In this post, I am comparing several options from the first wave of Pentium chipsets. The subject of the test is loosely defined as Socket 4 and Socket 5 PCI chipsets from Intel and other vendors from the 1993-1995 period. Specifically the following chipsets will be compared:
- Intel 430LX (Mercury)
- Intel 430NX (Neptune)
- Intel 430FX (Triton)
- Intel 430HX (Triton II)
- ALI M1451 Aladdin
- UMC UM8890
- SiS 501
- OPTi Viper-M
Despite comparing chipsets across different socket generations (Socket 4, 5 and 7), it is important to be using the Pentium CPU running at the same internal and bus speed. Fortunately, the ordinary Pentium 133 MHz (66 MHz Bus and 2x multiplier)can be easily matched against Pentium Overdrive 133 MHz for Socket 4 running at the exact same internal and bus speeds. In fact, the PODP5V133 CPU is internally a Pentium 54C, that is just adapted to run on a 5V motherboard which makes it a perfect choice for the test.
Several motherboards featuring a particular chipset were evaluated. I tried to pick the best one I could find. I ended up using the following boards:
- Intel Premiere/PCI board (Batman’s Revenge) with Intel 430FX chipset
- Intel Premiere/PCI II (Plato) with Intel 430NX chipset. I modified this board to support 2x multiplier.
- Intel Advanced/EV (Endeavor) with Intel 430FX chipset
- Gigabyte GA-586HX with 430HX chipset
- Gigabyte GA-586AM with UMC8891BF chipset
- Microstar MS-5107 with ALI M1451 chipset
- Ellitegroup SI5PI AIO with SiS 501 chipset. This motherboard has been modified to support 2M cache.
Some of these boards (notoriously all Intel motherboards) have only a limited BIOS tweaking options. In absence of direct BIOS settings I used hidden settings, BIOS modifications and chipset register programming to ensure that all performance-critical options are enabled so that the chipset performs at its best.
- Diamond Stealth 64 PCI VGA card with 2MB DRAM and S3 Trio64V+ chip
- 16MB EDO ram 50ns. Configured as either FPM or EDO (where supported).
- Adaptec AHA-2940U PCI SCSI controller with 512MB + 1GB disks emulated using SD2SCSI v6 device
- MS-DOS 6 and Windows 3.11. MS-DOS running in real mode.
Intel 430LX Mercury
The 430LX is the original Pentium Socket 4 chipset from Intel launched alongside with Pentium 60/66 CPUs. It is a 4-chip design with one system controller (82434LX), two LBX accelerators (82433NX) and a PCI/ISA SIO bridge (82378ZB). Optionally, in an EISA configuration there could be 82375EB + 82374EB chips raising the total number of chips to five. And that’s without any I/O or IDE controller which often was a discreet infamous CMD640 or RZ1000.
The chipset can be configured with either 256kB or 512kB of either asynchronous or rarely synchronous burst SRAM modules. The cache is often soldered on the motherboard with no option to upgrade. Unlike other chipsets doesn’t need the external tag and alter chip which is integrated into the chipset. The memory controller can handle up to 192M of RAM thanks to its 6 RAS lines. Memory timing is relatively slow at (7/11/14)-4444. EDO modules can be used, but EDO timings are not supported. On the other hand, all of the supported memory can be cached.
Performance-wise the chipset delivers average performance. It is obviously beaten by all Intel successors, but performs favourably against 3rd party options, especially considering it only has 256k of L2 cache.
Intel 430NX Neptune
The 430NX Neptune is scaled up 3.3V version of the Mercury chipset. Together with Pentium 90 CPU, it was launched in March 1994. The feature set is similar to Mercury. Nevertheless, the chipset also has some significant improvements. With the help of the integrated APIC in the P54C CPU, it supports multiple CPU configurations. It has two more extra RAS lines which together with 16M chip support improves installable memory size up to 512M of fully cached RAM. It officially supports 50 MHz bus speeds (used by Pentium 75 chip) with optimised refresh and timings. The cache interface is identical, however, the memory controller has one clock decreased lead-off latency improving memory cycle to (6/10/12)-4-4-4.
In benchmarks the chipset shows up to 15% higher RAM throughput which translates to approx. 3-4% higher performance compared to 430LX.
Intel 430FX Triton
The popular 430FX Triton is a member of next-generation Socket 5 chipsets designed exclusively for 3.3V P54C chips. It was first truly mainstream chipset which appeared on many different motherboards and sold at large quantities. Compared to the 430NX Neptune chipset it succeeded, it has many new features and performance enhancements. It supports the new pipelined-burst L2 cache configuration with noticeably reduced L2 latencies. This configuration is much more common than the synchronous burst cache variant found on some earlier chipsets and both provide substantial gains over traditional asynchronous L2. It was common for the cache to come installed on dedicated COAST (Cache On A Stick) slot allowing for easier upgrades. On the other, there were also some low-end cache-less motherboards or motherboard with traditional asynchronous SRAMs. I’ve done a comparison of different 430FX cache options in this post.
