Original Link: http://www.anandtech.com/show/541
Last August the market had its eyes fixed upon AMD's do-or-die creation, the Athlon. The success of this "new" processor would either bring about a dramatic change in the desktop x86 microprocessor industry or simply add to the list of AMD processors that Intel's Pentium III was capable of outperforming.
Luckily for AMD, the Athlon turned out to be quite a success as it completely dominated the 600MHz Pentium III that was available at the time of its launch. But history would repeat itself as it took no longer than three months for Intel to respond with their Pentium III based on the new "Coppermine" core, which included a full 256KB of on-die L2 cache running at clock speed.
The performance delta that existed between the Athlon and the Pentium III (Coppermine) seemed to grow over time as the Athlon's performance was being hindered by the fact that the Athlon had an L2 cache running at a fraction of its clock speed (it ended up that the L2 cache was always running at or below 350MHz) in comparison to the Pentium III, whose smaller 256KB L2 cache was closing in on the 1GHz mark.
Although the Athlon did have a very large L1 cache advantage over the Pentium III (128KB vs. 32KB), only half of that 128KB is a data cache and that 64KB was not enough to compete with the Coppermine's 256KB on-die L2 cache.
It was obvious that AMD needed to move to a new version of the Athlon's core that featured an on-die L2 cache, just like Intel had done with the transition from their Katmai to Coppermine cores. That transition would also allow AMD the opportunity to drive the costs of their CPUs even lower since they could do away with the Slot-A packaging that would become useless with all of the CPU's cache on-die – not to mention the performance boost gained by taking that L2 cache off of the processor card and moving it to its full speed location on the CPU's die itself.
We originally expected the first 1GHz CPU from AMD to feature an on-die L2 cache running at clock speed; unfortunately, we were considerably disappointed as every single Athlon released this year has featured the same old off-die L2 cache running at as low as 1/3 of the clock speed of the core itself (in the case of the 1GHz Athlon).
Things are about to change. Intel's Willamette is still a few months off, but before Intel's true architectural successor to the Pentium Pro comes along, AMD will have the chance to regain the lead that made the Athlon so popular in the first place. Nine months ago we concluded our review of the AMD Athlon with a chart illustrating the potential future of the Athlon. That future is almost upon us, and it's time to take another look at the evolving Athlon core to set a few things straight about AMD's role in the desktop x86 microprocessor industry, at least for the rest of this year.
of the Athlon - August 1999
On April 27, 2000, AMD officially announced the name of their upcoming low-end processor based on a derivative of the Athlon core. The AMD Duron, as it will be sold as, will feature what we've all come to know as the Spitfire core.
The Duron will be packaged as a 426-pin Socket-A part and will feature clock speeds comparable to that of Intel's new Celeron, currently residing in the 533 – 600MHz range, though it is expected to hit 667MHz by June or July. The new Celeron will be the Duron's true competitor and you can expect to see the Duron priced competitively against Intel's Celeron on a clock for clock basis.
Fortunately, the Duron won't be as crippled as the new Celeron, since it will feature the same 100MHz DDR EV6 Bus (effectively 200MHz) that brought the Athlon so much attention upon its release. As we proved in our review of Intel's new Celeron, its 66MHz FSB and 66MHz memory bus is definitely holding back the performance of the CPU, and unless overclocked, this will give the Duron an easy performance advantage over the Celeron.
The Duron will feature the same 128KB L1 cache from the Athlon in addition to a full speed on-die L2 cache. While AMD isn't ready to disclose exactly how large that L2 cache is, they are willing to say that the Duron will feature more total on-die cache (L1 + L2) than the Celeron (which features 32KB L1 + 128KB L2). At the same time, the Duron will obviously have less L2 cache than the upcoming Thunderbird part (256KB), so you can safely guess as to what the actual cache size of the Duron will be.
The reason AMD is being so secretive about the L2 cache size of the Duron is because they want to be able to explain exactly why they chose that particular L2 cache size for the processor without having people draw their own conclusions without any factual basis for it.
From a performance aspect, the Duron will be able to hold its own in comparison to Intel's Celeron and will provide some pretty powerful performance considering it is a "low-cost" part. Contrary to popular belief, the Duron will not outperform the current Athlons on a clock for clock basis, but they will come very close to the full performance of the current Athlons.
Given that the performance of the Duron is very close to that of the present day Athlon, if you look at a graph (above) comparing an Athlon 600 to a Celeron 600, you can easily see how the Duron will have no problem dominating Intel's Celeron. The remaining question is what happens when you overclock the Celeron to 850MHz or beyond; only then will the Duron have problems keeping up because it currently takes an Athlon 800 or an 850 to match or beat the performance of a Celeron overclocked to 850MHz on a BX board.
AMD's New Athlon – Thunderbird
The Thunderbird has been one of the most talked about processors cores simply because it was thought to be exactly what AMD needed in order to regain the performance advantage over the Pentium III.
