS'
Theoretical South Bridge Bandwidth Requirements
|
|
Device
|
Bandwidth
|
AC'97 Audio |
2MB/s
|
Integrated Ethernet MAC |
12MB/s
|
Legacy PC Functions (ISA, KB, Mouse, etc...) |
16MB/s
|
USB |
2
x 12MB/s
|
PCI |
133MB/s
|
ATA/100 |
2
x 100MB/s
|
Total |
387MB/s
|
While those are real peak bandwidth requirements, what isn't made clear is that it is current impossible to hit 100MB/s on any single ATA/100 drive; saturating the 133MB/s PCI bus is also very difficult without the use of a RAID card and a few hard drives; and finally USB 1.0 is 12Mbps per channel, not 12MB/s indicating each channel requires 1.5MB/s of bandwidth and not 12MB/s as SiS indicates in their 645 presentations. For a more realistic comparison, here's our take on the North/South interconnect bandwidth requirements for today's PC:
Realistic
South Bridge Bandwidth Requirements
|
|
Device
|
Bandwidth
|
AC'97 Audio |
2MB/s
|
Integrated Ethernet MAC |
12MB/s
|
Legacy PC Functions (ISA, KB, Mouse, etc...) |
16MB/s
|
USB |
2
x 1.5 MB/s
|
PCI |
50MB/s
|
ATA/100 |
2
x 20MB/s
|
Total |
123MB/s
|
SiS will eventually offer a replacement to the 961 South Bridge in the form of the 962 which will introduce SiS' own IEEE-1394 (Firewire) controller to their Pentium 4 chipset. This IEEE-1394 controller is integrated into the single chip 745 which is due out soon. The 745 is the successor to the 735 that implements the enhanced memory controller from the SiS 645.
Definitely not a reference board
The board SiS supplied us with for testing was far from a conventional reference board. Not only did the board not meet the ATX specification (it is physically too large) but it was definitely not optimized for performance. Before you take that the wrong way, let's explain.
The SiS board pictured above is an internal test board, outfitted with one too many slots to fit in any ATX case. This is mainly so that the engineers can test the board with all 6 PCI master slots as well as the ACR and AGP slots. You'll also notice that the board features three DIMM slots which is very rare when a manufacturer is attempting to tune a board for optimal performance.
Maintaining the highest performing memory timings with three DIMMs on a board is very difficult, which forces most manufacturers to tune for performance by stripping one DIMM off of the board. This is usually acceptable for chipset reference boards, but generally a practice not undertaken by motherboard manufacturers because of a demand for more memory slots. This is also one of the reasons that reference boards generally outperform the first generation of motherboards based on that chipset when they actually hit the streets.
Not only were we surprised when we saw that the 645 board was given to us with three DIMM slots, but it also worked perfectly fine with three DIMMs installed. The icing on the cake was the fact that it worked fine with three DIMMs installed in DDR266 and DDR333 mode. Even the SiS engineers present with us hadn't validated the board for use with three DDR333 DIMMs which definitely surprised us.
Motherboard manufacturers have expressed quite a bit of interest in the 645 platform and unless VIA can turn things around with Intel soon, you should expect to see (according to SiS) some of the bigger names announce boards this month and ship before year's end. For comparison purposes we used a retail P4X266 motherboard from Shuttle (AV40R) and the i850 based ABIT TH7-II RAID. We used an updated BIOS currently available on ABIT's website for the TH7-II RAID that improved performance of the board a few percent from when we first looked at the board in our Pentium 4 2.0GHz review.
With that said and done, it's time to benchmark…
The Test
Windows 2000 Test System |
||||||
Hardware |
||||||
CPU(s) |
Intel Pentium 4 2.0GHz | |||||
Motherboard(s) | ABIT
TH7-II RAID (Intel 850) Shuttle AV40R (VIA P4X266) SiS Engineering Sample 645 Motherboard (SiS 645) |
|||||
Memory |
256MB
PC800 Mushkin RDRAM |
|||||
Hard Drive |
IBM Deskstar 30GB 75GXP 7200 RPM Ultra ATA/100 |
|||||
CDROM |
Phillips 48X |
|||||
Video Card(s) |
NVIDIA GeForce3 64MB DDR |
|||||
Ethernet |
Linksys LNE100TX 100Mbit PCI Ethernet Adapter |
|||||
Software |
||||||
Operating System |
Windows 2000 Professional Service Pack 2 |
|||||
Video Drivers |
|
Memory Performance
Normally we can take a look at the memory bandwidth and latency scores and generally predict the outcome of the rest of the review, at least when it comes to chipsets. However, with the Pentium 4 the situation isn't as predictable. It was clear from our P4X266 Review that the the 3.2GB/s of memory bandwidth offered by the i850 chipset wasn't all being used, but how much is necessary for today's Pentium 4 processors and applications? We'll find out throughout the course of this review, but as usual we'll start with some theory:
|
In spite of having a 19% theoretical bandwidth advantage, the i850 is able to offer a 26% memory bandwidth lead over the DDR333 SiS 645. It's difficult to say exactly what we can attribute this to but we can hazard two guesses. Either RDRAM is simply more efficient than DDR SDRAM on the Pentium 4 platform or the SiS 645 platform has some untapped potential that is not being exploited in the reference board. We're leaning towards the latter judging by SiS' warning that the board we were testing hadn't been tuned for performance and our own observations on the matter. Time will only tell once production boards hit the streets.
