Original Link: http://www.anandtech.com/show/1040
Comdex 2002 Wrapup - AMD, NVIDIA & Transmetaby Anand Lal Shimpi on November 23, 2002 3:42 PM EST
- Posted in
- Trade Shows
With the week over, Comdex has come to a close and now that we're all back after a good three days of covering the show we're able to provide you with the rest of our highlights.
For this piece we've got information from AMD, NVIDIA, Transmeta and even a quick look at the new Palm Tungsten T.
AMD - "I'm Impressed"
You would think that for a company in as much trouble as AMD is right now that the mood over there would be significantly degraded. It turns out that the exact opposite is the case and it is quite evident by our meeting today with them. Before our meeting we were quite concerned about AMD (as are many of the motherboard manufacturers) but afterwards we have renewed confidence in their ability to not only survive but do well in this market; just as you can imagine though, their survival and success rests entirely on Hammer - both on the desktop and server sides.
So what did we learn from AMD today? Quite a bit
uPGA Mobile Athlon XPs - Finally
We showed off AMD's uPGA mobile Athlon XP packaging back during Computex and we actually first saw the "new" packaging technology exactly one year ago during Comdex. Luckily for AMD, the uPGA packaged mobile Athlon XPs will begin shipping in the first quarter of next year.
The uPGA socket retains the same electrical characteristics of the current 462-pin socket (Socket-A) for Athlon XP CPUs; it will simply be more compact.
The major benefit of the smaller packaging is that mobile AMD processors can finally be used in thin-and-light notebooks instead of the full sized solutions that are currently available.
It won't be until AMD can reduce power requirements before they can be considered to be real competition for Intel in the mobile market. The idea of a mobile Athlon 64 (Hammer) brings music to our ears, and the ears of some manufacturers. NVIDIA won't be introducing a mobile version of their nForce chipset until Athlon 64 does go mobile for the reasons we've just mentioned.
400MHz FSB & Barton
As we reported earlier, AMD is strongly considering migrating to a 400MHz FSB for Barton. If you ask AMD for their official stance though, it is much like their stance on the 333MHz FSB - "we're evaluating it."
It turns out that Barton will ship with a 333MHz FSB initially, and that we received directly from AMD. Remember that even at the end of 2003, shipments of Athlon XP and Athlon 64 will be an even 50/50 split - so a good deal of the market will be serviced by Barton. In an attempt to keep Barton more competitive, the 400MHz FSB would be implemented but only if chipset and motherboard designs can be reliably produced to support the spec.
The biggest issue you encounter with a 400MHz FSB is maintaining signal integrity in the traces between the CPU and North Bridge. With 200MHz DDR signaling (effectively 400MHz), noise and trace length become even bigger factors thus putting the pressure on motherboard manufacturers to work even harder. The implementation side of things shouldn't be a problem however, as most motherboard manufacturers are working on 200MHz QDR signaling for Intel's upcoming Springdale and Canterwood projects. It'll take some work, and it's too soon to have it in January but 400MHz FSB is a very simple reality to produce.
What all manufacturers, partners or not, are waiting for is AMD's Athlon 64 and Opteron parts. As we mentioned yesterday, with Intel slowing down their clock speed ramp AMD has a great opportunity to make a splash with Athlon 64 on the desktop.
AMD is fairly certain that they will be able to deliver Athlon 64 to the desktop at the end of Q1 2003 or the beginning of Q2 2003. Realistically, it will be closer to the start of Q2 2003 but we'll keep that tight range open for discussion. We reported earlier this week that current Athlon 64 samples are running at 1.4GHz (software and hardware developers just received 1.8GHz samples) and today AMD reassured us that they will have a performance leading part at launch. With the current 1.4GHz CPUs weighing in at around the speed of a 2.2GHz Pentium 4, we'd expect great things from a 2.0GHz processor on a fully tuned platform at launch; luckily AMD will be up against no faster than a 3.20GHz Pentium 4 at launch.
We asked the question of whether AMD expects to see a 64-bit desktop version of Windows at the Athlon 64's launch and their answer was an expected "it would be nice" but clearly unlikely. It was quite clear that Microsoft was never going to have a 64-bit desktop OS ready for the Athlon 64's launch had it been the end of this year, and even as we move into 2003 it will definitely be after Athlon 64 is out before Microsoft makes the move.
