Introduction

You all can remember a time when Intel was considered to be the mainstream and performance market segment leader, with their processors found in the more expensive retail systems.  In the value market, AMD reigned free as their K6-X line of CPUs couldn’t cut it in terms of the performance that was necessary for most gaming or professional level systems. 

Intel began to slowly eat away at AMD’s value market share after the introduction of their Celeron A in late 1998.  With a high performing desktop CPU under their belt, and now a high performing value CPU beginning to ship as well, Intel set the tempo for what would carry them through the next two years. 

Unfortunately, a few bad decisions on Intel’s part left the company with a mainstream and performance market segment platform that no one wanted.  Instead, they turned to AMD for their Athlon for use in the majority of $1,000+ retail systems.  AMD continued to deliver very well, and Intel continued to lose a great percentage of the performance desktop and mainstream market segments to the Athlon. 

One thing that didn’t change very much at all, in spite of the release of AMD’s Duron processor, was Intel’s grasp on the value market segment.  Still using the same basic Celeron that was released in late August 1998, although on a different manufacturing process with a new core, Intel has managed to continue to succeed in an area that used to be dominated by AMD. 

At Comdex 2000 AMD provided us with a presentation that listed some fairly interesting statistics which illustrate perfectly Intel’s position in the market.  The average system price for a retail AMD system in October was a little over $1,300 while the average system price for an Intel system in the retail market was almost $400 less.  This clearly shows that while the Athlon is being found in most AMD systems, the Duron is underutilized.  At the same time, it shows that, on Intel’s side, the Celeron is driving the average price of Intel systems down considerably. 

In an attempt to continue their successful focus on the sub-$1000 retail market, Intel has released another Celeron processor, now bringing the number of clock speeds introduced since its release in 1998 up to 16 with the Celeron 766. 

The Chip

The Celeron 766 remains unchanged from the previous Celeron processors.  It is still manufactured by disabling half of the cache on low-yield Pentium IIIs that have bad cache blocks, thus giving the Celeron half the cache as well as making that cache half as associative as the Pentium III’s cache.  This ends up hurting the Celeron severely as it, clock for clock, is unable to outperform the Pentium III, even when both are clocked at the same FSB/memory clock. 

The explanation for this is simple, the less associative a processor’s cache is, the lower hit-rate it has, and the lower the hit rate, the more times the CPU has to go to main memory to fetch needed information.  In the case of the Celeron, the memory bus is running at a horrendously slow 66MHz which just amplifies the downsides to its 4-way set associative L2 cache versus the Pentium III’s 8-way set associative L2. For more information on cache mapping techniques, check out our one page guide here.  

Celeron 766 Specifications

·        0.18-micron Coppermine128 core

·        766MHz clock speed – 11.5x clock multiplier

·        32KB on-die L1 cache running at core speed

·        256-bit Advanced Transfer Cache – 4-way set associative 128KB on-die L2 cache running at core speed

·        Advanced System Buffering

·        370-pin FC-PGA Socket-370 GTL+ CPU interface running at 66MHz

·        1.65v core voltage

The Celeron 700 is really no different from the 566/600MHz parts that have been on the streets for a while now.  It is still based on the same 0.18-micron Coppermine128 core, features the same amount of L1 cache (16KB Instruction + 16KB Data cache) and the same 128KB inclusive L2 cache. 

CPU Specification Comparison
	AMD Duron	AMD Athlon			Intel Pentium III		Intel Celeron
Core	Spitfire	K7	K75	Thunderbird	Katmai	Coppermine	Mendocino	Coppermine128
Clock Speed	600 - 800 MHz	500 - 700 MHz	750 - 1000 MHz	750 - 1200 MHz	450 - 600 MHz	500 - 1000 MHz	300 - 533 MHz	533 - 766 MHz
L1 Cache	128KB				32KB
L2 Cache	64KB	512KB		256KB	512KB	256KB	128KB
L2 Cache speed	core clock	1/2 core	2/5 or 1/3 core	core clock	1/2 core	core clock
L2 Cache bus	64-bit					256-bit	64-bit	256-bit
System Bus	100 MHz DDR (200 MHz effective) EV6				100 - 133 MHz GTL+		66 MHz GTL+
Interface	Socket-A	Slot-A		Socket-A Slot-A (OEM only up to 800MHz)	Slot-1	Slot-1 Socket-370		Socket-370
Manufacturing Process	0.18 micron	0.25 micron	0.18 micron		0.25 micron	0.18 micron	0.25 micron	0.18 micron
Die Size	100mm^2	184 mm^2	102mm^2	120mm^2	128mm^2	106mm^2	153mm^2	106mm^2
Transistor Count	25 million	22 million		37 million	9.5 million	28 million	19 million	28 million

