Celeron 1

The Celeron 1.2 GHz processor has been around for a couple of years, now; so, this little review is rather late in coming. It was stimulated by my curiosity about whether these "Tualatin" Celerons offered any sort of an upgrade path for systems built around a PIII processor with the earlier "Coppermine" core. My PIII-1000 MHz system is getting rather long in the tooth and is having trouble keeping up with current games. This isn't too surprising, since it is over three years old. A 1 GHz processor pretty much represents the top of the line for this system's I815E motherboard. Eventually, Intel went on to produce faster PIII processors (up to 1.4 GHz), based on the Tualatin core, but these new PIIIs were not compatible with the earlier motherboards that supported the Coppermine PIIIs. These new PIII-S processors came with 512KB of L2 cache integrated into the core and were primarily aimed at the entry level server market. They have remained relatively expensive, even up to today. (A PIII-S 1.4 GHz processor sells for over $200.) They also found there way into notebooks as the PIII-M line of processors. Apparently, Intel never really targeted these beefed up PIIIs for the desktop market, probably because it didn't want them directly competing with their newly launched Pentium 4 line of processors.

Many hardcore computer enthusiasts figured out how to use fine pieces of wire for reconfiguring some connections on the their motherboard's socket in order to hack their way around the incompatibility between these Tualatin processors and older motherboards, such as the I815E. Eventually, PowerLeap made things easier by producing adapters that allow the Tualatin processors to work on older motherboards. (Note: the adapter described in this review and pictured above is the PL-Neo/T, and it requires a motherboard that fully supports the Coppermine PIII in the FC-PGA package. Motherboards, such as those that use the i440BX chipset or the i810 chipset, require one of PowerLeap's other adaptors, such as the PL-370/T (for socket 370) or PL-iP3/T (for slot 1).)

	PIII Coppermine	Celeron Coppermine	PIII-S Tualatin	Celeron Tualatin
Process Tech	.18 micron	.18 micron	.13 micron	.13 micron
L1 Cache	32 KB	32 KB	32 KB	32 KB
L2 Cache	256 KB	128 KB	512 KB	256 KB
Bus Speed	100/133 MHz	100 MHz	133 MHz	100 MHz

Comparing the Celeron Tualatin to the PIII Coppermine, we can see that they are rather similar. The Tualatin Celeron is primarily held back by its 100 MHz bus speed, as well as the introduction of some extra latency in the memory timings. However, the Tualatin core incorporates better prediction logic about what to store in it's memory cache, which should be an advantage with many applications. It is manufactured using the newer, .13 micron process, so it uses a lower core voltage and produces less heat.

XbitLabs goes into more detail about the history of the Tualatin core and how the PIIIs and Celerons based on this core differ from their predecessors based on the Coppermine core. See their review of the Celeron 1.2 GHz for more information.

Both the PIII-S and the Tualatin Celeron have a product line that goes up as high as 1.4 GHz, but the Celeron 1.2 GHz model seemed like the best choice for use with an older i815e motherboard. As already noted, the cost of the PIII-S processor remains relatively high. If one wanted to spend that much money on an upgrade, it seems to me that it would be better spent towards making a complete upgrade (motherboard, processor, and memory). As noted above, one of the drawbacks to the Celeron is that it is hampered by being run on a 100 MHz bus. Therefore, I was interested in a Celeron that I thought stood a good chance of being overclocked to run on a 133 MHz bus. Looking at Overclockers.com CPU Database, it appeared that 1.6 GHz was a fairly typical overclock for the Tualatin Celerons (without requiring any exotic cooling or sharp increases in voltage to the core). The Celeron 1.2 GHz, with its locked multiplier of 12 (12x100 MHz=1200 MHz or 1.2 GHz), would be running right at 1.6 GHz when the bus speed is bumped up to 133 MHz (12x133=1596 MHz).

When installing a Tualatin processor using the PowerLeap PL-Neo/T adaptor, you first press the processor into the adaptor, and then you install the adaptor into your motherboard's ZIF socket. Since you must press the pins of the processor into adaptor, once they are joined, it will be very difficult to separate them. The inclusion of the adaptor between your motherboard's socket and the processor means that the processor is now raised up above the socket by about 4 mm (the thickness of the adaptor). The Tualatin processor also has an integrated heat spreader over the core, which makes this processor slightly taller than the earlier Coppermine processors with the exposed core.

As a result of these factors, your heatsink clip is going to have an extra stretch in order to reach down to the clip points on the socket. I was able to get my Alpha PAL35 heatsink reinstalled by using a bit of brute force on the clip, but this might be a problem for some clips. The PL-Neo/T comes with a heatsink and a special clip that is capable of having the length of its reach adjusted. Unfortunately, this clip wasn't compatible with my Alpha heatsink, though it looks like it would work with the heatsink included with retail versions of the Tualatin Celeron. (My Alpha heatsink uses a copper core, along with aluminum radiators, so I thought it might be a bit more effective than either the PowerLeap or Intel heatsinks, which are all aluminum.)

