Original Link: https://www.anandtech.com/show/1243



Introduction

IDF this year is shaping up to be a very informative week at the beginning of many changes sweeping through the PC world. We are about to see new chipsets, busses, processors, architectures, sockets, and memories, and some of those announcements are starting to be made here.

The highlights of day one consist of PCI Express graphics card announcements, the state of DDR2 (and a little about RAMBUS as well), and, of course, the fact that Intel has officially announced its adoption of x86-64 for its Nocona Xeon core.

Without further ado, let's check out what's new in computing.



PCI Express Product Launches

NVIDIA and ATI have officially begun the switch to PCI Express by launching their initial products. NVIDIA implements PCI Express via a bridge chip while ATI's solutions are native. Our ATI Roadmap article shows the lineup we can expect to see come out of ATI this year. NVIDIA launched four PCI Expressed solutions dubbed Geforce PCX:

GeForce PCX 5950
GeForce PCX 5750
GeForce PCX 5300
GeForce PCX 4300

The PCX 4300 is an iteration of NVIDIA's MX value line, the 5300 and 5750 are bridged 5200 and 5700 cards respectively.

Apparently ATI demonstrated a "next generation" GPU at a presentation they gave on Tuesday showing off 2 to 3x performance gains over their current generation cards running DX9. Of course, it was an unreleased game on unreleased hardware, we didn't get to run the benchmarks ourselves, and apparently the performance of the next gen part was just at acceptable levels.

ATI is really pushing the idea that native PCI Express gives them an advantage in things like running multiple HD streams at higher frame rates than AGP or a bridged solution allow, but again, we need hardware to test in order to verify these claims.

We will have longer meetings with both ATI and NVIDIA over the next couple days, and we will be sure to report on what learn.

DDR, DDR2, RDRAM, and XDR

RAM technology continues to moving forward; many vendors were showing off their DDR2 memory at the Technology showcase, but we are going to save that coverage for later today.

We were able to speak with Samsung about what they are doing with DRAM at the moment. Samsung really has their bases covered making SDR, DDR, DDR2, 4 channel RDRAM and XDR. The upcoming XDR chips were on display up at the RAMBUS both across from a demo of Toshiba chips running at very high speeds (the bandwidth of XDR is 6.4GB/s).

4 Channel RDRAM will be supported by a SIS P4 chipset, and we are told to expect good things from P4 performance when dropped in a SIS system with 4 channel RDRAM. Hopefully we will be able to put the theories to the test sooner rather than later.



Nocona and Prescott: Intel adopts x86-64

There was an incredible amount of speculation that Intel would make a formal announcement about their use of AMD's 64 bit extensions to the x86 ISA at IDF this week. Intel has indicated for a long while now that they would adopt 64 bit for the desktop "when it makes sense." Of course, the time frame we have been given for such a thing making sense has always been much further down the road than this year, and the tradition continues still. Even though Intel has announced that its Nocona (90nm Xeon) processor will have 64 bit x86 extensions enabled, they are targeting this squarely at the workstation/server market and have still not made the decision to move x86-64 to the desktop.

It is the case that x86-64 can be released in a P4 form anytime Intel wants (though it may end up being later rather than sooner) since the Nocona processor is based on a Prescott core with its 64 bit hardware enabled in a Xeon package (and with Xeon sized caches and features). We are looking into the method with which Intel has disabled the 64 bit extensions in current versions of Prescott, but we don't have conclusive data as of yet. We suspect, though, that the extensions are disabled much the same way that clock speeds are locked (so that neither enthusiasts nor remarkers can add value not included straight from Intel).

As far as what is actually going on architecturally, we still need to do a little digging. We do know some things for fact. Intel's implementation of the x86-64 extensions will be completely compatible with AMD's. The extensions are in the current version of Prescott in a disabled state (and Intel is still determining an appropriate time to release a 64bit enabled P4).

We still have questions about Intel's ALU design and how it supports the new extensions, as well as whether or not Nocona will have a larger trace cache than Prescott. Needless to day, there are still plenty of things we don't know yet.

Up until last year (with the release of the Athlon64), Intel had four options for more accessible 64 bit computing: bring the Itanium's EPIC (Explicitly Parallel Instruction Computer) based IA-64 to the desktop, develop a desktop 64 bit ISA based around fast emulation of x86, create their own 64 bit extensions to the x86 architecture, or adopt AMD's extensions to x86 for their future processors. Let's take a look at these options to try to understand how we got here today.

IA-64 on the Desktop?

The Itanium processor has shown itself not to be the worlds greatest x86 emulator. The problem with bringing IA-64 to the desktop is that the architecture just isn't well suited to x86 style code. Much of the power of IA-64 can only be exploited via very complex compilers designed to optimize and schedule IA-64 code. It is possible to translate, dynamically recompile and optimize code on the fly (ala Transmeta's code morphing or some of the edgy Java JIT compiler research), but the complexity of IA-64 doesn't make this process any faster.

Optimization and scheduling is an NP-complete problem (as is the famous traveling salesman problem) which means that the faster and more efficient you want your code to run, the longer it will take to compile. On the bright side, there is generally a point of diminishing returns at which the virtually infinite loop of optimization and scheduling can be stopped without sacrificing too much. Of course, doing all this on the fly and hoping to come up with near optimum code at high speeds is a very difficult problem to solve.