The Thunderbird, like the Duron, will take the current Athlon and move its L2 cache on-die. For cost and performance reasons, the Thunderbird core will feature a total of 256KB L2 cache running at clock speed versus the 512KB of L2 cache running at between 1/2 and 1/3 clock speed on the current Athlons (K75 core). Combine this with the Athlon's 128KB L1 cache and AMD once again has the cache advantage over Intel.
Since the Thunderbird will be moving the L2 cache on-die, there is no longer a need for the Slot-A processor card and we will thus see the Thunderbird available in a 426-pin Socket-A package.
The biggest disappointment here is that, although we all remember this slide below indicating that the Thunderbird would be available in a Slot-A package as well as a Socket-A package, AMD wants to move to Socket-A very quickly and will only be supplying OEMs with Slot-A Thunderbirds.
According to AMD, you shouldn't be able to go out and purchase a Slot-A Thunderbird, meaning that you will most likely need to get a new Socket-A motherboard for use with the Thunderbird. While it is inevitable that there will be some Slot-A Thunderbirds leaked from the OEM channels to on-line vendors, for the most part you can expect the Thunderbird to be a Socket-A solution only.
AMD is pretty far behind with the transition to the cheaper Socket-A packaging for their processors and they want to make the move as quickly as possible. If you remember, Intel took quite a while to make the transition to FC-PGA 370, and even today, you can find a good number of Slot-1 Pentium IIIs. Since AMD didn't start the transition at the end of last year like Intel, they have to act very quickly in making this move. This will definitely upset the many Athlon users with Slot-A motherboards that will be relatively useless when the Thunderbird hits the streets this June.
Upon its release, the Thunderbird will replace the K75 core as the Athlon processor, so when June rolls around and the new Thunderbird core is shipping in volume, the processors that the Thunderbird core will be found in will still be called the Athlon. The way you'll differentiate between the new Athlon and the old Athlon will be because only the new Athlon (Thunderbird) will be available in a Socket-A package while the older Athlon (K75) will be solely a Slot-A part.
The performance of the new Athlon (Thunderbird) will definitely be greater than the Pentium III (Coppermine) on a clock for clock basis, although it may not be able to claim the same against the upcoming Willamette.
The third derivative of the Athlon core will be the Mustang, which will be out sometime in Q3/Q4 2000.
The Mustang core will offer some tweaks and enhancements over the Thunderbird core, including support for larger on-die L2 cache sizes (~1MB) as well as AMD's PowerNow! technology. AMD has not disclosed any of the other feature enhancements as they are concentrating on the Thunderbird/Duron parts right now. However, closer to the release date of the Mustang, you'll hear more about it.
The Mustang will actually be quite a bit more than simply the server/workstation Athlon part; the high end market will simply be one area that the Mustang is geared towards.
The Mustang will eventually find itself in every Athlon derived processor after its release, starting off as the core for the Athlon Ultra (High End server/workstation part), but by varying the amount of on-die L2 cache, the Mustang core will find its way in desktop Athlon parts as well as in mobile Athlon solutions.
This means that the Thunderbird will be fairly short-lived if it is to be replaced by the Mustang as the core of choice for the Athlon later this year, and it also means that AMD is going to be able to compete with Intel much more seriously than they have in the past in the mobile market with the mobile Mustang.
The mobile Mustang, going under the code name Corvette, will feature a smaller L2 cache than the desktop Athlon (Mustang) part and a lower voltage, which makes sense since it will be a mobile part.
Athlon goes Mobile
The Corvette will put AMD's PowerNow! technology to good use. You may remember PowerNow! technology from last Comdex when we referred to it as AMD's Gemini technology, which was talked about around the same time as Intel released their first mobile Pentium III processors with Speedstep. AMD really wants to illustrate that their PowerNow! technology is dramatically different from Intel's Speedstep, and it actually is.
Intel's Speedstep technology allows a Speedstep enhanced processor to switch clock speeds/voltages when plugged into a wall versus when running off of battery power alone. For example, Intel's mobile Pentium III 650 with Speedstep technology runs at a core voltage of 1.60v when plugged into a wall outlet (i.e. not running off of battery power). When the laptop switches over to battery power, the clock speed drops from 650MHz down to 500MHz while pulling the core voltage down to 1.35v in order to conserve power.
AMD's PowerNow! takes this one step further. Instead of simply decreasing the clock speed and the voltage of the CPU when running off of battery power, the PowerNow! technology in combination with the motherboard's BIOS allows for the dynamic adjustment of clock speed/voltage of the CPU during actual program usage.
For example, say we have a 600MHz mobile AMD processor equipped with the PowerNow! Technology. Upon launching a program such as MS Word, the processor will most likely operate at close to its full speed, but immediately after the loading process is complete, the CPU will drop to a slower clock speed and lower voltage. During this time, if you decide to start up another application or begin doing some very CPU intensive calculations (ok, maybe not in Word) the PowerNow! technology will increase the CPU's operating frequency in order to compensate.