The memory interface is also much more refined. EDO rams are now supported which significantly improved read timings to (7,9,12)-2-2-2. The latency when using standard FPM improved too.
On the other hand, there were also some compromises. The 430FX was designed for mainstream computers. Compared to the Neptune the chipset no longer supports multiple CPUs. The number of RAS lines went down to 5 which reduced the maximum RAM size to 128M. 430FX motherboards have 4 SIMM slots only. The first RAS line is reserved for on-motherboard RAM which some had. The full tag space is no longer integrated into the chipset and the cacheable limit is only 64MB. These look like big losses, but 64/128M was still aplenty in 1995.
Even more significant changes can be seen in the south-bridge where the 82371FB (PIIX) chip introduced bus-master DMA support which was clear step up from previous buggy CMD640 based designs.
Performance of this chipset is top-notch second only to the 430HX. Both in low-level memory tests, PCI transfer speed and application it consistently beats the competition. However, if the Triton chipset is configured with asynchronous cache it doesn’t perform nearly as well and the performance drops to average numbers. It is not without surprise that due high latency of writes, asynchronous SRAM configs can be beaten even by cache-less configurations. Clearly, the Triton was designed to work with a pipelined-burst cache in the first place.
Intel 430HX Triton II
The high-end successor of Triton technically belongs to the Socket 7 era. But I thought it is a good idea to have it included as a reference. The 430HX is a significantly refined Triton designed for high-end use in servers and workstation. It came in a new BGA package. The chip count was reduced from 4 to 2 bringing back the traditional north/southbridge design. Multiple CPU support is back. On the other hand, the chipset finally dropped the support for asynchronous and synchronous SRAMs. Pipelined-burst SRAM is the only L2 option.
As for the RAM controller, 8 RAS lines are back improving the maximum RAM size to 512M. Depending on the configuration of the tag SRAM the cacheable limit is either 64M or 512M. Both EDO and FPM configurations have reduced latencies and all sorts of buffers are deeper.
The chipset officially supports PCI 2.1 specification namely PCI. Even more, changes are in the south-bridge (PIIX3) controller where we see the first-ever implementation of USB.
The 430HX is an undisputed winner of the competition. However, despite reduced timelines and deeper buffers, the lead over 430FX isn’t that big. Most likely a faster CPU than P133 is needed to see its full potential.
The first non-Intel chipset in my comparison is a solution from SiS. Released in 1994 it is a competitor of the 430LX chipset. It supports both P5 as well as 3.3V P54C CPUs. Three chip design consists of 85C501 system controller, 85C502 Local Data Buffer (the equivalent of Intel’s LBX chips) and 85C503 System I/O chip.
The chipset offers flexible L2 cache configurations from 64k to 2048k. The L2 is typically realised using asynchronous chips, optionally configured in a dual-banked mode requiring not less than 18 SRAM chips. A much less common configuration uses synchronous burst SRAM chips which allow achieving tighter cycles. Depending on the cache size up to 128M of RAM is cacheable.
The memory controller only has 4 RAS lines, however, 16M chips are supported so that the chip can support up to 128MB of RAM. The read latency is the same as on the 430LX, write latency is somewhat improved. EDO timing is not supported (but EDO RAMs can still be used in FPM mode).
Performance of this chipset is mediocre. In stock 256k L2 configuration, it finished my tests as the second last. Half-way between the worst ALI M1451 and Intel Mercury. The results are due to relatively slow L2 timing higher than 4-2-2-2 with 15ns SRAM chips.
However, the support 2MB of L2 cache configuration can make the chipset competitive when upgraded.
ALI Aladdin M1451
Another 430LX competitor from ALI supports 5V P5 Pentiums (although I’ve seen it adapted for P54C chips). Unfortunately, no datasheet could be found so one has to guess its specification from features on a very few motherboards that used this chipset. M1451 is budget two-chip design with M1451 north-bridge and M1449 south-bridge.
The chipset supports 256k – 1024k of asynchronous L2 cache.
The 6 RAS lines supports up to 196MB when 16M chips are used. EDO timings is not supported.
Except for a decent memory write performance, this chipset showed the slowest overall performance with particularly bad RAM read speed. That caused it to be the loser of this competition. However, the application performance, especially Quake, can be greatly improved by upgrading L2 cache to 1MB. When doing it the performance can match 430LX chipset with 256kB.
Viper-M is a Pentium chipset made by well-known 386/486 era chipmaker. One of several OPTi chipsets produced in early Pentium days. The M in the name stands for Multimedia which was a buzzword of the day. On paper the chipset is a very capable proposition, matching or exceeding Triton feature set. Supports both asynchronous as well as pipelined burst cache thanks to the standard COAST modules. Integrated IDE with bus master support. A bit unusual is support for both PCI as well as VL-Bus slave ports. Although the VL-Bus is not a true local bus as on 486 but a feature of the host bridge.