My Asus CUSL2 motherboard misidentified the processor as a PIII, after I booted the system up, but it also gave a little error message about not properly recognizing the CPU. This isn't surprising, since the motherboard was never intended to support this processor, so no BIOS updates have ever included the proper ID code. However, this is strictly a cosmetic issue and will not affect the performance of the processor.

The PowerLeap adaptor fixes the core voltage at 1.55v, which is slightly higher than the 1.5v that this processor's specification calls for. I considered this to be a good thing, since I wanted to overclock the processor. Although my BIOS menu supported making changes to the core voltage when a PIII processor installed, this option disappeared after the Celeron and its adaptor were installed, leaving no way to make further changes to the voltage. (Apparently, one of the advantages of the PL-370/T adaptor is that you can make changes to the processor's core voltage, probably by adjusting some jumpers on the adaptor.) The PL-Neo/T has only one jumper, for selecting either 100 MHz or 133 MHz bus speed. I selected 100 MHz, so the processor would default to its normal speed. I overclock the processor by changing the bus speed in the BIOS.

In this section, we'll take a closer look at how the Celeron 1.2 GHz stacks up against a Coppermine PIII 1.0 GHz. We'll look at the Celeron when it is running at its default speed of 1200 MHz on a 100 MHz and when it is running at an overclocked 1600 MHz on a 133 MHz bus. We'll do a quick sampling of synthetic benchmarks, number crunching tasks, and a couple 3D benchmarks. While this is hardly an exhaustive protocol of benchmark testing, it should give us a sense of what a computer user might expect from this processor.

We start out with a look at SiSoft Sandra's CPU benchmarks. Basically, the results confirm the similarity between the PIII and the Celeron, since the results scale proportionally with the increased clock speed of the processors. In other words, the Celeron 1.2, which is running at a clock speed 20% faster than the PIII 1000, and produces scores that are 20% higher than those of the PIII 1000.

Sandra's other CPU benchmark, which makes use of Intel's SSE instruction set, tells us basically the same thing as the previous CPU benchmark.

Turning to ScienceMark's CPU benchmark, we get a somewhat different picture. Here the Celeron does not seem to scale in a manner equivalent to the PIII. While the Celeron at 1600 MHz is running 60% faster than the PIII 1000 MHz, it is only producing 36% more floating point operations per second.

ScienceMark can also show us the same results in a different way by representing how many floating point operations occur with each processor cycle. As we can see, the PIII is getting more done with each tick of the system clock. It may be that the ScienceMark's CPU benchmarks are more memory dependent than Sandra's are. Let's see why that might make a difference.

We'll start out with Sandra's benchmarks, again. These results seem to highlight the fact that the PIII 1000 and the Celeron 1600 both have their system memory running on a 133 MHz bus, while the Celeron 1200 is relying on memory running on a 100 MHz bus. Thus, the scores for the PIII 1000 and the Celeron 1600 are pretty similar, while the scores for the Celeron 1200 lag behind.

ScienceMark indicates that perhaps there are some differences between the Celeron and the PIII, after all, since the Celeron at 1600 MHz scores a bit less than the PIII 1000, despite both using 133 MHz memory bus. Let's dig into memory subsystem a bit more.

Sandra attempts to break out the performance of the L1 cache by using a relatively small block of data, 16 KB, which should fit into the 16 KB L1 data cache of these processors (meaning that there should be little need to use either the L2 cache or system's main memory). We see how the bandwidth of the integrated L1 cache, which runs at the same speed as the CPU's core, scales proportionally with increases in the processors' speeds.

Using a larger block of data, which at 128 KB is too large to fit into the L1 cache, but small enough to be contained by the 256 KB L2 cache, we can take a look at how the L2 cache performs. Sandra indicates that L2 memory performance scales according to the processor's clock speed, regardless of whether it be a PIII or a Tualatin Celeron.

And, ScienceMark's benchmark confirms this. However, ScienceMark can give us a look at how much Latency is involved in these memory cache operations, i.e., how quickly they are being performed.

We see that with the L1 cache, the latency, measured in clock cycles, is the same for the Celeron and the PIII, so as the processor runs faster, the time required for memory operations goes down. Turning to the results for the L2 cache, we see something interesting.

Intel has introduced an extra wait state in the operation of the L2 memory cache on the Celeron processor. This is probably a way to hobble the Celeron a bit more, so it doesn't infringe upon the performance of Intel's top of the line processors. Whether this makes a significant difference or not may depend upon the kind of data being worked with. The Tualatin Celerons have better predictive algorithms for determining which data should be kept available in the L2 cache, and this may offset the extra latency, at least when the predictions prove right. But, if they miss, the Celeron's performance will be slowed compared to that of the PIII.