Making sense of it all

If you remember back to our review of the Duron 800 we pointed out that at 800MHz the Duron was not considerably faster than any Celeron that would be released in the next six months.  Not to mention that at 800MHz, AMD would continue to extend their clock speed lead over Intel’s Celeron.  Intel’s philosophy with the Celeron, especially after the Duron started outperforming it at lower frequencies, was that clock speed sells, however that has obviously taken a back seat in this case. 

In spite of AMD’s performance and clock speed advantage in the value market segment, they unfortunately don’t have enough design wins for the Duron.  We have discussed the main reason behind this time and time again, there is no integrated video platform for the Duron while the Celeron has the i810, enabling it to be found in sub $600 systems.  According to Intel, the Duron isn’t really hurting the Celeron’s sales because of this, which is the reason that the Celeron won’t move to a higher performing 100MHz FSB until the first quarter of 2001. 

Intel’s assessment of the Duron’s impact on Celeron sales is partially correct and partially incorrect.  In Europe, the Duron sales have simply been incredible, however in the North American markets the Duron continues to suffer in favor of the Celeron.  The next chance AMD will get to push the Duron into retail systems will be early next year, conveniently around the same time when Intel will hope to bridge the performance gap somewhat by giving the Celeron a much needed 100MHz FSB. 

With all of that in mind, it makes perfect sense that Intel would release another 66MHz FSB Celeron before the year’s end.  Unfortunately it seems like it’s too late in the year to actually have the Celeron 766 found in any retail systems, and at the same time, with the processor just being recently introduced, it hasn’t even had time to drop in price to a reasonable level along with the rest of the Celerons.

Assuming we give the Celeron 766 the benefit of the doubt, and expect that it will drop down to the $100 arena shortly, how does it stand as an upgrade?  Obviously for a freshly built system, the Duron is the best option, but what about for the person that already has a BX board and maybe an older processor?  Does it make more sense to scrap your motherboard and pick up a KT133 + Duron or is the Celeron 766 able to breathe cheap new life into your existing system?  That is the focus we will be taking in this review.  We already know that the Celeron, clock for clock, is completely dominated by the Duron so there’s no point in going down that path of comparison any further, instead we’ll be looking at the value of the Celeron 766 as an upgrade, and compared to other $100 CPUs from AMD/Intel. 

The Test

New Benchmarks

To kick off the performance tests we have the first benchmark from Ziff Davis Media.  For those of you that may remember, the media giant CNet acquired ZDNet in a huge buyout, leaving the magazine portion of what used to be Ziff Davis all alone.  Acquired by a holding company, this magazine portion of what used to be Ziff Davis (responsible for PC Mag, etc…) is now operating under the name Ziff Davis Media.  With the cooperation of ZDBOP, now called eTesting Labs, they have produced a new set of 2001 benchmarks.

We had the opportunity to speak with a few representatives of Ziff Davis Media at this year’s Comdex and discussed with them our concerns over the future of benchmarking, especially with the recent fears surrounding BAPCo’s relationship to Intel and the impartiality of their benchmarks being jeopardized by that.  It is good to know that Ziff Davis Media completely understood our concerns and are looking to restore a bit of sanity to the benchmarking world. 

As you may or may not know, the old Ziff Davis Winstone benchmarks (Business & High End) used to be the de-facto benchmarking standards.  However the pace of hardware improvements has outlasted the benchmarks greatly, thus causing limitations in the benchmarks to surface and the numbers to begin to lose value.  Since December 20, 1999 we stopped using Business Winstone 99 because of this very reason, and switched over to Content Creation Winstone 2000.

Ziff Davis Media and eTesting Labs have brought back the next set of benchmarks, their 2001 suite.  Now focusing much more on “hot spot” operation (performing longer tasks at once, instead of constantly switching tasks like the old tests), the 2001 benchmarks should be much more accurate indicators of overall system performance. 

Without further ado, let’s see how these new benchmarks stack up.

Remember that we aren’t shooting for a clock for clock comparison, rather aiming to compare processors that fall around $100 in price, although some of these are a tad more than that figure.  Also remember that we are giving the Celeron 766 the benefit of the doubt and expecting its price to drop down to within a similar range shortly, so how does it perform?

