The market of Socket 478 chipsets looks quite busy now: the manufacturers are shifting toward dual-channel mainstream products, and they can twice cream the sales if they announce their new models before the release of Pentium 4 of 800MHz bus - by asserting that the chipset certainly supports future processors and, then, by making a newer version with enhanced, guaranteed and certified support. Socket A is in different situation: although NVIDIA launched the newer version of the dual-channel nForce, neither VIA nor SiS (who said "ALi"?) seem to plan that.
Everyone got tired of waiting for a new AMD's solution; Barton is not a match for development of new system logic and it must work flawlessly with existent solutions. Besides, the chipset makers have already announced that their products are ready for operation with the next-generation AMD CPUs - SiS755 and Crush K8, and VIA has already developed discrete and integrated versions for Athlon 64 - K8T400 and K8M400, as well as redesigned the previous one, that is why K8T400M based boards will also get onto the shelves.
Today we are going to have a look at the remaining Socket A chipsets. The only thing to be mentioned is that VIA is going to launch a new version of its KT400A in spring, - it will be the A-version of KT400, i.e. it's not going to add functionality to the north bridge (though it can boost memory speed), but it will certainly attract attention to VT8237 south bridge. We will return to it on its release, and now let's go with what NVIDIA and SiS offer.
NVIDIA nForce2 chipset was already discussed though that time it was discrete nForce2-S/ST (it was good that that version appeared first because it made good prospects in competition against VIA KT400). nForce2 took the lead but showed that the dual-channel memory access with the throughput equal or greater that that of the CPU bus is not much beneficial, and often doesn't justify the rise in price. Obviously, it helped to decide whether it made sense to bring out Socket A dual-channel chipsets for VIA and SiS (Socket 478 counterparts are already in production), and they left the battlefield for the Californian stubborns.
Boards based on the integrated nForce2 arrived only by the Xmas, mostly with MCP-T south bridge (nForce2-GT). This chipset has identical functions with nForce2-S/ST, but it supports integrated video core "in the person" of NV18 (GeForce4 MX 440-8x). Note that the integrated video core works at a lower clock speed than on external video cards, and the buffer in the system memory used as video memory works at a frequency also lower than 256 MHz of NVIDIA's external solution (GeForce4 MX 440 with AGP 8x).
Well, nForce2-G(T) based board will cost equally to the more efficient tandem of nForce2-S(T) board and NV18 graphics card, and the integrated solution will remain advantageous only for owners of slim desktop cases who can't use an external AGP card. But at least, they get such things as nView (video image coming to two receivers), DVI digital interface, integrated TV-Out and hardware DVD decoding.
SiS746 and SiS746FX
We tested SiS Socket A chipsets almost two years ago last time, at the time of single-chip SiS735. But the company wasn't hanging around, though our test lab couldn't fetch any of its products. What we finally received is not a mainstream solution, but it's the occasion to have a closer look at the latest products of Silicon Integrated Systems. The Socket A market is not crucial for SiS because its Pentium 4 chipsets, especially SiS650, turned out to be a success, and this market segment is much more attractive. Besides, productiveness of SiS's factories is almost depleted...
SiS735 was followed by one more single-chip SiS745 which increased memory support up to 3 GB, added DDR333 support and replaced Fast Ethernet MAC controller with FireWire. But today such solution is not attractive as it's incompatible with processors with 333MHz bus or DDR400, and the missing support for USB 2.0 and network leaves it in the low-budget sector.
Starting from SiS746 the company separated the bridges and created a line of pin-compatible south bridges. Whether it was needed from the technical standpoint or it was more convenient for performing the orders is hard to say. But now any board maker can use a cheap and functional version.
Either the integrated FireWire controller made the finished product pricey or the market wasn't ready for such an "elite" feature of traditionally cheap SiS chipsets, but the company is touting SiS963L (deprived of this controller) as a chip-companion for its SiS746 and SiS746FX. But SiS963L does look suitable as a south bridge if we compare these solution with VIA's counterparts instead of more expensive nForce2. Here are its brief specs:
The north bridge key features look decent. AGP 8x is supported in both 746 solutions; DDR400 and FSB 333 MHz are officially supported only in SiS746FX, but SiS746 works with them as well; Barton compatibility is also expected. So, the two latest SiS chipsets can be a good match for Socket A. Now let's compare performance of the competitors.
When we compared nForce2 based boards their scores hardly differed. And the today's tests keep the tradition. That is why the diagrams will show only two data types: on the discrete chipset version and integrated one.
Elitegroup K7S7AG sweeps the floor thanks to the external Xabre200 accelerator integrated on the board. The accelerator addresses the north bridge through AGP and it has its own memory also soldered to the textolite. So, in the non-graphics tests its results should be considered as those of the chipset. In the 3D graphics tests we will take into account the pair of SiS746+Xabre200 and add scores of the integrated video core nForce2-GT (in other applications the integrated graphics doesn't affect the nForce2's speed due to the dual-channel memory access).