One interesting feature of the Viper-M chipset is that next to usual write-through and write-back also supports a unique adaptive write-back mode where a write-back cycle is converted to write-through depending on the state of the memory interface which might be beneficial in certain situations.
The Viper-M proved to be the most problematic chipset in the group. I got it on an Acer branded motherboard with only very limited tweaking options available. It showed compatibility issues with popular apps using the popular DOS4GW extender. It was necessary to replace it in order to run benchmarks.
Performance is a bit lagging though. Relatively slow write performance resulted in weak Quake score. One interesting fact is that at least on the motherboard tested there was no difference when used with pipelined-burst or asynchronous cache. The performance was identical. Clearly the pipelined burst cache did not enjoy the benefit of improved cycles.
A popular 486 chipset maker – UMC also produced a limited number of chipset for Pentium. I was keen to test this one as there are some claims that this is actually a crude modification of 486 chipsets delivering abysmal performance. Upon reviewing the performance of UM8891BF variant of the chipset I can confirm this is not the case. At least for the board and revision of chipset I tested.
The UM8890 consists of UM8891BF/UM8892BF memory controller and UM8886 PCI/ISA bridge and IDE controller is more or less equivalent of Intel Triton chipset with a similar feature set. Unlike some claims that UMC8891.
There is also AF revision of the UM8891 which doesn’t support pipelined-burst cache or EDO but I did not have one for tests.
It supports asynchronous as well as pipelined burst cache up to 1024k. It has 4 RAS lines supporting up to 128MB of RAM which can be either FPM or EDO with x-2-2-2 timings.
Given my test motherboard (GA-586AM) has both COAST slot as well as DIP sockets for asynchronous SRAMS, I was able to test the chipset was tested in several configurations. First with pipelined-burst 256kB cache and then with asynchronous cache configured to 256k, 512k and 1024k. The pipelined burst was clearly the fastest with 1024k configuration getting close in application tests.
First let’s have a look at L2 cache and memory read/write speeds as measured by speedsys tool
As you can see Tritons are well in lead followed by the UMC chipset where the difference between PB and async option is smaller than one might think. The rest of the group shows comparable performance with some interesting differences in individual members. But please bear in mind that all the numbers aren’t equal in this graph. The L2 read speed is by far the most important figure to look for here. That gives an early warning to Aladdin whose relatively poor showing is saved by relatively decent write performance.
Next is my favourite test. The framerates in Quake 1.06 timedemo at 320×200.
Both Triton chipset is still undisputed leaders followed again by the UMC. The SiS and especially Aladdin both show quite poor performance in default config with Aladdin having 33% lower performance compared to 430HX. However, Quake is the test where slow bus speed can be somewhat compensated by having large cache (so that the slow RAM is accessed less often). Upgrading L2 cache size nicely improved performance of SiS and Aladdin chipsets and made them competitive again.
Winstone 95 is a nice Windows 3.1 based benchmarks that should indicate performance in real-world business-oriented productivity apps of the era. Like Microsoft Word, Excel, Wordperfect, Paradox and others.
The rankings are unchanged here. Triton chipsets followed by UMC and then everyone else. ALI and SiS are still lagging behind. Again their performance can be improved by upgrading L2 cache size, although the improvement is smaller than in Quake.
This story is open ended as I intend to continue in my research. Specifically I’d like to do the following:
- Add more chipsets. Such those from OPTi (Python, Cobra), VLSI, UM8890 revision AF, SiS 5501 and possibly more
- Source better SRAM chips or motherboards sporting SiS501 and ALI Aladdin that would allow more aggressive L2 timings
As you can see the choice of chipset can make a significant difference. The gap between first and last chipset in Quake test is 38%. Which is equivalent of two CPU classes.
The results clearly show that Intel Triton (either the original or in Triton II variant) is the undisputed king of this generation of chipsets. It just shows how big deal the introduction of this first-party chipset was for the Pentium CPU.
The third-place forUMC8891 chipset came as a surprise after reports about its horrendous performance and outdated 486 design. Perhaps I was just lucky in chip revision lottery?
The original Mercury and Neptune chipsets delivered a solid performance which was adequate when they came to the market. For a long time they were etalons everybody wanted and compared with. The small improvements in 430NX proved to be useful for it to beat the 430LX.
And finally SiS and Aladdin. Those chipsets are lagging behind the Intel offerings they tried to compete with. Back in the days, they provided high value given by their price and level of integration. However, nowadays the support of 1024k and 2048k L2 cache gives them tuning potential if you can find the required SRAM chips. The SiS 501 is better of these two if you have to choose.