TechRadar.com - Deep into technology

AMD’s triple-core Phenom X3 is finally here. It may have sounded like an April Fool, but AMD has finally delivered on its announcement back in September 2007. Will the new holy trinity provide the renewed fortunes AMD needs to combat Intel?

The triple-core advantage

Although AMD’s Phenom X4 can’t quite compete with Intel’s Core 2 Quad even at the same clock speed, its real issue is the low frequencies currently available. The fastest Phenom X4 runs at 2.5GHz, whereas Intel’s Core 2 Extreme QX9770 is now on 3.2GHz. A clock speed issue is essentially a yield issue – AMD hasn’t perfected its Phenom production process sufficiently to produce enough high-clock CPUs just yet.

Although AMD would like to spin things another way, touting the benefits of three cores, the arrival of the Phenom X3 can only be an answer to this yield problem. AMD doesn’t have a special production line producing triple-core CPUs. They are all quad-core parts with one core disabled. The only reason AMD would want to do that is if one core isn’t performing up to the standard of the other three.

With its current designs, Intel can’t easily produce triple-core processors. Its quad-core parts are essentially two dual-core CPUs packaged together and joined at the FSB. So a triple-core design would mean two cores on one half, and one on the other – technically feasible, but hardly elegant.

So from AMD’s point of view, the Phenom X3 makes good sense. The Phenom X4 is hardly expensive – even the flagship 9850 Black Edition is only £150. But the entry-level Phenom X3 8450 is already below £100, putting it in the same league as AMD’s quickest dual-core Athlon 64 X2s.

Only a handful of Intel dual-core processors are cheaper – the Allendale 65nm ones with 2MB of L2 cache and 800MHz FSB rather than the usual 4MB and 1,066MHz of Conroe, plus the 45nm Core 2 Duo E7200, which has 3MB of L2 and a 1,066MHz FSB, rather than the usual 6MB and 1,333MHz respectively.

That’s likely to be an easy retail sell, now that true megahertz have been hidden behind mysterious model numbers in the processor market. For many punters, three cores for the same or less money than two sounds like a winner.

Intel strikes back

Intel’s response has been to slash its quad-core prices. And herein lies the rub for AMD. The 65nm Core 2 Q6600 can now be snapped up for a modest £134. The Q6600 has been a proven favourite amongst overclockers for nine months already.

So now the choice is dual-core, AMD triple-core for the same money… or add £40 and go quad-core. Intel’s other quad-cores are considerably more expensive, particularly the 45nm variety. The Q6600 is Intel’s single-handed spoiler.

If the Q6600 was available for £100, that would have been game, set and match to Intel. The current price leaves a little doubt, and a decision for potential buyers to make between price and an extra core.

So AMD should sell a few X3s. But only to those who don’t realise that an extra core is going to be of benefit primarily for tasks like professional 3D rendering. For most everyday software – and virtually all games – a higher-clocked dual-core processor will almost certainly be quicker.

It hasn’t taken long for some of Nvidia’s recent acquisitions to bear fruit. Just a few months ago the company got its hands on trailblazing games physics company Ageia. Now it appears the PhysX engine is already running on Nvidia graphics cards.

Kudos to CUDA

The port to Nvidia hardware takes advantage of CUDA, a C programming interface which allows software other than graphics applications to directly address the stream processors on GeForce 8 series and above.

There are over a hundred of these on a top-end GeForce 8 or 9, each one capable of thousands of floating point operations a second. So the potential for calculating physics algorithms (amongst many other things) is immense. Basically, anything which involves lots or repetitive floating point operations could be accelerated by a graphics card.

Nvidia wasn’t the first to announce it was making its GPUs available for wider processing tasks than just 3D acceleration. ATI launched its Stream Computing  initiative in September 2006, closely followed by Close to Metal a couple of months later. The latter brought general purpose computing on graphics processing units (GPGPU) to every day graphics cards.

CUDA was Nvidia’s response. It runs exclusively on GeForce 8 series cards and later (plus their professional Quadro FX siblings). The software developer kit for creating CUDA-powered applications only arrived in February 2007, alongside the Nvidia Quadro FX 4600 and 5600.

A physical future

Both ATI and Nvidia have been talking about running physics on their graphics cards for some time, and even showing off working demos. The implication was that GPUs would exceed the capabilities of dedicated hardware.

Looking at the performance figures quoted by Tom’s Hardware, this is now the case. Nvidia’s presentation mostly focused on how much faster than an Intel multi-core processor its PhysX port was. But it also showed better performance than dedicated Ageia PhysX hardware. Considering Nvidia now owns Ageia, the PhysX hardware looks likely to be seeing the end of the line – which many predicted anyway.