As AMD mentioned, there are a good number of betas and preview versions of the 64-bit OS floating around but nothing close to final yet. By far the biggest issue Microsoft is encountering is driver support; unlike the server market where you don't have to worry about much backwards compatibility with older hardware, with a desktop OS Microsoft must make sure that all sorts of older peripherals will work just fine with the new OS. Unfortunately that means that all device drivers have to be ported to x86-64 as you cannot run 32-bit device drivers in a 64-bit OS. Getting that sort of driver support will take some work, although all of the latest products should have no problems with driver availability (NVIDIA has already released IA-64 GPU drivers and we're certain an x86-64 version wouldn't be too much to ask).
We also had a number of questions about the Athlon 64's architecture and how the on-die memory controller changed the rules of the game. The biggest question on everyone's mind is how difficult is it to support DDR400 and future memory technologies?
It has always been AMD's official stance that the Athlon 64 will support a single channel of DDR333 SDRAM through it's on-die memory controller. Obviously when the Athlon 64 is out, DDR400 will be picking up some momentum (thanks to Intel's endorsement of the technology) and an eventual migration to DDR400 support would be nice. Unfortunately DDR400 support will only be available through a new revision of the chip, although we're hoping that AMD may forego DDR333 support in lieu of DDR400 at launch, assuming that DDR400 is in enough supply. The reason for a new revision of the chip is only to optimize internal timings for DDR400 operation; remember that with the memory controller now on-die it is no longer necessary to make sure the memory runs in sync with the FSB frequency, as there is no "FSB" to speak of. The results of this are that internal datapaths within the chip's memory controller must be optimized at new frequencies in order to guarantee top performance.
What you can take away from this is that it doesn't mean that DDR400 won't work on first-generation Athlon 64s, but it may not run at peak performance. We brought up the idea of our more "adventurous" (read: all) readers running the first Athlon 64's out of spec with DDR400 memory settings and they confirmed that it could be a possibility. It would be up to the motherboard manufacturers to expose the appropriate settings in their BIOSes. An overclocked Athlon 64 with DDR400 support would not, in theory, outperform a new revision with optimized DDR400 support.
Implementing DDR400 support in the Athlon 64 is a relatively trivial matter compared to supporting a new memory technology like DDR-II for example. The major differences between DDR-I and DDR-II exist in the interface, or the in the way the controller talks to the memory; the signaling is vastly different as are the electrical characteristics of the interface. The end result is that the move to integrate DDR-II support is much more difficult but manageable of a task for AMD. Once again, DDR-II support would come in a future revision of the chip; and unlike DDR400 support, you wouldn't be able to overclock your Athlon 64 to support DDR-II.
The fact that AMD will now have an on-die memory controller means that they are less likely to introduce support brand new memory standards on their own but rather wait until industry support is there and then release Athlon 64 support. This could work to the industry's advantage as we won't have another company supporting another set of varying memory standards. The reason AMD will have to be cautious of what they support is simple; just think about what would have happened to Intel's CPUs if they had integrated a Rambus memory controller in them back in the i820/Pentium III days.
Packaging was another question we've been curious about; the 754-pin Athlon 64 will definitely be the most expensive desktop x86 CPU from a packaging standpoint, but how will that translate into CPU and overall system prices?
AMD admitted that the CPU would come with a premium because of the packaging issue, with the biggest problem being power delivery to the core (read our BBUL article for more information on what factors influence packaging technology). Remember that the only way power gets into the die itself is through approximately 30% of those 754 pins that connect to tiny wires on the inside of the package which finally bring power to the chip. On the bright side, AMD did mention that Athlon 64 chip prices would be competitive with Pentium 4s; indicating that the Athlon 64 would compete with the higher end P4s on price. On the system side, AMD expected Athlon 64 systems to be selling for $1600 - $2300 at launch, which is fairly reasonable.
Looking towards the future, by the end of 2004 AMD expects close to 100% of their desktop production will be Athlon 64. The benefit of producing such large numbers of CPUs is that AMD will be able to significantly reduce the price premium of the Athlon 64's packaging, by then, hopefully to 0% over an Athlon XP.