|
Here we see a clear latency advantage by SiS only when operating in DDR333 mode. The 645 actually offers higher latency memory accesses than the P4X266 in DDR266 mode. In spite of this, none of the solutions come close to the 850 in terms of memory latency. Now it's time to see how these figures translate into real world performance.
Business & Content Creation Application Performance
It's been with us since the start and we continue to use it; Winstone has always been a great benchmark as a way of measuring performance as seen through the eyes of the normal user that is bound by the same I/O limitations (HDD performance) as everyone else.
|
If we say that the only significant performance advantages are those above 10% we can conclude that the SiS 645 does have a significant performance advantage over the P4X266 as it is 11% faster in DDR266 mode. Switching to DDR333 buys you another 2% performance improvement which is pretty much unnoticeable. The Intel 850 is pretty much as fast as the SiS 645 with DDR266 SDRAM.
|
Content Creation Winstone is a bit less disk limited than Business Winstone and thus stresses difference in memory bandwidth even more. Here we only have a 9% advantage when comparing the two DDR266 contenders, but the Intel 850 moves up a notch to be on par with the DDR333 equipped SiS 645.
Internet Content Creation & Office Productivity Performance
SYSMark 2001 is another favorite, especially considering that it is definitely not as disk-bound as Winstone although both are equally useful.
|
The SYSMark definition of Internet Content Creation involves much more than just running a few content creation applications. A good portion of the performance in this benchmark comes from running an avi through Windows Media Encoder while using Dragon Dictate, a speech recognition program, along with a handful of other content creation applications. While you would expect all of this to put a heavy load on memory bandwidth, it doesn't. In fact, there is only a 2% spread between the top and bottom performers here. Strength in content creation comes from strength in the Pentium 4, not just its memory bandwidth.
|
We get a repeat of the results here under the office productivity tests; there simply isn't a large enough spread to even bother splitting hairs between these contenders.
|
The overall performance picture that's painted is identical to what we saw in the previous tests; the SiS 645 equipped with DDR333 SDRAM comes out on top but the performance advantage isn't significant (only 2% over the P4X266, less than 1% over the 850).
Video Encoding Performance
In order to test video encoding we used Flask coupled with the latest official release of the DivX codec (v4.01, from www.divx.com) and encoded a 120MB MPEG-1 video. The encoding resolution was 352 x 288, the Fast SSE2 iDCT codec was chosen, deinterlacing was enabled as was the highest quality filtering, and we did not decode audio. The end result is the following performance graph measured in frames per second:
|
Again, we aren't seeing a large performance difference between the contenders. The 645 with DDR SDRAM does come out on top once again, but barely holds any sort of lead over the competition.
|
...and the same holds true for MPEG-2 video encoding.
3D Gaming Performance
Next we turn to the 3D gaming scene to show us a use for the added bandwidth of DDR333 over DDR266:
|
Quake III Arena has historically been a great showcase of improvements in memory controllers and memory bandwidth in general; thus it is no great surprise that the 645 with DDR333 SDRAM holds a 8% lead over the P4X266 with DDR266 SDRAM. The lead is cut in half when both platforms are outfitted with DDR266 SDRAM and is reduced even further when you increase the resolution beyond 640 x 480, so the lead again diminishes in value.
|
Also based on the Quake III Arena engine, the recently released Wolfenstein MP Test was included in this review using the second of two benchmarks we created for our GeForce3 Titanium Review last week. This demo, 'atdemo8' is more platform and video card limited than the other more CPU limited demo we created but even then, the span of scores doesn't exceed 4%.
|
To finish things off, we have another 4% lead by SiS with the 645.