Under business applications, in spite of its crippling 66MHz FSB and memory bus, the Celeron 766 is still approximately 7% faster than a 500MHz Pentium III.  For business applications however, a 7% improvement is hardly worth an upgrade.  This illustrates a major theme of this review, which is our take on the age-old adage, don’t judge a book based on its clock speed. 

The two AMD solutions, in spite of being lower in price and lower clocked than the Celeron 766 that is available today, offer noticeably superior performance.  This is due to a number of factors, the 200MHz EV6 bus that both the Duron and the Athlon share as well as their more associative, larger, L2 caches. 

The 14% improvement carried by the lower clocked Duron over the Celeron is something that is most expected, and we have proved this countless times.  However, for the user that currently owns a BX, or even an LX board with proper BIOS support, the Celeron is a much more attractive option since you get rid of the cost of a new motherboard.

As we mentioned at the start of this review however, in the case of a user constructing a new system from scratch, the Duron is a much more attractive alternative.

Content Creation Winstone 2001 is a much better benchmark to compare performance differences in, since the type of applications being run tend to be much more performance demanding than your casual set of business/office applications. 

Here we see that the clock speed advantage over the Pentium III 500E does appear to be giving the Celeron a decent advantage, 12% to be exact.  Before questioning why we even bothered including the Pentium III 500E as a comparison point wait until we take you through all of the benchmarks. 

It is clear though, that as a general purpose CPU, without any focus on gaming or professional level applications, the Celeron 766 could be a decent upgrade to an older Pentium III system. 

Again we see the Duron’s sheer dominance which is why we desperately want to see the Duron used in more retail systems. 

SYSMark 2000 takes a different approach to benchmarking than the two Winstone tests we just took you through.  Instead of focusing on multitasking, which we believe is how the majority of enthusiasts use their computers, SYSMark 2000 focuses on testing the performance of a single application without any task switching. 

Without any focus on task switching, the memory usage of the tests go down (keeping one program in memory is easier than keeping two or more) and thus the stress on the memory bus goes down as well.  This allows the Celeron to play catch-up to the Duron, decreasing its advantage to under 9%, and its SSE optimizations aid its performance as well. 

The Pentium III 500E is also closing in on the Celeron 766, trailing by only 6% this time around.  This will be the last time that the Pentium III 500 loses to the Celeron 766, at least for the next few pages.  Surprised?  We told you there was a reason for including such a “slow” processor in this comparison.

What happened to the Celeron 766’s lead over the Pentium III 500E?  No longer is it a 6 or 7% lead, instead the Pentium III 500E outperforms it by over 5%.  Explaining the performance difference is simple. 

The Celeron has four things counting against it here, and they all help to nullify any performance gains its 766MHz clock would give it. 

For starters, the Celeron is still using a 66MHz memory bus.  If you remember back to our original Celeron 600 Review, we pointed out that one of the two most memory bandwidth intensive tests we ran was the Quake III Arena benchmark. 

Secondly, confounding the effect of the 66MHz memory bus is the effect of the 66MHz FSB.  We mentioned in the same review that Quake III Arena can appreciate a very high bandwidth FSB.  More recently, we pointed out similar findings in our Pentium 4 Review courtesy of the processor’s 400MHz FSB. 

The third factor working against the Celeron here is that its L2 cache has a lower hit rate than the Pentium III because of the fact that it is only 4-way associative.  The Pentium III enjoys an 8-way associative L2 cache, and the Duron/Athlon have a 16-way associative L2 cache.  The more associative a processor’s cache, the greater the hit rate. 

Finally, the Celeron only has half of the L2 cache of the Pentium III.  We can prove that this is effecting its performance by looking at the 7% performance drop from the Athlon to the Duron.  Both CPUs run at the same clock, have the same FSB/memory bus clock, the only difference is that the Athlon has a L2 cache that is 4 times larger than that of the Duron.  The Celeron’s 128KB L2 is dwarfed by the Pentium III’s 256KB L2, which also plays a role in its performance disadvantage. 

Even at 1024 x 768 x 32, a resolution that we generally declare as memory bandwidth limited by our testbed’s GeForce2 GTS graphics adapter, the Celeron comes out just slightly slower than a Pentium III 500E. 