Athlon 64 - Motherboards Galore
AMD had a huge display of Athlon 64 motherboards at their suite (the same suite as last year's Comdex). This is greatly in contrast to the launch of the Athlon back in 1999 where the only motherboards to be found were really from FIC and AMD with their reference boards.
Athlon 64 - Motherboards Galore (continued)
Athlon 64 - Motherboards Galore (continued)
Athlon 64 - Motherboard Prices
We've been talking all this week about how the only thing the industry is waiting on with Athlon 64 are the CPUs from AMD; everything else, chipsets, motherboards and even OEMs are in place. There is much truth to this as virtually all of the motherboard manufacturers we visited had their Athlon 64 solutions up and running. AMD was imposing NDAs on them, not allowing them to show the systems running but the NDAs were not very strict. As you can tell from our coverage earlier this week, we were able to benchmark one setup and others were simply masked by hiding the name of the processor. This is usually the case at Comdex, an AMD or Intel will have a product but won't let their partners show it - inevitably the product is shown, it just requires a bit of finesse, both on the part of the manufacturer and on us as well.
We brought up the issue of CPU cost, but now let's talk motherboard prices. One of the benefits of AMD's Hammer architecture is that there is no on-board North Bridge, which significantly reduces the complexity of motherboard design. The area around the CPU and North Bridge is by far the most difficult part of the motherboard to design, and the most expensive. By eliminating the external CPU-North Bridge connection, AMD has effectively made motherboard design much easier.
The end result is that all Athlon 64 boards (at least all that you see here) are 4-layer designs and should be fairly cheap. In terms of chipset costs, there's only a South Bridge and AGP controller to pay for unless you have a solution with on-board/integrated video. The AGP controller is noticeably cheaper than a full North Bridge, so chipset cost is reduced a bit. As we mentioned before, board design is simplified and manufacturing costs are kept to a minimum thanks to AMD's 4-layer reference design (remember that the Athlon launched with a 6-layer board initially, while Athlon 64 is able to come to market on a 4-layer design immediately). If you add all of those factors up, we would expect initial board prices to be cheaper than current nForce2 & 845PE motherboards.
By far the biggest push from AMD is on making Opteron the success it could be on the server side of things.
A 4-way Opteron running at 1.4GHz
AMD has published estimated SPEC benchmarks of the Opteron and compared it to Xeon, where the Opteron measures up quite favorably. We asked AMD if they honestly felt that this would translate into real world performance (e.g. our DB server tests) and they responded with an interesting "we feel that the SPEC numbers are a bit conservative." Quite bold claims from a company that hasn't yet demoed a system above 1.4GHz but their confidence was quite inspiring and interestingly enough, believable.
We'll see what happens come crunch-time but we rest assured that the hardware behind Opteron is there, it's merely a question of if they can get it out and in great enough quantities.
Newisys had their 2-way 1U Opteron server up and running a 64-bit version of Windows.NET Server beta at AMD's suite:
Because of the on-die memory controller each Opteron CPU gets its own set of 4 memory slots as you can see to the left and right of the CPUs below (the leftmost CPU has its memory slots covered by a SCSI cable):
The board itself is fairly clean, with the majority of the components you see being used for Newisys' own remote management interface. Everything past the edge of the power supply is all that is needed to support the CPUs; such a simplified design is made possible thanks to AMD's on-die memory controller with the Opteron CPU. Since servers have no need for an AGP controller, the only other chip that's necessary is a South Bridge which you see directly behind the first Opteron processor.
Newisys' management interface allows remote monitoring over LAN of the server's components.
Above you can see the first CPU's set of memory slots; note that memory is installed in pairs because of the dual channel DDR memory controller that exists in the Opteron processor.
AMD's 64-bit Push
One area AMD was possibly too optimistic on was what 64-bit memory addressability would bring to the desktop. Promises of bigger, more intelligent, photorealistic games for example are honestly not going to be made possible by the move to 64-bit. But AMD insists that 64-bit does more for the desktop than Hyper-Threading; maybe that's just AMD-PR talking but it's still too far fetched for us to swallow.