3D Rendering & Animation Performance
For our readers that dabble in the professional 3D realm, here we have a breakdown of the six viewsets that make up SPECviewperf 6.1.2. For more information on exactly what each viewset simulates click here.
|
The most interesting thing here is that there is actually a decent sized lead by the DDR333 equipped SiS 645 over the competition (8%). Memory bandwidth dependencies couldn't be the explanation here since that would mean that the Intel 850 chipset would have to be at the top of the charts. The only other thing that comes to mind is a somehow superior AGP controller.
|
With DDR333 SDRAM the SiS 645 is able to offer performance rivaled only by the Intel 850 here; assuming close to a 14% lead over the VIA P4X266.
|
The P4X266 has some serious issues in the Data Explorer viewset as it is clearly outperformed by the competition. Even when equipped with DDR266 SDRAM the SiS 645 outperforms it by close to 37%.
|
Again we see more performance issues from the P4X266 in this test; it could very well be the Shuttle board being tested since we didn't see these types of issues when we looked at VIA's reference P4X266 board.
|
While not doing as badly as it did in the earlier tests, the P4X266 still finds itself at the bottom of the charts and the 645 with DDR333 SDRAM holds a 10% advantage. When paired with the same DDR266 SDRAM, the advantage is more than cut in half and reduced to 4%.
|
Final Words
It's the start of a new month and we get another DDR chipset for the Pentium 4 platform. The only difference this time around is that the SiS 645 has Intel's blessing which is worth its weight in gold to motherboard manufacturers. Ever since Intel and VIA filed suits against each other, most motherboard manufacturers have stepped out of the conflict and refused to release P4X266 based solutions. With the 645, SiS may be the knight in shining armor that a lot of the manufacturers are looking for to provide them with a solid DDR solution for the Pentium 4. The official support for DDR333 SDRAM is an added perk that will give you another 0 - 5% performance improvement in most cases, but definitely not needed to get good performance out of the SiS 645.
The SiS 645 doesn't hold much of a performance advantage over the VIA P4X266 or the Intel 850 but it should be pointed out that it is technically faster than both of those solutions. And the timing on the 645, albeit a little later than SiS would have liked to gain the full effect of a "legitimate" Pentium 4 chipset, couldn't have been better. Motherboard manufacturers have clearly expressed interest in the solution and will most likely pursue 645 based designs unless things settle between Intel and VIA very soon. There is one kink in this plan for chipset domination however, and that is Intel's own DDR solution for the Pentium 4.
Due out in the first quarter of next year (remember that's less than 4 months away) is Intel's 845-D chipset with support for both DDR200 and DDR266. If the SiS 645 isn't found on motherboards until November or December in mass-quantities, then it may be difficult to force many people to go with a SiS based solution instead of waiting two months for an Intel solution. Granted that the SiS 645 based motherboards will be cheaper, but considering how reliable the first round of 845 motherboards were and the early proclamation of the 845 being the "next BX chipset" by most manufacturers in Taiwan, it will be difficult to justify the 645. This brings us to another point about chipset releases. Somewhere along the lines it became "acceptable" for chipsets to be released without a sign of motherboards for weeks or even months to come. Compare this to an Intel chipset launch where motherboards (quite a few) are available on launch day and you'll see why we're always uneasy with talking about performance when the only board available is one supplied to you by the manufacturer of the chipset. In the case of the SiS 645 we actually expect shipping boards to be faster than this engineering sample, but competing chipset manufacturers should start coordinating their chipset launches better with the motherboard manufacturers that will be implementing those chipsets.
The first manufacturer producing 645 based motherboards is, as you'd expect, ECS. You'll remember that ECS also had the world's first widely available SiS 735 based motherboard which actually turned out to be a pretty reliable design from ECS. Their 645 based motherboard should be very cheap and we're hoping will boast a similar if not slightly improved overall build quality to the ECS K7S5A (SiS 735). Boards like the upcoming offering from ECS, as well as solutions from ASUS and MSI (provided that they actually hit the market) could redefine the entry-level Pentium 4 standards. Instead of having to deal with an underperforming PC133 SDRAM equipped 845 board, a DDR266 board with the SiS 645 could be had. For OEMs, a similar scenario exists but with the upcoming 650 (w/ integrated video) as the chipset of choice. Even in spite of Intel's upcoming 845-D, there won't be another Pentium 4 chipset with integrated graphics other than the P4M266 for quite some time; and an integrated graphics core could definitely saturate the remaining bandwidth offered by DDR333.
Setting concerns regarding Intel aside, it does look like SiS does have another winner with the 645. This is oddly reminiscent of the release of the 735 chipset; the only question that remains is whether this spark will ignite another fire within VIA causing a P4X266A to be released down the road. It's unlikely that VIA can get more performance out of their memory controller without some more serious modifications, potentially adding DDR333 support.
Will VIA shorten the life-span of yet another SiS chipset? From an OEM standpoint, probably not because of the strengths in SiS' upcoming integrated solution. But when looking at the 645 itself, chances are that if VIA doesn't, Intel will.