When we were first introduced to AMD's Duron, it renewed our faith in the recent axiom: value doesn't have to mean slow. Originally introduced to us by Intel's Celeron, the idea of "value" CPUs coming dangerously close in performance to their counterparts in the performance desktop market segment has been one that we have hung on to ever since.

One thing we learned from our experience with the Duron was that because of its very close ties to the Athlon (same FSB and same memory bus), the Duron has no problem competing with the Celeron in terms of performance. In fact, it took an overclocked Celeron running at 850MHz, a clock speed that the Celeron won't hit until sometime next year, to come close to the performance of a normally clocked Duron running between 600 and 700MHz.

The reason Intel is falling behind, performance-wise, in the value market segment is because the Celeron is being horribly crippled by its 66MHz FSB and 66MHz memory bus. Giving the Celeron 100MHz FSB support would definitely help it perform closer to the levels that we saw with the first Durons, however it still lacks something, which is hindering its performance.

As we discovered in our original review of the Celeron, there were obviously a number of factors that contributed to the Celeron performing slower than the Pentium III on a clock for clock basis. The most noticeable reason was the fact that the Celeron only ran at a 66MHz FSB and was thus limited to a 66MHz memory bus (on BX boards), but even when overclocked using a 100MHz FSB setting, the Celeron was still coming in well below an equivalently clocked 100MHz FSB Pentium III.

It turns out that the missing link was the L2 cache subsystem of the Celeron. As you all know, the Celeron features exactly 1/2 the L2 cache of the Pentium III (128KB vs. 256KB). The way Intel apparently produces the Celeron is by taking the Pentium IIIs that may have a number of bad cache blocks during their production and effectively disabling half the cache.

This explains why the Celeron has an identical transistor count to the Pentium III in spite of having half the L2 cache (cache increases transistor count dramatically) and it also explains why the Celeron's L2 cache is only a 4-way set associative L2 cache while the Pentium III features an 8-way set associative L2 cache.

This difference in cache mapping does create a noticeable performance delta between the two chips; for a more thorough explanation on the reasons behind this, take a look at our comparison of cache mapping techniques in our AMD Thunderbird Review.

With those factors all hurting the performance of the Celeron, it isn't surprising that the Duron is able to dominate it in virtually all aspects. And today, the Duron even surpasses the Celeron in terms of clock speed as well. Now running at 750MHz, the Duron has an additional selling point that OEMs and system integrators can push: clock speed advantage. While most AnandTech readers know that operating frequency is only as important as the CPU that's running at that speed, your average buyer is only going to compare various systems based on a handful of specifications, including clock speed.

The Chip

Since the first Duron 700 we reviewed had no problem overclocking all the way up to 900MHz, there is no reason that AMD would have to make any dramatic changes to their Duron core in order to allow it to hit 750MHz.

· 25 million transistor 0.18-micron Duron core with Aluminum interconnects
· 750MHz clock speed - 7.5x clock multiplier
· 128KB on-die L1 cache running at core speed
· 64-bit exclusive 64KB on-die L2 cache running at clock speed
· 462-pin Socket A EV6 CPU interface running at 100MHz DDR (effectively 200MHz)
· 1.60v core voltage

Our original Duron sample (700MHz) was a 1.60v part, but it seems as if AMD is shipping Durons rated at 1.5v as well as 1.6v. We received an OEM Duron 700 that had a 1.5v core voltage setting, but our most recent 750MHz sample was rated at 1.60v.

With the Thunderbird being rated at between 1.70 - 1.75v, it makes sense that as the Duron increases in clock speed, its core voltage will see a similar increase until it reaches the level of the Thunderbird since it is based upon the same core but with less 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 - 750 MHz	500 - 700 MHz	750 - 1100 MHz		450 - 600 MHz	500 - 1000 MHz	300 - 533 MHz	533 - 600 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)	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

The Duron processors are still coming out of AMD's Fab25 facility in Austin, Texas since they are using Aluminum interconnects versus their high speed Copper counterparts being produced in the Fab30 plant in Dresden. It seems like Copper may be necessary for the Thunderbird/Duron core to reliably reach 1GHz and beyond, which could explain why we're only seeing 1GHz+ CPUs coming out of Dresden.