AMD did have the new 64-bit port of the latest Unreal engine up and running at their suite. From talking to Epic it was clear that they had intentions of supporting x86-64 from the start and like almost everybody, are simply waiting for AMD to release the CPUs. Epic has said that they will ship a 64-bit version of Unreal Tournament 2003 when Athlon 64 makes it to retail, and the performance increase from moving to Opteron is significant. You get a 15% speedup from the on-die memory controller and another 15% from compiler optimization for the additional registers in x86-64 mode.
If you look at how desktop memory sizes have grown in the past 5 years, you'll notice that on the average enthusiast's desktop we've seen a 4x - 16x increase in memory size. If we continue that growth trend you'll see that it will take another 3 - 5 years before the enthusiast desktop reaches the 4GB memory addressability limit of current 32-bit architectures. Being able to use more than 4GB of memory will, without a doubt enable some pretty amazing things on the desktop end but there are other limitations to get around first before memory size becomes an issue.
We asked AMD what they thought of Intel's Hyper-Threading technology and if they would ever support it. As we expected, AMD said that if the technology did take off and received industry support that they could conceivably support it (or maybe a similar technology). For now however AMD sees their 64-bit ISA (Instruction Set Architecture) and affordable multiprocessing as ways of improving performance, not Hyper-Threading.
Transmeta shows off their "Banias killer"
Transmeta approached us in the elevator lobby of the Venetian and invited us up to their suite so we could see their "Banias killer."
As most of you know, Banias is Intel's upcoming processor that was designed from the ground-up to be a mobile CPU. The base architecture of Banias is supposed to be a shift back to the Pentium III pipeline but with a number of architectural improvements (e.g. micro-ops fusion). One of the biggest selling points when it comes to Banias is its supposedly very low power operation, while offering very high performance (the launch Banias 1.6GHz will outperform a Pentium 4 1.6GHz).
When it comes to low-power operation, Transmeta's CPUs do have quite a bit to bring to the table, but they've generally been weak performers when compared to Intel's mobile Pentium III. Transmeta gave us a presentation as well as a live demo of what they call their "Banias killer" or their Astro processor.
The Astro processor is a logical evolution of the TM5x00 design and is referred to as the TM8000. Although not specifically mentioned, we can assume that Astro has the same on-die North Bridge architecture as the current Crusoe CPUs, although maybe with a faster memory controller. Technical details on Astro are sparse at best but the major improvements include:
- 256-bit wide instruction bundles (can bundle up to 8 32-bit instructions per clock vs. 4 in the current design)
- Shortened critical paths within the TM5x00 CPU design to provide for higher clock speeds.
- Increased IPC through larger cache and higher execution throughput
The reason that details are scarce at this point is that the CPU will not be in production until Q3-2003, just under a year from now. The CPU is being demonstrated now to give the world an indication that Transmeta will have a Banias competitor, although not until much after the Intel's official launch in Q1.
Despite the fact that the CPU won't be out for another 9 months, Transmeta did have working silicon in a stable enough state to run some preliminary benchmarks. The CPU they demonstrated was running at 1.067GHz as reported by the system's BIOS and they compared it to a mobile Pentium 4 running at 1.80GHz.
The benchmarks weren't performed very scientifically at all; they involved manually timing the start-up of Microsoft Word and PowerPoint as they attempted to open multi-megabyte files. In all cases the Astro was faster than the mobile Pentium 4 however what invalidated the results was the fact that Transmeta outfitted the Astro system with a desktop hard drive and the Pentium 4 laptop had a slow notebook drive. Transmeta attempted to justify the comparison by saying that the desktop hard drive was only an ATA-33 drive but as you all know, peak transfer rates are far from the best gauge of a drive's performance (otherwise ATA-133 drives would have taken off).
There is potential with the Astro platform and by the end of next year we'll have an idea of how well the chip performs, but until then we're just going to have to wait and see. It's too early to call Astro a "Banias killer" especially considering that it will be available 6+ months after Banias' release.
Transmeta's Impressive TM5800 Product List
Although not a very high performing solution, Transmeta's TM5800 CPU has a significant advantage over Intel's mobile solutions - physical size. Because of the CPU's on-die North Bridge and the tiny size of the chip itself (courtesy of a software based instruction decoder), motherboards can be made very small. The cool operation of the CPU and small form factor packaging enable some very tiny designs, which is what we're seeing on the market today.
Crusoe based motherboards can be very small thanks to the on-die North Bridge and small packaging.
