BlueSmoke - Review : Pentium 4 Chipset Roundup

		Date	: Apr 2nd, 2002
		Category	: Chipset
		Manufacturer	: Various
		Author	: Jin-Wei Tioh

Earlier this year, Intel introduced the first 0.13-micron process based Pentium 4, based on the "Northwood" core. Prior to this, Pentium 4s were based on the 0.18-micron process "Willamette" core. The die shrink accomplished several important goals : reduce the 217mm^2 surface area of the Willamette (by upwards of 70%), power consumption and heat (due to small, faster switching transistors), and finally cost price per unit. What's more important is that the new cores have fantastic overclocking potential, surely catching the attention of performance users, as the Pentium 4's performance rests heavily on its ability to reach higher clock speeds than the competition (a direct consequence of its "deeper" pipeline architecture). With the planned increase of the Pentium 4's quad-pumped FSB, and the implementation of Hyper-Threading technology, the future certainly looks bright for the Pentium 4 platform.

Barring the future improvements however, what can an end-user do if he or she wants to get a top performing Pentium 4-based solution today? The performance of the P4 is critically dependant on memory bandwidth, which was why the P4 was launched on the RAMBUS memory platform. Although RAMBUS offered an incredible amount of bandwidth, it had a higher latency, causing its performance to be mixed rather than clearly superior. Coupled with the high price of RAMBUS RIMMs, and the P4 platform looked very unattractive. Intel tried to alleviate this with the i845 SDRAM chipset. Although the costs decreased dramatically, the use of SDRAM left the P4 starved of memory bandwidth, hurting its performance. Performance pundits criticized Intel the lack of P4 DDR support, and VIA promptly responded with their legally controversial P4X266 chipset., the very first DDR chipset. Intel finally bowed to market pressure, releasing the i845 DDR chipset several months later, and also granting use of the P4 bus to SiS, who launched the SiS645 chipset, a DDR-333(!) chipset.

So as it stands, the competition really boils down to this : i845DDR, P4X266 or SiS645? Join us as we take a close look.

The introduction of the P4X266 chipset sparked a legal uproar between VIA and Intel. Intel claims VIA had no right to use the P4 bus, whereas VIA claims that they already have that right due to Intel's previous agreement with S3 (which was acquired by VIA). But enough of law, this isn't Harvard. What we're more interested in is the technical aspect of the chipset.

It shouldn't have been all that difficult for VIA to come up with the design for the P4X266. It is the exact Pentium 4 equivalent of the KT266 Athlon chipset. The North Bridge uses the same AGP 4X controller, V-Link interface and DDR memory controller as the KT266 chipset. Probably the only unique part of the P4X266 is the Pentium 4 bus interface in place of the Athlon bus interface.

As suggested by the North Bridge, the P4X266 is also a two chip solution. The South Bridge used is the VT8233, sporting support for Ultra-ATA 100 and 6 USB ports. The VT8233 also features VIA's integrated network controller, and for brand name Ethernet support, the VT8233C has an integrated 3Com controller. The datapath between the two components is VIA's V-Link, an 8-bit bus having a peak bandwidth of 266MB/s. Like many of VIA's chipset, the P4X266 support asynchronous operation of the memory bus and FSB. This is a required feature due to the Pentium 4's FSB being clocked at 100MHz (quad-pumped to 400MHz), while DDR266 (or PC2100) SDRAM runs off a 133MHz clock.

The VIA P4X266 chipset test platform used was a PE11-S motherboard from the VIA Platform Solutions Division. The PE11-S is a solid contender as a motherboard, having good layout, good stability (no crashes observed during tests) and feature set : 3 DIMM slots and 5 PCI slots (all full-length capable).

