18 years ago
published this review on the AMD Duron by Dan Mepham :
AMD Duron Review
Introduction
In the past, business hadn’t been too great for AMD. Ever since the introduction of Intel’s Pentium, AMD found itself playing catch-up, first with the K5, then the K6, and K6-2, none of which were able to achieve the level of performance of their Pentium counterparts. The only AMD processor up to that point that was capable of giving the equivalent Intel processor a run for its money was the K6-III, which proved to be too expensive for the masses.
The Athlon changed all that. Not only was an Athlon cheaper than an equivalent Pentium III, but it performed almost identically, if not better. However, while AMD finally started to get a grip on the high-end market, the low-end market was quietly being stolen away from them, thanks primarily to Intel’s Celeron. While the K6 series was still the best option for upgrading Socket-7 users, the Celeron was clearly the best choice for new value/OEM systems.
It became very clear that AMD needed to do something to regain low-end market share. While the Athlon actually had the price point to be considered a competitor to the Celeron in some cases, associating the Athlon name with low-end processors was a route AMD preferred not to take. Another solution was needed.
That solution arrived a few short weeks ago. AMD introduced a low-end variant of the Athlon called the Duron to compete with Intel’s Celeron. Today we bring you a long-overdue review of AMD’s new Duron. Read on, as we see how it stacks up against Intel’s tried and tested Celeron.
Basic Overview
Features:
600, 650, 700 MHz clock.
462-pin Socket package
0.18 µm die process
100 mm2 die
25 million transistors
200 MHz (100 MHz DDR) EV6 Bus
128 KB First Level Cache
64 KB Full-speed Second Level Cache
The Duron, unlike the Athlon, is available only in Socket-style packaging. The primary reason for this is cost, as socketed processors are physically less expensive to produce than Slot-style ones. Since the Duron features on-die L2 cache, AMD is free from reliance on external L2 SRAMs, and is able to move back to the less-expensive socket-style interface. With the Duron and Athlon (Thunderbird), AMD has introduced a new 462-pin socket called Socket-A, which is mechanically incompatible with Socket-370, to insure a Celeron isn’t accidentally placed on a Duron motherboard, or vice versa. While Socket-370 heatsinks can be clipped onto a Socket-A interface, the Duron/Athlon and Celeron have different clip and pressure requirements, so a heatsink designed for Socket-A processors, not Socket-370 processors, is essential.
The Duron, like recent Athlons and Coppermine Pentium IIIs and Celerons, is fabricated on a 0.18 micron die process. The smaller die size allows for faster switching times, as well as less power dissipation, allowing processors to be clocked much higher. AMD has had extremely high yields of processors at the 1 GHz barrier, although the Duron is currently only available in 600, 650 and 700 MHz speed grades. This places the Duron in the 600-700 MHz range, while the Athlon (Thunderbird) is available in the 700-1000 MHz range, in order to minimize competition between the two. Faster variants of the Duron will arrive in time, and availability should be high given AMD’s current yields.
The Duron features some 25 million transistors, of which 21-22 million are devoted to logic and the L1 area, while the remaining 3-4 million compose the 64 KB L2 storage area.
Cache Structure
The Duron’s cache structure is significantly different from that of the original Athlon, Celeron, or Pentium III, and by exploring those differences, we can help explain the performance of each.
Like the original Athlon and the new Thunderbird, the Duron features 128 KB of First Level cache (64 KB Data, 64 KB Instruction). This 128 KB of L1 cache is 2-way set associative, like the Athlon’s. In contrast, Intel’s Celeron features only 32 KB of L1 cache (16 KB Data, 16 KB Instruction), but is 4-way set associative. The increase in set associativity translates into a higher hit rate, but the Celeron’s smaller L1 size serves to negate any real performance increase.
Where the Duron differs from the new Thunderbird is in its L2 cache. The Duron features only 64 KB of Second Level cache, unlike the Thunderbird's 256 KB. The Duron’s L2 is 16-way set associative, as is the Thunderbird’s. In contrast, Intel’s Celeron features a larger 128 KB L2 cache, but is only 4-way associative. As with the L1 cache, these differences should partially cancel out, as a smaller but easier to search cache should have a similar hit-rate to a larger but harder to search cache. In the case of both the Duron and the Celeron, the L2 cache is on-die, and runs at full clock speed.
Unlike the Celeron, however, the Duron’s cache is exclusive. In an exclusive cache structure, data currently in the L1 cache does not need to be mirrored in the L2 cache. In essence, the Duron can have completely different data in its L1 and L2 caches. In contrast, the Celeron features an inclusive cache, meaning that data in the L1 must be mirrored in the L2 cache. In the case of the Celeron, the 32 KB of instructions/data in the L1 must be mirrored in the L2 cache, essentially reducing the L2 storage available to 96 KB.
The final difference in cache structure between the Duron and Celeron lies in the L2 bus width. While the Duron features the same 64 bit bus as the original Athlon, new Coppermine Celerons feature a much wider 256 bit bus. Essentially, this gives the Celeron four times as much L2 bandwidth as the Duron--a definite advantage. It is not exactly clear why AMD chose to stick with a 64 bit L2 bus for its Athlon and Duron, while Intel moved to a 256 bit path for Coppermines. AMD’s engineers were either unable or unwilling to modify the L2 bus width at this point, which certainly would have bolstered the Duron/Athlon’s performance.
Test Setup
Die sizes and cache structures all look good on paper, but the only real indicator of a processor’s potential is performance (and price).