Transmeta's suite was full of current TM5800 design wins and we snapped a few pictures for you:
Fujitsu had a pair of notebooks on display; their small form factor is pretty much the norm when it comes to Transmeta design wins:
Innolabs had a Tablet PC running the TM5800; PaceBook has a similar offering with the PaceBlade:
Transmeta's Impressive TM5800 Product List (continued)
One of the more impressive solutions was this Sharp notebook:
The Mebius Muramasa from Sharp only weighs 2.09lbs in the configuration you see above, but the real beauty of the notebook is how thin it is:
As you can tell by this picture, the Mebius is well under an inch thick.
Even when closed the notebook is impressively thin.
Another neat feature is this docking station:
The notebook charges itself when put in this docking station but it also establishes a USB2.0 connection to the docking station. You can then plug the docking station into your network and access the hard drive of the Mebius on any other computer as a removable disk drive.
Although the notebook is currently only available in Japan, Sharp has announced plans to bring it to the rest of the world including North America.
Our final stop in the Transmeta suite was at this tiny notebook from Sony. It's about 2/3 the width of a Picturebook and as you can see by the picture above, it is extremely tiny.
NVIDIA's Personal Cinema 2
Albatron had a demo of NVIDIA's Personal Cinema 2 product on display:
There was a good deal of hype surrounding the original Personal Cinema release but when push came to shove, it didn't have the software or the hardware to compete with ATI's All-in-Wonder line.
Personal Cinema 2 fixes some of the hardware issues behind the original Personal Cinema. For starters there is a tuner on-board, making the breakout box much simpler. The connector to the breakout box is unusually reminiscent of the latest All-in-Wonder cards but we still see no options for component output, even though it should be simple to implement in theory.
NVIDIA has also moved to a remote manufactured by X10 (the same company that manufactures the ATI Remote Wonder) that communicates using radio frequency (RF) instead of infrared (IR) transmission. The benefit, just as we saw with the All-in-Wonder Radeon 8500/9700 is that you don't need to have line-of-sight between the remote and the receiver and can control the machine even if it is sitting in another room.
NVIDIA has effectively caught up to ATI on the hardware side but there are still some key issues that remain even with Personal Cinema 2:
1) Software and ease of use has yet to be proven and was definitely not a strongpoint of the original Personal Cinema. Especially after using Microsoft's Windows XP Media Center Edition it's clear that even ATI has a long way to go before these cards can be used as set-top PVR replacements.
2) No manufacturers have plans to couple the technology with NVIDIA's highest speed cards. It is easy to implement on anything up to even a GeForce FX yet manufacturers insist on pairing Personal Cinema 2 up with GeForce4 MX. There is a much greater market with the GeForce4 MX but there are a good number of folks willing to spend more on a flagship part with Personal Cinema 2 functionality.
NVIDIA is progressing with Personal Cinema, but it seems like they will have to still play catch-up to ATI here.
NVIDIA Changes plans on Ti 4800
NVIDIA's marketing had a meeting to discuss the future of the GeForce Ti 4800 and Ti 4800-SE; it turns out that NVIDIA has reversed their decision and will sell the Ti 4800 as a GeForce4 Ti 4600-8X in North America. For the rest of the world, the Ti 4800 and Ti 4800-SE will continue to be manufactured under the new name.
As we mentioned in our first piece on the new naming, the Ti 4800 is nothing more than an AGP 8X version of the Ti 4600 and the Ti 4800-SE is an AGP 8X version of the Ti 4400. With no difference in clock speeds it's good to see that NVIDIA will be selling the AGP 8X card under a Ti 4600-8X name in North America, but let's just hope that the rest of the world doesn't get too thrown off by the naming system.
Palm's Tungsten T
Palm has finally moved to a faster processor with their latest PDA - the Tungsten T. The Tungsten T is ARM powered using a TI processor and features a 65K color display, integrated Bluetooth and SD expansion support.
We got a chance to hold the unit at Palm's booth on the show floor:
The Tungsten T is quite small when the graffiti area is hidden and feels very well built.
Extending the cover exposes the graffiti writing area and makes the PDA about the size of most Pocket PC devices.
We hope you enjoyed our coverage of Comdex 2002; assuming that there even is a Comdex next year, we'll definitely be here to cover it.