Around last September, Intel introduced the i845 chipset. To address the cost concerns of the i850 RDRAM platform, the i845 supported standard PC133 SDRAM. The Pentium 4's 64-bit, quad-pumped 100MHz FSB offers 3.2GB/s of peak bandwidth to the processor. Ideally, the available peak memory bandwidth should match of even exceed this figure to minimize any bottlenecks. PC133 SDRAM can only offer a peak bandwidth of 1.06GB/s, only 1/3 the amount. As the Pentium 4's performance depends heavily on memory bandwidth, the i845 severely impeded its performance. Still, the i845 sold well, pulling market share away from AMD. Informed consumers however, shied away from the P4 altogether, opting instead for an AMD solution.

After a legal row with VIA, Intel finally bowed to market pressure and released their highly anticipated i845D chipset, being essentially the same chipset as the i845, but with DDR capability allowed to be utilized. It is also a two chip solution, consisting of the i82845 MCH (Memory Controller Hub) and the i82801BA ICH2 (I/O Controller Hub). The i82845 MCH supports either PC133, DDR200 or DDR266 memory, and obviously also asynchronous memory bus and FSB operation for PC133 and DDR266 support. The i82801BA is a carry-over from i850, i815 and i810 chipsets, which accounts for the reason why the feature set of the i845 seems a little dated; only 2 USB host controllers (4 USB ports), ATA-100, and a maximum memory capacity of 2GB (compared to the 3GB or higher limit of the P4X266 and SiS645).

However, some motherboard manufacturers like MSI and EPoX have equipped their boards with additional features such as ATA-133 RAID. Such is the case with our Intel i845D chipset test platform - the MSI 845 Ultra-ARU. An excellent contender as a motherboard, the 845 Ultra-ARU has good layout (aside from the RAID headers being placed right at the bottom), good stability (no crashes observed during tests) and feature set : 3 DIMM slots, 5 PCI slots (all full-length capable) and hardware-based 6 channel integrated audio.

SiS came into the Pentium 4 chipset market from out of nowhere, blind-siding quite a few parties (yours truly included). Without warning, I received a reference SiS645 motherboard together with a stick of Micron DDR333 memory marked "Engineering Sample". This is probably one of the most exciting chipsets of the year, and with good reason. What with the good performance reputation of the SiS735 and SiS745 Socket-A chipsets, many wonder just what SiS can do in the Pentium 4 chipset arena.

As mentioned earlier, SiS has a Pentium 4 bus license, which puts them in a much better legal position than VIA. On paper, the SiS645 looks absolutely fabulous. A two chip solution with the 645 North Bridge, and the 961 South Bridge, offering 6 USB ports and ATA-100 support. It's made special by its memory controller - an advanced version of the one in the SiS 735. What's so special? It officially supports the operation of the memory bus at 166MHz, ie. DDR333 with a maximum memory capacity of 3GB. This gives the Pentium 4 2.7GB/s of peak memory bandwidth, still shy of 3.2GB/s, but better than the 2.1GB/s offered by DDR266. Like the i845 and P4X266, the 645 supports the asynchronous operation of the FSB and memory buses with three different settings : 100/100, 100/133, and 100/166MHz. The North and South Bridges are interconnected with SiS's MuTIOL technology. It is a 16-bit wide bus clocked at 266MHz, designed for multiple isochronous virtual channels. Thus, it is able to offer a peak bandwidth of 533MB/s both upstream and downstream, being the most advanced chipset interconnect on the market so far. Although it is questionable whether such a large headroom is required, there's no harm done.

Incidentally, to make the 645 more attractive for OEMs, it is pin compatible with the forthcoming 650 North Bridge. The SiS650 is basically just the SiS645 with integrated SiS315 video.

Our SiS645 test platform is the reference motherboard supplied by SiS. The board was not only shocking because it supported DDR333, but because it was too large to meet ATX specifications. Outfitted with 6 PCI slots, 1 ACR slot and 1 AGP slot, it has one slot too many to fit in any ATX case. An interesting observation is that both the North and South Bridges lack any cooling solution whatsoever. We performed several stability tests to verify its stability, and the board remained stable even after going through several "makeworld -j4"s under FreeBSD.