But the news can only be good. It’s hard to deny the potential of hardware physics acceleration. Anyone who has played CellFactor on a PhysX-equipped system will attest to the extra possibilities for fun from a world where you can throw stuff around, send piles of boxes flying, and see cloth fluttering in the wind. It’s another step towards more immersive realism.

In the past, the barrier to more elaborate physics adoption was a classic Catch 22 situation. Games developers were reluctant to harness its full capabilities until enough people had the hardware to take advantage of this. CellFactor is essentially an Ageia-funded technology demo turned into a full game. But gamers weren’t going to buy the hardware until the games supporting it arrived.

Now, however, anyone with a GeForce 8 or later will have the capability. So instead of the reported 150,000 PhysX owners, there will be 50 million Nvidia -owning gamers potentially interested. That’s a much stronger argument for PC games developers.

In other words, Nvidia’s announcement is just what the market needs for physics to make that leap from clever niche to mainstream gaming technology. With 3D graphics now so photorealistic that further enhancements have become boring, physics has the potential to revolutionise gaming again. After all, the real world isn’t just skin deep.

There have been a lot of rumours in the last few months about Nvidia developing its own CPU, although nothing official has come to light just yet. But a recent Nvidia announcement hints that the processor market really could be hotting up over the next few years, irrespective of whether AMD manages to get back to its former glory.

In the announcement, Nvidia has brokered a deal with VIA to produce chipsets for a new low-cost platform incorporating VIA’s forthcoming Isaiah processor. At the beginning of the year, the VIA Isaiah made the news, for promising to double or even quadruple the performance of existing VIA C7 processors, whilst still keeping the power requirement low.

What Nvidia will bring to the party is of course its capable integrated graphics chipsets.

The hope is to produce a platform which has comparable (or even better) CPU performance than an Intel Celeron, with significantly more powerful GPU capability than any Intel integrated chipset can muster. And all for $40 (£20), allegedly.

One of the rumours about Nvidia’s CPU intentions involved Nvidia buying VIA, to get hold of the x86 licence held by the Centaur processor division. Clearly, although those talks haven’t come to fruition just yet (apparently due to VIA quoting too high a price), they have borne other fruit.

Why does Nvidia need a CPU?

Looking at things purely from today’s point of view, Nvidia is still doing well. It just about has the high ground with its GPUs (although ATI’s next generation is imminent), and its latest chipsets have finally added support for DDR3 memory. So you don’t need to sacrifice the option of SLI anymore, by choosing an Intel chipset, if you want the latest memory support.

But in the long term, Nvidia will be facing Intel’s Larrabee, and the prospect of graphics moving over to ray-traced rendering on a heavily multi-core CPU. So Nvidia really does need to be in the CPU business sooner or later.

The VIA announcement clearly isn’t about Nvidia actually producing its own CPU. But it does put Nvidia more directly in competition with Intel. Although their chipsets have been fighting in the same market for years, there has been as much synergy as combat on the platform front.

But single-manufacturer platforms are increasingly becoming the focus of attention. AMD can now supply you a complete package of CPU, graphics and chipset. When Larrabee arrives, Intel will be able to do the same thing. That will leave Nvidia out in the cold.

Splitting the Intel Atom

The new Isaiah-based platform is currently being compared to the Celeron, although Intel’s Atom will be a more likely foe. With the success of the iPhone, and the increasing focus on even cheaper PCs, this is clearly an area of potential growth.

Virtually everyone has a computer nowadays who wants one, apart from those on extremely low incomes. So the question is how to sell computers to the poor, and how to get those of us who already have powerful desktops and notebooks to purchase yet more computing devices – either handheld or secondary ones in every room of the house.

Low-cost platforms are the answer to this, and Nvidia wants a piece of the pie.

The VIA announcement could also be a step towards a new platform alliance in the years to come. So will we be seeing ‘nVIAdia’ in the near future? The financial issues remain unresolved, so a merger remains questionable. But it certainly looks like Isaiah could be heralding a second coming for Nvidia.

Hardware enthusiasm has just taken another step towards the mainstream. Buying a fast PC is merely a matter of money, but tweaking a system for ultimate performance still requires considerable knowledge and patience.

This is where Nvidia's new Enhanced Performance Profiles 2.0 (EPP 2.0) are supposed to step in. Launched with the new nForce 790i chipset, EPP 2.0 provides high-speed memory settings beyond the standard JEDEC ones.