In our benchmarks, we’ve shown the Celeron using a 66 MHz FSB and 100 MHz memory clock, which represents a more realistic environment, as most new systems (133A, i815) are capable of asynchronous memory operation, and feature PC100 memory at least.
We’ve also shown the Duron using a 100 and 133 MHz memory clock. If purchasing from an OEM, the most common configuration would be with PC100 memory, which is why we’ve opted to show 100 MHz memory benchmarks. However, those building on their own will have the option of slightly more expensive PC133 SDRAM, and since the KT133 is fully capable of asynchronous memory operation, we thought it best to include those, too.
Performance - SYSMark & 3DMark
In SYSMark2000, our business application test, the Duron outpaces the Celeron by a significant margin, almost 10%, and even more when paired with PC133 memory.
3DMark2000 paints a similar picture, with the Duron pulling ahead by a greater margin with PC133 memory.
Performance - OpenGL & Direct3D
It doesn't take a PhD to understand these graphs--for a gaming platform, the Duron is the better option, without doubt.
Performance - Professional
The Duron, based on the more-powerful Athlon core, shows superior professional level performance. While neither of these processors would be the first choice for a professional environment, the Duron is certainly the better of the two.
Performance Analysis
The graphs really speak for themselves in this case: at default clock speeds, the Duron is clearly faster than the Celeron in all circumstances. Not marginally faster, but significantly faster, around 25% plus in most cases. When paired with common 2-2-2 PC100 memory, the Duron has no trouble soundly trouncing the Celeron, and when coupled with the somewhat rarer (and more expensive) 2-2-2 PC133 memory, it pulls even farther ahead.
We also expect the Duron to benefit greatly from DDR memory when it becomes available, primarily due to its smaller cache size. The less cache, the more dependent the system will be on memory performance, and so DDR memory should produce a sizeable performance increase for the Duron.
Given the large performance gap between the two, even an overclocked Celeron will have trouble competing with a non-overclocked Duron. Furthermore, our 600 MHz Duron was able to run with complete stability at a whopping 900 MHz, a speed which guarantee it will leave the Celeron in the dust.
Summary
In summarizing, we’d like to take the time to dispel a few common myths.
Myth 1: An overclocked Celeron is faster than a Duron.
Not at all. The Celeron is severely bottlenecked by memory performance, and it would take a Celeron pushing an FSB of 110+ MHz (which is only realistic for the Celeron 533A, and possibly 566) to even approach Duron levels of performance. Besides, comparing an overclocked Celeron to a non-overclocked Duron is hardly a level playing field. The Duron can be overclocked as well, and if our 600 MHz piece was any indication, overclocked very, very well, too.
Myth 2: Celerons are still more overclockable than Durons.
Not really. Granted, Celerons may be slightly easier to overclock in some situations due to their low FSB, and there are currently really only two or three Duron motherboards that allow the multiplier to be changed, but in terms of actual overclocking potential of the processor itself, the Duron may be the better of the two. Early reports show Durons hitting 850-950 MHz with relative ease, which puts them neck-and-neck with, or even faster than, the Celeron in terms of raw MHz.
Myth 3: The Celeron is dead.
Not any time soon. Some sources have observed the Duron’s performance, and declared that the Celeron is dead. This is a bit of a hasty judgement, and not a very accurate one. The Duron is much faster than the Celeron, there can be no denying that, but speed isn’t everything, especially in the low-end market where both the Duron and Celeron are targeted. For example, consider a small OEM. That OEM could purchase an i810e board with on-board sound, video, and even LAN, add an inexpensive Celeron 500, some PC100 memory, and have an extremely cheap and easy system. Someone wishing to do the same with a Duron will likely have to work a bit harder to find a Duron motherboard (and will pay more for it than for an i810e board), and will certainly have to add his/her own video card, network card, and so on. Furthermore, the Celeron is a perfect upgrade path for LX, BX, EX, and many other users, whereas a Duron means buying a new motherboard. It is for reasons like this that we will continue to see Celeron systems for a long time. The Celeron isn’t dead, although it has had the wind abruptly taken from its sails.
A quick look around the internet will show 600 MHz Durons shipping for less than $85-90, whereas 600 MHz Celerons are currently running about $125. Of course, Duron motherboards are typically more expensive than Celeron boards, and so that price advantage is partly negated, at least until Socket-A motherboards become more available.
For overclockers, the Duron looks to be a wonderful value. 900 MHz from an 85-dollar chip is an incredible and very realistic goal given an appropriate motherboard, such as Asus’ A7V or Abit’s KT7. The Duron does dissipate a fair amount of heat, so an appropriate heatsink is a necessity. Again, we’d like to stress that mounting a heatsink on a Socket-A processor is a very delicate task, and should be performed with extreme care.
In summary, AMD sought to produce a solution competitive with Intel’s Celeron, and it's succeeded. The Duron is faster than the Celeron, cheaper than the Celeron, and arguably even more overclockable than the Celeron. The only thing stopping the Duron from being a hit with OEMs is motherboard support, a problem which has plagued AMD for the past year. Should cheap, readily-available, integrated Socket-A motherboards begin to emerge in volume (KM133?), the Celeron will have a real fight on its hands. Let's hope AMD doesn’t have the same problem Intel had 18 months ago when the Celeron 300a started stealing market-share from the high-end Pentium II 450...
Pros:
Benefits from the extremely powerful Athlon core.
Faster than an equivalent Celeron.
Cheaper than an equivalent Celeron.
Highly overclockable.
Cons:
Poor motherboard support compared to Celeron.