So while the Celeron was a fairly decent upgrade for SOHO/business users that had older Pentium II/III systems, it isn’t as attractive for gamers.  But before we pass judgment let’s take a look at how the numbers shape up in other gaming benchmarks.

The Celeron immediately hops back to its original position ahead of the Pentium III 500E under MDK2.  The MDK2 test is not nearly as intensive as Quake III Arena, in terms of the stress it puts on the Celeron’s FSB and memory buses, thus allowing its clock speed advantage to kick in and give it close to a 10% advantage over the Pentium III 500E. 

The Duron continues to dominate as does its older brother, nothing out of the ordinary there.  Even as a gaming platform the Duron puts the Celeron to shame, it wasn’t long ago that the Celeron was doing the same to AMD’s previous flagship, the K6-III.  What goes around, comes around, and it’s definitely making AMD one happy camper. 

At 1024 x 768 x 32, the performance standings remain relatively unchanged, illustrating that MDK2 isn’t as quick to limit performance as the Quake III Arena tests we just finished analyzing. 

We have always known UnrealTournament to be a very texture intensive game.  This translates into a larger dependency on having a fast L2 cache, as well as a fast memory bus.  The Celeron holds the advantage of a 766MHz L2 cache, versus the Pentium III 500E’s 500MHz cache, however it loses in that it is crippled by a 66MHz memory bus. 

When the smoke clears, the Celeron 766 is no faster than a Pentium III 500E. 

The same holds true as we increase the resolution, these two contenders aren’t going anywhere. 

Expendable, another very memory bandwidth/L2 cache intensive game leaves the Celeron at 766MHz just barely faster than the Pentium III at 500MHz. 

How does the Celeron 766 measure up to the competition as an entry-level professional OpenGL workstation?  Unfortunately not too well at all.  Being severely crippled by its FSB and memory bus frequency, the Celeron is far from a good solution for users that find themselves depending on high performance in applications like Pro/Engineer, AutoCAD, Microstation, etc…

For those users, it may very well be worth the added cost to upgrade your entire system to a relatively inexpensive Duron or Athlon workstation since high performance is definitely more valuable when your job depends on it. 

Final Words

What should you take away from all of this?  For starters, the Celeron, now at 766MHz, is mainly an upgrade chip, since there is clearly a better option if you are building a new system from scratch.  For those of you that need us to spell it out, the Duron is obviously a much better alternative not to mention that, clock for clock, it is cheaper than the Celeron.  A Socket-A platform carries quite a bit of upgrade potential as you will be able to drop an Athlon in there later on, and if you pick up a DDR board or a motherboard with true 266MHz FSB support (based on the VIA KT133A chipset) you will be able to use future Athlons as well.

But let’s talk about where the Celeron 766 would fit, and where you would be better off sticking to an older Pentium III.  As we discovered, business/content creation applications do get a small boost courtesy of the Celeron’s clock speed.  Remember that at 766MHz, both of the Celeron’s caches are running at 766MHz as well.  For applications that can reside in the processor’s L2 cache, a 766MHz clock speed is definitely desirable. 

If you happen to own an older BX or even a LX based motherboard, with something slower than a Pentium III 500E (especially if you have an older Katmai Pentium III, without any on-die L2 cache), the Celeron makes perfect sense as an upgrade.  Again, provided that your usage patterns are best represented as a general purpose or home office/business user. 

Once you start getting above the Pentium III 500E level, the Celeron 766 begins to lose a lot of its appeal.  For those users, you’re probably better off either upgrading to another Pentium III or waiting until the 100MHz FSB Celerons come out and picking up one of them. 

Gamers will want to stay away from the Celeron 766 as it won’t be your best option at all.  The only gamer that will want to consider it will be the user that has a Pentium III 450, without any on-die L2 cache, and is looking for a marginal performance boost.  Even then, don’t expect too much out of the Celeron 766 as a great gamer’s CPU. 

In the end, we’re glad to see that the Celeron will be gaining a 100MHz FSB next year, it will give it more of a chance against the Duron, and although it will lose the performance battle, it will provide current BX motherboard owners a cheap upgrade path which is always a good thing.

What we would like to see on the AMD side is larger presence of the Duron processor in the retail market. The processor is such a good, high performing design, it is a shame that it is a well kept secret among hardware enthusiasts alone. Hopefully with a solid value platform to run on next year, the Duron will be able to provide system integrators and OEMs with exactly what they need as the keystone to a high performing sub $1000 system.