For now, all Durons are still using Aluminum interconnects and are thus being produced in AMD's Fab25 plant.

Overclocking

Overclocking the Duron is governed by the same limitations as the Thunderbird; for more information on adjusting the clock multiplier of the Duron, read our 1.1GHz Thunderbird Review and our AMD Overclocking Investigation.

There have still been reports of Durons being shipped without the multiplier pins present on the CPU. We are still waiting to come across some of these CPUs, and when we do, we will keep you all updated on the situation.

Test Notes

• Due to the recall of the processor, we have stopped including 1.13GHz Pentium III scores.
• In order to remain focused on the Duron price range, we have only included CPUs clocked at 900MHz and lower in this comparison
• For purposes of simplifying the graphs, the 133MHz FSB Pentium III scores are taken on an i815E platform. The performance of the Pentium III on the i815E is close enough to the performance of the same processor on a VIA 133A or BX133 platform that it warrants the removal of the other two platforms from the comparison.
• The RC5 benchmark will return in our next major CPU review with updated performance scores using the latest RC5 core.

The Test

The first thing to keep in mind as you look at this performance comparison is that the Duron's chief competitor isn't present in a non-overclocked form. Intel's Celeron, currently available at up to 700MHz (66MHz FSB), doesn't even outperform the "slowest" contender here, the Athlon 600. The Celeron 700 scores a 123 under SYSMark 2000, and the 566MHz entry level part comes in at 111.

The entry-level Duron 600 is already as fast as the original K75 based Athlon 600, and the new 750MHz part proves to be around 4% slower than the Thunderbird 700. On a clock for clock basis, the Duron is around 8 - 10% slower than the Thunderbird with its larger L2 cache.

If you're making the comparison to the older K75 based Athlon's with their off-die L2 cache, the Duron is around 2 - 5% slower on a clock for clock basis.

In spite of the fact that the Duron is "only" delivering 90 - 95% of the performance of its older brothers, keep in mind that the Duron is currently priced at around 30% less than an equivalently clocked Thunderbird.

Compared to Intel's Celeron, the Duron offers clearly superior performance, with a competitive price. The only chance the Celeron has of coming out ahead of the Duron is if you overclock the processor using the 100MHz FSB, but if you overclock the Duron the balance shifts yet again.

As a gaming platform, the Duron is another success. While it does not provide performance equal to that of its two older brothers, for its intended market segment, the Duron helps to redefine the phrase "the most bang for your buck."

Intel's Celeron is also in this category of a value processor that has no problem providing a decent level of performance, and prior to the introduction of the Duron it was clearly the best choice for those that didn't have the budget to spend on one of the faster CPUs. Unfortunately for Intel, the Duron has the Celeron beat in almost all categories.

The only advantage the Celeron currently holds over the Duron is that it has a very affordable motherboard platform with integrated video that OEMs and system integrators can use in very low-cost systems.

For hard core gamers the last thing you want is a platform with an integrated video solution, but if gaming isn't your forté then the Duron's lack of a low-cost integrated platform is a major downside.

As we've said time and time again, before considering a new CPU you always have to first take into account any other bottlenecks that may be present in your system. In this case, the video card is obviously limiting the performance of the system as a whole.

The GeForce2 GTS used in our test system is exhibiting signs of a memory bandwidth limitation thus making AMD's newly released Duron 750 perform just as well as a Pentium III 800 or even an AMD Athlon (Thunderbird) running at 1.1GHz.

Testing Quake III at 640 x 480 is an ideal way to examine the power of a CPU using the game as a benchmark, however running it at 1024 x 768 x 32 is much more of a real world test of overall system power for use in games.