There is still the issue of getting that very wide, cache hungry, complex architecture down to a price point consumers can afford even if Intel is able to bring it out with performance that makes it a better option than native x86 CPUs.

IA-64 really would be a kludge on the desktop without turning it into an ultra emulated x86 with 8-way-hyperthreading and a +5 Dynamic Compiler of Doom to try to exploit its potential for achieving high IPC (Instructions Per Clock - how much work gets done every clock cycle). Honestly, by the time we get to a point where something like this is feasible, we will probably already be enjoying multicore CPUs and it still won't be any faster than cheaper native solutions.

Intel has informed us that they have not thought of IA-64 as a desktop architecture in any way for the present or the future. Popular opinion seems to suggest that either x86-64 is causing problems for IA-64, or that IA-64 was supposed to come to the desktop but that Intel's plans have now been thwarted. From our brief analysis of IA-64, it seems to us to be a particularly bad idea to even think about moving it to the desktop space. It seems to make more sense that Intel definitely had other plans up their sleeves. Let's take a look at our other two options.

Intel x86 Extensions?

Intel has extended the x86 instruction set in the past, and they could have done so once again when 4GBs of memory just wasn't enough for desktop users. AMD may have jumped the gun slightly on 64 bit computing as most of us don't even get to see beyond 2GBs of RAM in our systems today. Of course, by jumping the gun they were able to add some performance enhancing characteristics to their current line of CPUs that could potentially make them more attractive to customers than Intel's CPUs.

Unless Intel added even more impressive architecture changes than AMD, out pacing what is already out there, there would be little reason to buy an Intel processor. Another very big advantage that comes with AMD's preemptive push toward 64 bit computing is that by the time Intel could bring out its own extensions, operating system and software support will already be there for x86-64. We've seen first hand how long it has taken Microsoft to bring AMD's extensions to Windows XP, and there is no reason to believe that MS would be quicker to support Intel's enhancements (branching an OS, rewriting parts of it, and putting it all together with support and documentation is not a quick and easy task).

All signs point to the fact that it is very unlikely Intel could expect to release separate extensions to x86 that are not compatible with AMD's extensions. Which brings us to the last and final option Intel had for 64 bit and x86.

One ISA to Rule Them All

AMD's x86-64 based processors perform really well in 32 bit mode (and better performance in 64bit mode). AMD is moving all their product lines over to x86-64. Microsoft is on the verge of releasing an operating system to support x86-64 on which both 32 bit and 64 bit applications can run. Driver and application support are getting better everyday. It certainly seems like x86-64 is the place to be for desktop 64 bit architectures.

There have been rumors about Yamhill for a very long time, Intel and AMD have licensing agreements that give Intel access to AMD's extensions, and Intel will already have software support as they make their move (this verified by Steve Ballmer of Microsoft).

The only downside is that some in the industry seem to think that Intel be submitting to AMD on some level by adopting their extensions. While this kind of perception could help AMD, we at AnandTech believe that credit should go where credit is due: AMD designed a very solid and effective extension to x86 regardless of whether or not Intel wanted to adopt it. As for Intel running with x86-64, more power to them as well: there exists a better option than current x86 technology, and by implementing x86-64, Intel shows a commitment to providing high performance processors to its customers (at least in the workstation and server space).

Final Words

We will have much deeper coverage of the PCI Express based solutions when we actually have hardware to play with, so, for now, the official product launches are the stories here.

We are hoping to get sample hardware in for both DDR2 and 4 channel RDRAM (supported on P4 via a SIS chipset) as soon as possible so we can see how their performance stacks up against our old friend DDR. XDR DRAM is a little further down the road right now, and we don't really know where this insanely fast RAM will end up. We do know, however, that Sony and Toshiba have licensed XDR for undisclosed products ... I wonder what that could be ...

If AMD had simply decided to extend the range of addressable memory with their x86-64 extensions, this playing field would look entirely different. It would be easy for Intel to stand behind their mantra of "we will provide 64 bit support when it is needed" (a line we had been hearing for well over a year). Frankly, at this point, 64 bit support is not a necessity on the desktop.

But performance is the bottom line. By adding performance enhancing aspects to the architecture, AMD added another level to the issue. Intel can't simply say that 64 bit isn't needed (even if it isn't) because now the issue isn't do I need 64 bits as much as how many registers exist to which my compiler and I have direct access. Intel doesn't need to adopt 64 bits for the desktop; Intel needs to adopt the performance enhancements of x86-64. It just so happens that, in order to do so, 64 bits will come along with the package.

Intel knew it would have to launch x86-64 eventually. By building the functionality disabled into Prescott, they are saving a little face. Intel can launch Nocona and say prove to the world that they also have 64 bit x86 support, while at the same time supporting their position that 64 bits is not yet needed on the desktop.

It really does makes so much sense for Intel to go with x86-64 in its processors that we really aren't surprised by this development. The timing of the announcement did throw us off a little bit, but, in looking back on how the events unfolded, we understand why we are where we are.

Now all we have left to do is figure out exactly what's going on under the hood. As we learn more about the fuzzier parts of x86-64 (and anything else interesting for that matter), we will be brining you updates. Stay tuned for more IDF, including coverage of the Technology Showcase.

Log in

Don't have an account? Sign up now