We will see PowerNow! used in AMD based notebooks this summer, but these notebooks will be using K6-2+ or K6-3+ processors, not the Corvette.
Chipset & Motherboard Support
If the fact that the Thunderbird isn't really going to be available in a Slot-A package got you, you're going to love the chipset/motherboard support for the new Thunderbird & Duron processors.
The Duron will obviously only run on a Socket-A motherboard. The only two chipsets that will officially support the Socket-A platform are the AMD 750 and the upcoming VIA KZ133. The KZ133 is nothing more than the KX133 with support for Socket-A. Once again, the KZ133 is nothing more than the KX133 with support for Socket-A.
This means that there is a possibility that the KX133 chipset may not work with Socket-A CPUs, which is quite unusual because AMD is claiming that their AMD 750 will work just fine with them. At the same time VIA will be pushing for the Socket-A KZ133 chipset, so it makes sense that they wouldn't want to push the idea of Socket-A backwards compatibility with their KX133 platform.
The Athlon (Thunderbird), as we mentioned earlier, will primarily be a Socket-A solution and will thus be limited to the same chipset solutions as the Duron. However, we also mentioned that AMD will be producing Slot-A Athlon (Thunderbird) processors for OEMs that have AMD 750 based designs and aren't willing to qualify a new setup in order to move to Socket-A.
So there is a chance that current AMD 750 based motherboards will be able to run the new Athlon (Thunderbird) processors with the proper BIOS updates and provided that you can actually get your hands on a Slot-A Athlon (Thunderbird). There is no word as to whether or not KX133 based motherboards will be able to work with these processors, although it doesn't really make sense if they weren't able to work on them unless VIA completely forgot to incorporate a portion of the AMD 750's design into the KX133.
Update 05/04/2000: We just received the following note from AMD's Drew Prairie:
"...the only thing I thought i
would point out is the part where you say kx133 boards likely won't support Thunderbird. They won't. Our plan is to only make the Slot A TBird part available to our OEM partners who have requested a Slot A TBird part and plan on using it in their current systems featuring the AMD-750 chipset."
The idea of a Socket-A to Slot-A converter (à la the current Socket-370 to Slot-1 converters) has come up more than once over here, and we decided to ask AMD about it. Officially, AMD is neither endorsing nor supporting such a converter card although they wouldn't deny that it could be done. The only reason Intel is supporting a Socket-370 to Slot-1 converter is to ease the transition to FC-PGA 370, but we have already established that AMD isn't looking to make this a smooth transition to Socket-A at all. We may see Socket-A to Slot-A converter cards in the future, but it is very clear that AMD wants to make this move to Socket-A as quickly as possible, allowing no time for any transitional tools such as a converter card.
The assumption has always been that, upon the release of the Thunderbird, everyone would be running motherboard platforms with DDR SDRAM support, but unfortunately, that won't be the case.
Remember that we just said that the only two available platforms for Socket-A processors at the launch of the Duron/Thunderbird will be the AMD 750 and the VIA KZ133. We already know that the AMD 750 doesn't support DDR SDRAM, and as we made it a point to mention twice above, the KZ133 is nothing more than the KX133 with support for Socket-A.
This translates into no DDR SDRAM platforms available for the Athlon at the launch of the Thunderbird – don't kill the messenger.
It won't be until the release of the Mustang in Q3/Q4 that we will see the introduction of the AMD 760 chipset with DDR SDRAM support, but at the same time, we should see the AMD 760MP (also known as the AMD 770), which will be the first multiprocessor capable Athlon chipset allowing for 2-way Athlon configurations.
At some point during that time, ALi and SiS should be releasing their Athlon chipsets, but realistically, don't expect either of those two companies to bring anything to the table that you will want to take over the AMD or VIA chipsets.
It won't be long before AMD regains the performance lead in both the low end and the high end desktop markets. Although the current Athlon is definitely performing quite well, its power is being severely limited by its slow L2 cache.
The release of the Thunderbird & Duron will help restore power to the AMD name, but the real thing to wait for will be the Mustang, which in the Athlon form will be the competitor to Intel's Willamette.
The lack of any motherboard solutions with DDR SDRAM support at the release of the Thunderbird will definitely be a downside, but so will the relative lack of any Slot-A Thunderbird parts. The reality here is that if you want to stay on top of the AMD x86 processor market, you'll probably have to upgrade twice, once to the Thunderbird and once again to the Mustang unless you can find it in you to wait for the latter.
The introduction of the Mustang should restore some balance to AMD's product structure, with clearly defined high end, mainstream and low end processor solutions falling under the Athlon Ultra, Athlon and Duron product families.
For the present day consumer, if you're buying an Athlon motherboard today with hopes of using it with tomorrow's Athlon processors, think again. It's an unfortunate reality, but it's better to know now than find out in two months when you're kicking yourself for picking up that new KX133 board just a few weeks earlier.
The CPU market is going to be a very interesting place come this fall, although AMD will get a bit of a head start this summer with the release of their Athlon (Thunderbird) and Duron parts.