Silicon Integrated Systems (SiS), earlier known as a manufacturer of successful (and not very) low-end solutions for consumer desktops, decided on developing chipsets with the characteristics corresponding to the high-end level. Isn't it better to make products competing against high-end chipsets rather than being a leader on the small-profit market? Today we will examine two SiS chipsets developed for the niche traditionally occupied by Intel, the niche of high-performance systems characteristics and performance of which are much more important than price tags. At present, there are two SiS solutions deserving the high-end title: SiS655 with dual-channel DDR400 controller, and SiS R658 - the first x86 chipset ever supporting Rambus DRAM not from Intel.
If we leave aside the marketing frills which present even holders of IDE connectors as "Great Achievement of Progress" and "Patented Technology of Our Best Company", we can see that SiS R658 consists of the conventional SiS963 south bridge (USB 2.0 + FireWire, and other standard stuff) and SiS R658 north bridge. The north bridge provides functionality typical of all modern chipsets; also, it tramples down the sacred things (Intel 850E) and provides support for both innovative PC4200 RDRAM and even more innovative PC4800 RDRAM. Besides, it features AGP 8x, ATA133, 6-channel AC'97 2.2, integrated 10/100 Ethernet MAx in short, everything that any modern x86 chipset offers for any ordinary (Socket 478, Socket A) platform. That is why SiS R658 can be interesting only in terms of its performance. It looks progressive only because this is the only non-Intel RDRAM chipset. The achievement is significant but it's rather a marketing and psychological bait than a technical innovation.
SiS655 is much more interesting. Intel's first dual-channel DDR chipset for Socket 478 platform wasn't asynchronous to some reason, though Intel never saves on R&D. There had to be something they were afraid of... SiS encroached on the holy of holies by making a dual-channel asynchronous chipset: irrespective of FSB clock speed, the memory bus speed can equal 266, 333 and 400 MHz (in DDR equivalent), plus lots of intermediate and overclocking frequencies. Moreover, the two channels can work in two modes: "128-bit" and "2x64" (as SiS names them). All other things are well seen on the chipset's diagram above and hardly differ from SiS R658; that is why we will take a closer look only at the memory controller of SiS655:
A bus throughput of today's Pentium 4 doesn't exceed 4.2 GB/s (at the bus speed of 533 MHz; at 400 MHz it makes 3.2 GB/s), and the dual-channel memory controller coupled with DDR333 (2x2.7=5.4 GB/s) must boost the performance in contrast to 2xDDR266 (4.2 GB/s, but the processor is not the only device using memory bus bandwidth), though 2xDDR400 (6.4 GB/s) will hardly be a match for the 533MHz FSB. Anyway, we'll see that in the tests. However, the diagram does not explain the difference between the dual-channel modes: which one is better?
The documentation of the chipset (given in the scarce marketing PDF file available at the site) does not clarify the things. I'll try to guess. So, "1x128" and "2x64": what do the figures hide? It would be logical if the "128-bit" mode provided 128 bit capacity somewhere. But where? The only thing I have in my mind is a true 128-bit data bus composed of the dual-channel controller and two 64bit memory modules. But remember that the 64bit data bus does match Pentium 4 because its bus is of the same width, and a 128bit bus is too wide for it. It's not clear whether the game is worth the candle.
The 2x64 mode, as I understand it, must provide the functionality SiS showed on its pictures. The controller is actually single-channel, but the speed should go up because they are two and can work in parallel. In this respect, SiS655 doesn't differ from old NVIDIA nForce/2 and young Intel E7205. Such dual-channel mode reminds VIA's Bank Interleave, though VIA uses interleaving of logic banks inside the chip in case of the on-chip pipelining, and SiS interleaves external physical lines.
There is one more explanation we got from Maksim Len (expert in memory subsystems). Under "1x128" and "2x64" modes he understands operation in the parallel and interleaved modes. The former mode supports simultaneous addressing and instant sampling from all channels (a single address bus is not necessary), in the latter mode the controllers work in turn and multiplex streams inside. But Maksim says that this is one of the logical suggestions, and it shouldn't be taken as true.
Here are brief characteristics of the boards tested today for the first time:
In the 128-bit memory access mode on SiS655 boards you will need identical pairs of modules (in size and in chip organization). In 2x64 mode you must only insert the modules into slots relating to different channels. The single-channel mode has no special requirements. According to SiS, mainboards must build on the B revision of any chipset to support Hyper-Threading technology of Pentium 4. SiS reference board based on SiS655, A0 revision, doesn't support HT, but Gigabyte SINXP1394 built on the same chipset revision detects and enables two logic processors. That is why I recommend to check the documentation of a board you are to buy. The reference board was also tested but the results showed nothing new.
First of all, we will analyze the scores of the competing chipsets when each works with the fastest memory supported, and then we will examine data on SiS655 and SiS R658, namely their operation at different memory clock speeds. You can thus estimate the comparative performance of the chipsets in the conditions best for them, as well as an effect of the faster memory.
3DMark 2001 SE
The spread in the values is very small: 6% at most. Since inaccuracy of measurements is considered 2-3%, the difference between the boards makes nothing. Formally, SiS655 has the worst result in the single-channel mode, and SiS R658 also shows unexpectedly poor scores. i850E is undoubtedly the champion.
Return to Castle Wolfenstein
The situation is similar to the previous test, with SiS R658 looking rather sad again. As expected, the single-channel i845GE and SiS655 demonstrate pretty low performance, while the dual-channel version of the latter shares leadership with i850E, with 2x64 being more efficient than 1x128.