The P4X266, i845 and SiS645 will be pitted head-to-head along with an existing platform for the Pentium 4, the i845 SDRAM. Before you examine the test setup, here is a comparison of the chipsets :

Pentium 4 Chipset Technical Comparison

	VIA P4X266	Intel 845	SiS 645
North Bridge	P4X266	i82845 MCH	SiS645
South Bridge	VT8233	i82801BA ICH2	SiS961
Memory Technology	SDRAM / DDR	SDRAM / DDR	SDRAM / DDR
Memory Type/s	PC133 / DDR200 / DDR266	PC133, DDR200, DDR266	PC133, DDR200, DDR266, DDR333
Effective Memory Clock Speed	133 / 200 / 266MHz	133 / 200 / 266MHz	133 / 200 / 266 / 333MHz
Peak Memory Bandwidth	2.1GB/s (DDR266)	2.1GB/s (DDR266)	2.66GB/s (DDR333)
Maximum Addressable Memory	3GB	2GB	3GB
Bridge Interconnect	V-Link	Intel Hub Architecture	MuTIOL
Peak Interconnect Bandwidth	266MB/s	266MB/s	533MB/s
USB Ports	6	4	6
USB 2.0 Support	No	No	Yes

Without further ado, we'll put the chipsets through their paces with our suite of benchmarks.

The DDR P4 chipsets have a common claim to fame - provide improved memory performance over SDRAM solutions. By far, the SiS 645 operating in DDR333 mode is the fastest contestant, roughly 15% ahead of the fastest DDR266 chipset, the i845. Operating in DDR266 mode, the P4X266, 645 and 845 are within 4% of each other, with the 845 being the fastest. The PC133 SDRAM equipped Intel i845 takes a beating, lagging severely behind.

Switching over to Cachemem, the SiS 645 operating in DDR333 mode has a clear latency advantage. Surprisingly however, the 645 actually has higher latency memory accesses (by far) in DDR266 mode. The P4X266 and i845 are roughly on par.

The disk performance of chipsets is often neglected in chipset comparisons. In this aspect, many hold Intel chipsets in high regard, and the results confirm this. Both the P4X266 and 645 lag the 845 by 6% on the High-End Disk WinMarks, and by a whooping 30% on the Business Disk WinMarks. As disk performance is also a significant factor in overall performance, these results should not be taken lightly.

End-users utilizing "external" (ie. non-integrated) drive controllers shouldn't notice this particular performance disparity.

ZD's Business Winstone 2001 measures real world system performance by seeing how fast a system can finish a set of common business tasks, including Microsoft Word 2000, Excel 2000, Access 2000, FrontPage 2000, PowerPoint 2000, Project 98, Norton AntiVirus, Netscape Communicator 4.7, Lotus Notes R5 and NicoMak WinZip.

All applications are run on a multi-tasking basis. The whole benchmark suite being run five times (highest score taken), with automatic hard drive defragmentation and rebooting in between each run. This effectively minimizes the margin of variation to within 3% or less. The final score indicates how much faster the system is compared to ZD's base system.

The SiS 645 is 7% faster than the P4X266 in DDR266 mode, with the 845 keeping very close pace. Switching over to DDR333 only yields a 2.5% increase in performance, pretty much unnoticeable by the end-user.

The FreeBSD "make world -j4" command stresses all major system components and subsystems. In addition to being an excellent stability test, it is utilized in this chipset comparison to measure overall system performance. The DDR333-equipped SiS 645 completed the "make world" in the fastest time, 5319 seconds. The P4X266 and DDR266 i845 lag behind by about 6 minutes. However, the performance turned in by the 645 in DDR266 mode lags the closest competitor, the P4X266, by a full 15 minutes! We can only speculate that this somehow correlates with its higher DDR266 memory latency noticed earlier.

Quake III Arena has always been sensitive to memory bandwidth in general. Thus, it comes as no surprise that the 645 with DDR333 SDRAM edges out the DDR266 competition by 8%.