Well, so much for theory, let's look at some benchmarks that strive to give these processors something more like real world work to do.

Prime 95 is a distributed computing project that searches for extremely large prime numbers. The client program includes a benchmark utility that lets you see how quickly your computer can complete some calculations of various lengths. As can be seen in the above graph, the PIII 1000 and the Celeron 1200 run about even on this task, while the Celeron 1600 is modestly faster.

ScienceMark includes several number crunching tests. One is an encryption benchmark in which lower times indicate faster or better results.

Here we can see a nice improvement over the PIII 1000's performance by the Celeron 1600, which is about 35% faster.

Above, we can see that the Celeron 1600 is about 45% faster than the PIII 1000 on a molecular dynamics simulation task and that the Celeron 1200 is about 20% faster.

However, on another simulation task, the memory bandwidth issue seems to significantly hamper the Celeron 1200, which is only 5% faster than the PIII 1000, while the Celeron 1600 is only about 26% faster.

The variations in how much the Celerons benefit from their faster clock speed, compared to the PIII, emphasize how much other factors, such as memory bandwidth, and perhaps memory latency, can play a role in a computer's performance. Turning to a couple of 3D graphics benchmarks, next, we can see this conclusion displayed yet again.

In the above results from the popular 3D Mark 2001SE benchmark, we can see how memory bandwidth comes into play. Despite the higher clock speed of the Celeron 1200 compared to the PIII 1000, the performance is essentially the same (similar to what we saw with the Prime95 results). The Celeron 1600, which has roughly the same memory bandwidth as the PIII 1000, is only able to show some modest performance gains over the PIII 1000 (about 26% at the lower 800x600x32 resolution).

The results from the DroneZ benchmark show the same trends, but even more dramatically. The Celeron 1200 is unable to keep up with the PIII 1000, while the Celeron 1600 is only slightly faster (about 12% at 800x600x32) than the PIII 1000. These game benchmarks seem to emphasize how much memory bandwidth can become a bottleneck for these processors running at speeds over 1 GHz. Clearly, PC100 SDRAM is choking the performance of the Celeron 1200, and it seems likely that PC133 SDRAM is unable to keep up with the needs of our Celeron when it is running at 1.6 GHz. Consequently, we don't see as much benefit from these faster processor speeds as we might have hoped for when they are presented with memory intensive applications.

In many respects, the Tualatin Celerons might be considered faster versions of the Coppermine PIIIs. Their primary handicap is that they are specified to run on a 100 MHz bus, but this can be overcome by overclocking, if your system supports 133 MHz bus speeds and you select the right Celeron. The extra latency in the L2 cache is another problem, but it doesn't seem to have a significant impact on performance, perhaps because the Tualatins incorporate better predictive logic into the functioning of their cache. Nevertheless, memory bandwidth appears to be the Achilles heel of these processors. PC100 memory isn't sufficient for keeping up with a 1.2 GHz processor, and PC133 memory doesn't fully support the capabilities of a 1.6 GHz processor.

So, do these processors qualify as a cost effective upgrade path for an older system? Well, that depends. The PL-Neo/T cost $25 and the Celeron 1.2 (boxed) cost $40. (The other PowerLeap adaptors for Tualatin processors, the PL-370/T and the PL-iP3/T, which have their own voltage circuitry, cost $55.) For about this same amount of money, you can purchase a relatively modern processor, such as the Athlon XP 1800. It's actual clock speed is 1.53 GHz, which is in the same ballpark as our overclocked Celeron. Of course, if you go this route, you'll need to make a complete overhaul of your system. You'll need to buy a new motherboard, DDR memory, and perhaps even a new power supply, in addition to your Athlon. Your total outlay will be closer to $300 by the time you are done. The advantage of going ahead and doing this more thorough rebuild of your system is that you'll end up with parts that complement each other well and perform better as a whole. As we have seen, replacing the processor in your system with a faster Tualatin Celeron may eliminate the processor as the main performance bottleneck, but this will reveal limitations in other areas, such as your system's memory capabilities.

To try to distill some guidelines, I would suggest that upgrading to one of these Tualatin processors is worthwhile, if you currently have a processor running at about half the speed of the Tualatin you are upgrading to and if you have some uses for this computer that don't require high end performance. For example, some folks like to keep a second computer running for various tasks, such as running distributed computing projects (e.g., SETI, Prime95, etc.), internet file sharing, backing up data from a main computer, or as a substitute for when their primary computer is offline. With modest expectations and appropriate tasks, this computer upgrade might be money well spent. However, if you are looking for a high end system with the computing horsepower to take on the demands of today's latest games, for example, you'll need to dig deeper into your pocket and purchase not only a faster processor (something on the order of a P4 2.4 GHz or an Athlon XP 2500 would be best), but you'll need to purchase a new motherboard and new system memory, as well.