Serious Sam: The Second Encounter
The results are not unexpected. SiS R658 has a little improved its performance. Does the memory speed depend on an algorithm used? Maybe, SiS again plays with latency?..
Rendering in 3ds max
Well, 3ds max is not that sensitive to difference in memory speeds; it depends mostly on a processor. If you want to make any use of this diagram, regard it as the proof of an inconsiderable difference :).
i845GE has failed here as it doesn't benefit from PC3200: actually, it works equally with PC3200 and PC2700 (DDR i845GE doesn't support the memory clock speed over 333MHz).
We have run all the subtests of SPECviewperf 7, but in only two of them the scores of the competing systems are noticeably different: 3dsmax-01 and drv-08. So, the first one.
Firstly, the dual-channel DDR controllers are able to beat the champion i850E equipped with PC4200 RDRAM, though it concerns only Intel's E7205; i850E performed rather poorly in this subtest - it failed to outdo even i845GE. SiS655 lost to everyone in all modes. Moreover, its dual-channel modes hardly differ from the single-channel one, - it's very strange taking into account big amounts of data pumped through the memory. However, there is one thing which will be looked into further. SiS' RDRAM is quite weak.
In drv-08, i.e. the engine is from Design Review, SiS R658 falls far behind all the rest. The advantages of Rambus DRAM of i850E are well demonstrated! SiS655 again loses to the single-channel i845GE falling behind its direct competitors. Dual-channel SiS655 goes on a par with iE7205 (minus 2-3%) with the HT technology disabled. It's curious that performance of Intel's chipset decreases in some applications with HT enabled, but in SPECviewperf it makes +-2%, but not -13%! Taking into account SiS recommendations of the chipset revision required, SiS655 A0 seems to have a tricky support of Hyper Threading.
The leaders for Socket 478 platform - Intel 850E and Intel E7205 - fight against SiS655 in its both dual-channel modes; the single-channel i845GE competes against SiS655 in the single-channel mode. SiS R658 is an outsider, - its specs draw this chipset into the high-end niche, but its performance rises not higher than the mainstream level. Inside the competing groups the difference is not tangible, though the comparison of two dual-channel modes of SiS655 shows that 2x64 is speedier.
In comparison with WinAce, here iE7205 drops its speed even more relative to SiS655, and SiS R658 is also much closer to the outsiders.
Low-level memory tests in Cachemem
On one hand, everything is correct: the single-channel chipsets lose to the dual-channel ones, in accordance with the specs. This is synthetics :). But... iE7205 has changed its memory read speed markedly. Let's see what is the memory write speed before making the conclusion.
iE7205 reads well, but writes not much faster than the single-channel i845GE. i850E, the champion in writing, has been keeping its leading position for several months already; it's only dual-channel SiS655 that came close to it and even outscored the old man in reading. SiS R658 - I'd call that funeral :).
SiS chipsets: PC4200 RDRAM vs. PC4800 RDRAM, DDR266 vs. DDR333 vs. DDR400
These six diagrams are given together as the dependences shown are very similar. So, what do both SiS chipsets have in common in replacing the memory with the faster one?
Sadly, but we don't understand why SiS rolled out its SiS R658... I just hope that the board we tested could get damaged during the transportation. Or it is a very deep beta version, and the engineers' aim was only to make it work. So, if this chipset becomes widespread and manufacturers start making boards on it, we will retest it.
SiS655 looks more attractive. Its single-channel mode can be collated with i845GE to estimate effectiveness of usage of the memory controller in the chipsets of two manufacturers, but let's leave it for those who are really interested in it taking into account that the diagrams show that clearly. But it doesn't make much sense to compare them because it's not difficult to get two memory modules of the same speed for a dual-channel mode, and the speed different is well seen. Besides, I won't focus on the difference between two dual-channel modes of SiS655, because the requirements for "2x64" are not so strict compared to "1x128", and the latter turns out to be faster. We will only compare "SiS655 2x64" to i845GE, iE7205 and i850E and define winners and losers. It may seem to be too primitive, but it will shows the true situation.
So, let's zip through the diagrams and draw up a small table. "+" means
victory of SiS655, "-" shows that the competitor wins, and "=" means a
As a mainstream solution SiS655 has gotten 8 pluses and 3 minuses, - it's a very good result taking into account Intel's traditionally strong positions. But to make a good showing on the market finished solutions shouldn't be dearer than i845PE based boards, since the price factor is very important in this sphere. In the hi-end sector (for desktops) the chipset lost to i850E with the scores 2.5 pluses vs. 8.5 minuses. That is why, in terms of performance SiS655 is a direct competitor exactly for Intel E7205: 4 pluses against 6 minuses. So, the resume is that SiS dual-channel chipset for Pentium 4 with DDR SDRAM support is a direct competitor for Intel's dual-channel chipset for Pentium 4 with DDR SDRAM support.
There is one more aspect to be mentioned: iE7205 works with DDR266, and SiS655 with DDR400. The question is who is going to win when Intel brings out dual-channel DDR chipset with DDR333 support (aka Springdale-P)?.. However, the May is still far away. And now it all depends on the price. Should SiS655 boards be cheaper than those on Intel E7205, they will sell like hot cakes because "this chipset is not always slower, and it works faster in a great deal of cases". Should they have unimaginable price tags, the speed-crazed guys will find the chipset to be "slower in most cases". So, users' opinion will depend on what they see on the price tags. At the same time, SiS655 has a good advantage: conventional and inexpensive DDR333 memory is quite enough for it, that is why the platform's cost won't be affected much by this component.
From the technical standpoint SiS655 looks interesting and original,
and if the price is going to be reasonable, it has every chance to succeed.
Write a comment below. No registration needed!