The Return to Castle Wolfenstein MP Test also showed big improvement in scores with higher available memory bandwidth. The 645 operating in DDR333 mode was 11% faster than the DDR266 competition. However, this gain decreases under Dronez and Max Payne, ostensibly because the GeForce2 MX became the bottleneck rather than the chipset.

3DMark2001 is also fairly sensitive to memory bandwidth, with the 645 with DDR333 SDRAM clearly leading the pack. Here, the Intel 845, which has showed strong DDR266 performance thus far, edges out both the P4X266 and 645 by 2%.

The POVray test was executed by rendering a chess set (available in our Downloads section) at 1024x768, with an anti-aliasing factor of 0.3. Here the SiS 645 outperforms the other chipsets in both DDR266 and DDR333 mode, a somewhat unusual showing as the 645 had been lagging the other DDR266 contestants thus far. The P4X266 seems to have some issues with this application, as it comes out just ahead of the PC133-equipped i845.

SPECviewperf is a portable OpenGL performance benchmark program which provides a vast amount of flexibility in benchmarking OpenGL performance. The SPECviewperf suite consists of 6 viewsets, each corresponding to a particular application, eg. Intergraph's DesignReview (DRV), IBM's Data Explorer (DX) and Alias/Wavefront's Advanced Visualizer.

In the Alias/Wavefront viewset, all three DDR chipsets are virtually on par, with the SiS 645 pulling ahead by a healthy 5% when operated in DDR333 mode.

In the remaining viewsets, the 645's additonal bandwidth when operated in DDR333 mode doesn't really buy it any performance advantages, at the most 1%, which is well within the benchmark's margin of error. The Intel 845 is exhibits very competitive performance here, second only to the DDR333-equipped 645. Here however, the P4X266 seems to lag by 6% to 9% behind.

The earlier poor success of the Pentium 4 was mainly due to the i850 and its RAMBUS-only architecture. At the time, a RDRAM DIMM cost significantly more than comparably sized DDR and SDRAMDIMMs.

The P4X266, despite its age, is still certainly a viable product. However, VIA's still unresolved legal entanglement with Intel effectively means that many motherboard manufacturers, especially larger ones such as AOpen, ASUS, and MSI will not touch the chipset with a 10-foot pole, leaving only VIA's own Platform Solutions Division and several smaller manufacturers as the producers of P4X266-based motherboards. This limits the choice for consumers.

On the other hand, the SiS 645 is the Godsend manufacturers are looking for to roll out DDR solutions for the Pentium 4. Official support for DDR333 SDRAM is an added perk which may or may not give up to a 7% performance increase depending on the application. Its DDR266 performance is generally competitive, despite its higher memory latency. With better performance than the P4X266, the 645 looks set to dominate the Pentium 4 DDR scene save one factor : Intel's DDR chipset.

The DDR266-equipped i845 shows very strong performance, coming in a close second to the DDR333-equipped 645. The same i82801BA ICH2 used since the i815 chipset is still going strong (save for ATA-133 support), turning disk performance numbers anywhere from 6% to a whooping 30% faster than both the 645 and P4X266. This is definitely an area in which SiS and VIA need to catch up on, although users utilizing non-integrated drive controllers won't see this issue. Considering the excellent reliability of the very first SDRAM-based i845 motherboards, accompanied by Taiwanese manufacturers' claim that the DDR-based i845 will be the next BX chipset, the i845 could displace the 645 as the chipset of choice, despite the 645's lower cost and higher performance.

Ultimately, the SiS 645 in DDR333 mode is the top dog in this comparison. Although as of yet unapproved by JEDEC, DDR333 or PC2700 DDR SDRAM isn't too difficult to find nowadays, with manufacturers such as Kingmax producing CAS 2 modules. The 645 reference motherboard performed flawlessly even under heavy use, which indicates that reliability should not be an issue with 645-based products. However, VIA is poised to release the P4X333 in the near future. Whether or not the 645 can fend it off remains to be seen, but for now, the SiS 645 remains the top choice for DDR Pentium 4 performance.