UnrealTournament makes for an excellent CPU benchmark since it scales so well with faster CPUs. Under UT the Duron is around 7% slower than the original K75 Athlon and 14% slower than the newer Thunderbird.

At 750MHz the Duron is just barely faster than the overclocked Celeron 850 and is only a step away from the K75 Athlon 700. It will take a 900MHz Celeron to beat the new Duron 750, and a Celeron clocked at 950MHz - 1GHz in order to compete with a Duron clocked at 800MHz. Intel's roadmap doesn't call for a Celeron faster than 800MHz until sometime next year, so it will be a while until the Duron is threatened by anything coming out of Intel's labs.

The Duron will most likely end up competing with its replacement from AMD before the Celeron begins to gain ground on it.

Unlike what we saw under Quake III, UnrealTournament is an example of a game that isn't memory bandwidth limited (on the graphics card) at 1024 x 768 x 32. In all actuality, quite a few of the games available today (especially those that aren't based on the Quake III Arena engine) aren't exposing the GeForce2 GTS' memory bandwidth limitations at this resolution.

In cases like this, a balance between CPU power and a fast video card is necessary. The Duron paired up with a GeForce2 MX would probably be the most cost effective solution that would offer an incredible amount of performance for the money. With the Duron priced under $100 for the chip, and the GeForce2 MX setting you back around $120, your motherboard will probably be one of the most expensive parts inside your case.

Expendable is our third and final game benchmark under Windows 98SE and the Duron exhibits behavior similar to what we've seen under Quake III Arena and UnrealTournament, to a greater degree.

The Duron is around 10% slower than an equivalently clocked K75 based Athlon and around 12% slower than a Thunderbird. The very small performance difference between the K75 Athlon and the Thunderbird indicates that Expendable is the type of game that has less of a dependency on a fast L2 cache, and simply requires a semi-large one.

The Duron's 64KB L2 cache and its 128KB L1 cache don't seem to provide the type of storage Expendable requires for its most frequently used data, resulting in a noticeable performance drop when compared to both the Thunderbird and the K75 Athlon.

In support of the aforementioned conclusion is the Celeron's performance under the Expendable test. At 850MHz the Celeron is just barely able to keep up with the Pentium III EB at 600MHz. The fact that the Celeron's 128KB L2 cache is only a 4-way set associative L2 cache (compared to the 16-way set associative exclusive L2 cache of the Duron) most likely also plays a part in the processor's poor performance (with respect to the Pentium III clock for clock) in this test.

An advantage the Duron holds over the Celeron here is that in spite of its 'small' L2 cache it has a faster path to memory (133MHz - courtesy of the KT133 chipset) meaning that all data that can't fit within the Duron's L1 and L2 caches has a 33% faster path to the system memory than on the Celeron's BX platform. Even on the VIA 133A platform with its memory bus running at 133MHz, the Celeron at 850MHz isn't able to outperform the Pentium III 600EB.

As the focus of our attention shifts from gaming under a consumer OS to productivity applications under a professional platform, the Duron has no problem making the adjustment.

Content Creation Winstone 2000 isn't a very taxing benchmark at all, not anymore at least. Today's processors have no problem handling the multiple applications the benchmark tosses their way. For those whose usage patterns are characterized by CC Winstone 2000 (multiple content creation applications open i.e. office applications, web browsers, image editors, etc...), the Duron proves to be just as good as its more expensive siblings.

The Thunderbird is around 5% faster than the Duron in CC Winstone 2000 and the original K75 Athlon is barely 2% faster than the "value" priced Duron, once again on a clock for clock basis.

The overclocked Celeron 850 is performing at around the same levels as Duron 700, however none of the regularly clocked Celerons can even outperform the lowest scoring Pentium III 600 and thus don't even appear on the comparison charts.