And now some information on the memory modes: NVIDIA's chipset works fast of all when the CPU bus and memory work at the same clock speed, KT400 offers no choice with FSB 333 MHz fixing the synchronous mode, and SiS746 feels really awful when coupled with DDR400 (at least, with CL 2.5). Only SiS746FX has proved that it supports DDR400 which slightly boosts its speed, that is why the diagrams will show the results of this chipset coupled with the fastest first-generation DDR SDRAM memory.
Low-level memory tests.
We expected that nForce2-G and nForce2-GT with no sound would have equal scores (it's like for nForce2-S/nForce2-ST), but we did expect any difference between the integrated and discrete chipset versions. However, the randomly picked boards finished with the equal scores. As to SiS chipsets, the newer one works faster with DDR400 than with DDR333, and, worringly, SiS746 falls behind by 15% in the equal conditions.
In general, nForce2 is leading, especially in write speed tests. SiS746FX nudges out KT400, and, in spite of much lower read speed, keeps up with VIA's chipset when coupled with DDR400 and lags by 5% when with DDR333. SiS746 falls far behind this group, while nForce2, with its DASP enabled in stream copying operations, outpaces the competitors by 20%.
The first non-synthtic test, MPEG4 video stream encoding, much depends on how fast the chipsets work with memory. Reading operations prevail, that is why KT400 outscores SiS746FX by 2-3%. nForce2 is certainly ahead (the closest competitor is away by 4-5%), and SiS746 yields to its successor by 9%.
WinAce archiving is the extreme memory speed test with the 4MB dictionary. As the read speed prevails, VIA's models thrive again; NVIDIA is ahead by 11%, SiS746FX loses to KT400 by 5-7% which, in its turn, outdoes SiS746 by 13%.
Let's estimate now the chipsets' operation with memory which is the component determining the chipset's contribution in the overall performance. Earlier we compared nForce2 and KT400, and now the tests prove NVIDIA is faster by 5-15% in the tests where the memory performance is important. SiS746FX performs well in the synthetic tests and falls behind VIA's only by 5% in the real applications with DDR400, CL 2.5. DDR333 makes the gap wider, but we expect decent increase with PC3200 memory with reduced timings.
SiS746 demonstrates rather poor performance. It falls behind SiS746FX by 15%; maybe the chipset is not ready for operation with processors bundled with FSB 333 MHz, maybe it's because of the manufacturer (but 15% would be too much!), or maybe the low scores are caused by the external integrated video accelerator. Besides, we tested the preproduction sample of SiS746FX. I haven't found the real cause so far, but we will try to test another board from another company as soon as possible.
The second part of the tests in 3D graphics will start with SPECviewperf, a recognized synthetic benchmark for professional accelerators. Like in Wstream, the outcome depends on the memory speed with the read and write operations being almost equal. NVIDIA leads by 14-16%, and outscores SiS746FX+DDR333 even by 20%. The scores of nForce2 IGP and Xabre200 are given just to complete the picture because they are definitely weaker than GeForce4 Ti 4600 and can't be considered good accelerators in 3D modeling. However, in ugs-01 Xabre200 outdoes its competitors by 50%, and IGP2 demonstrates 0.000 fps because of the specific accelerator's load. In the other tests the situation looks similar to what you can see on the diagram.
In 3D graphics the difference between the contestants is not great in the gaming video modes, and the fight of SiS and GeForce4 MX were commented in the reviews of our Video System section - here I can only certify that the former has a great advantage.
In real games the situation is similar and the gaps between the boards coincide with the average difference in the tests: the leading nForce2 beat KT400 by 5-7%, SiS746FX lost to the latter by the same 5-7% at worst. It's obvious that in higher resolutions and at higher quality levels all the chipsets go on a par making the spread in values a couple of percent at High(Quality)@1280x1024x32. I'm not going to compare the integrated graphics as it was done before; just remember that at Normal@1024x768x32 both boards demonstrate decent playability.
We have revealed nothing new about nForce2-G/GT. Briefly, this is the most efficient Socket A chipset, and its integrated version has a speedy graphics core and some features useful for home users. But it makes no sense to install nForce2 into office machines because of its high price.
SiS746FX nicely fits into this niche due to its low price, decent speed and a bit less functional south bridge. Besides, this chipset supports DDR400(PC3200) and Athlon XP with a 333MHz bus. Also, it has some reserve for speed increase by the time SiS746FX based finished products get into hands of home users. It can be used in powerful home computers as well if a user wants to economize. SiS746 performs worse but the final verdict will be made after the tests with "purer" conditions.
VIA KT400, the de-facto standard for new Socket A chipsets, takes an
intermediate position between NVIDIA's and SiS's products. Universalism
is valued at all times, and VIA's chipset hardly needs any more words.
We will return to this company when KT400A is launched.
Write a comment below. No registration needed!