As a "high-end" processor, the Duron keeps up with the best of them. While the 750MHz Duron is not able to outperform the Thunderbird at 700MHz, for the price, being outperformed by 3% isn't too bad.

In general, the Duron comes out around 5% under an equivalently clocked Thunderbird in the High End Winstone 99 tests. It is obvious that the tests prefer the faster L2 cache of the Duron to the larger L2 cache of the Athlon, as the latter only seems to outperform the Duron by 1 - 2% on a clock for clock basis. The Thunderbird, having the best of both worlds, a fairly large L2 cache as well as a fast one, manages to pull ahead of both of its relatives.

Quake III Arena under Windows 2000 is very similar to what we noticed under Windows 98SE, the only difference being that there is a slight performance boost courtesy of Windows 2000 combined with NVIDIA's solid Windows 2000 drivers.

Limited by the GeForce2 GTS' memory bandwidth, the Duron 750 performs just as well as an Athlon 900.

As is obvious by the benchmarks, the Awadvs-03 viewset is a benchmark of raw CPU power and isn't affected by L2 cache speed as much as it is by L2 cache size.

The Duron makes it very clear that value doesn't mean slow in this situation, something that even the overclocked Celeron can't do as is evident by its placement at the very bottom of the chart.

The performance under the DRV-06 viewset seems to be more dependent on video card and platform than the CPU, explaining the relatively small differences between CPUs in the benchmark.

The DX-05 viewset is another case in which the 200MHz FSB (100MHz DDR) of the Athlon combined with the complexity of its core helps achieve the performance advantage over the Pentium III and the Celeron.

Final Words

For the first time since the introduction of Intel's Celeron, AMD has a value segment processor that can not only outperform the Celeron across the board but is also shipping at higher clock frequencies. This is a very big win for AMD, not only does Intel have to concentrate on regaining control of the lost ground their Pentium III was unable to hang on to they also have to worry about competing with a very powerful chip that is currently carrying a two digit price tag.

The Celeron will be getting an upgrade to the 100MHz FSB sometime next year, but until then, even a 1GHz Celeron will barely outperform a 800MHz Duron, not to mention the difficulty it would have matching a 1GHz Duron in performance.

In spite of this, the Duron still has problems making its way into systems. For AnandTech readers, building a Duron system isn't a problem at all. You find a place to purchase the processor and pick a vendor to get your motherboard from and you're already half way there. For OEMs and system integrators however, the process isn't that easy.

While we may not like the i810E chipset, it has sold quite a few Celeron processors simply because it is a cost effective motherboard platform that features integrated audio and video.

With a CPU as cheap as the Duron, it is almost silly to expect that a manufacturer would stick a Duron in a motherboard that costs twice as much as the CPU unless that board had some serious levels of integration. VIA's KT133 chipset with integrated video, the KM133, is long overdue and every day that goes by without motherboards based on the chipset is another day that the Duron suffers from the lack of a home. Whether it is ALi, SiS or VIA someone needs to release a low-cost chipset for the Duron and it needs to happen soon.

Until then, the Duron is still a very viable option for individual users that would rather not be stuck with integrated graphics and audio.

When it comes to productivity, office and content creation applications, the Duron is around 10% slower than the Thunderbird and 5% slower than the older K75 Athlons on a clock for clock basis.

For gamers, a Duron will leave you with around 85 - 90% of the performance of a Thunderbird, and as long as you have a powerful enough video card you shouldn't find yourself in a situation where your gaming performance is suffering at all.

In the high-end arena, depending on the applications you're using, the Duron can offer performance equal to that of a Thunderbird (because of the fact that they share the same core as well as the same bandwidth friendly EV6 bus) while in other situations the performance difference can be much more noticeable.

In the arena that the Duron was originally intended to compete in, the value segment, the Duron does a wonderful job, there's no question about that. It's because it does such a wonderful job that it is forced to comparisons with its more expensive brothers. But isn't that how it always is when you're the youngest in a family?