Extra performance for free

Hardware is often capable of considerably higher settings than the current standard. Processors almost always run at least a few hundred megahertz faster than they are rated. Graphics card GPUs and memory usually do too. The results are extra performance for just a little bit of trial and error.

Memory has this ability as well, and most performance modules are aimed at this. If your motherboard only supports DDR2-800 as standard, then sticking in high-end memory with a faster rating won't do you much good. The standard latency timings (indicating how fast it responds to requests) may be better when running as DDR2-800. But you will get PC2-6400 bandwidth whether you like it or not - unless you have a poke in your BIOS and improve matters.

But memory overclocking really is a black art, by far the most complex and least rewarding of all the core hardware tweaks you can perform. Getting your CPU to run faster involves a juggling act between clock multiplier and FSB, with a bit of extra voltage to keep things stable.

You will get noticeably faster application performance in return. Graphics overclocking mostly involves increasing core and memory speeds until things fall over, unless you enter the shady world of hardware modification. This will provide faster gaming.

Memory, in contrast, has at least six variables to consider, and maybe more. You increase the speed of the memory by advancing the bus speed it is running on, but there is a clock divider to consider too. Then there are at least four timing values, which will have to be increased to allow the faster clock speed. Maybe a bit of extra voltage will keep things stable. Oh, and you might find increasing the motherboard Northbridge voltage beneficial, too.

With the trade-offs between timings and brute sustained throughput, you can never be certain that your tweaks will actually be faster, either. Only extensive performance testing of different configurations will tell you for sure. See what I mean about a black art?

EPP 2.0 and XMP - memory overclocking made easy

This is where EPP 2.0 comes in. The memory manufacturer will have tested its modules and found the optimum settings for you. Memory already has a bunch of standard settings stored in EEPROM on the module itself. These are called the SPD, and the motherboard reads this to work out how to configure the memory automatically.

Less than half the 256-byte EEPROM is actually used to store the SPD settings, however. So some of the remainder can be used to store the extra EPP 2.0 settings. A compatible motherboard will then have a single BIOS option to enable this configuration. Instead of juggling umpteen settings, you can simply select EPP 2.0 mode and the memory will be configured for you. Voila! One-click memory overclocking.

As its second-generation name implies, EPP 2.0 is not a brand new idea. It is essentially the DDR3 version of EPP, aka SLI Memory, which came out in 2006. Intel has also launched a similar idea in the shape of XMP, for its '3 series' P35 and X38 chipsets, plus the new X48 chipset which is currently arriving. Like EPP 2.0, XMP lets you choose a performance configuration beyond standard SPD with a single BIOS setting. Both formats are receiving a healthy level of support, particularly from Corsair.

Unfortunately, though, neither EPP 2.0 nor XMP really go far enough yet. They set your memory timings, FSB speed and voltage to values the RAM manufacturer knows its modules can cope with. But the rest of the system is still up to you. In particular, the FSB chosen by the profile could put your processor and Northbridge beyond their capabilities, without a bit of extra voltage.

So although EPP 2.0 and XMP are a step in the right direction towards performance overclocking for the masses, there's still quite a bit of a way to go before your granny could give it a whirl. Unless she has a PhD in computing, that is.

As predicted last month, AMD has launched a slew of new processors, but with very little fanfare. This is not surprising, considering even the fastest one, the Phenom X4 9850, runs at just 2.5GHz. That's not exactly going to put the wind up Intel.

Indeed, our sister title Maximum PC over in the US put the new Phenom up against its clock-for-clock quad-core adversaries from Intel, and the results were not pretty. The Phenom X4 9850's 3DMark06 CPU results, PCMark05 score, and virtually all other benchmarks were still behind Intel's entry-level Core 2 Quad Q6600, which runs at just 2.4GHz. The new Core 2 Quad Q9300 running at the same 2.5GHz clock as the Phenom was even further ahead.

Not as cheap as chips after all

At this point, we'd like to be able to say 'but the AMD processor is much cheaper'. Unfortunately, that isn't true either, despite the prices AMD quoted at launch. Looking at current retail pricing (from www.scan.co.uk), you can pick up the Intel Core 2 Quad Q6600 for £147.92, where the AMD Phenom X4 9850 is around a tenner dearer at £157.45. The Q6600 has another price drop due in April too. The Core 2 Quad Q9300 is more expensive at £198.56, though.

At least the Phenom X4 9850 is a 'Black Edition', so it has an unlocked multiplier. This will make it easy to overclock without having to fiddle with memory speeds and other variables. But even without this facility, the Q6600 is already a legendary overclocker. Where the 9850 is a 125W part, the Q6600 is 95W, and lower thermal design power (TDP) usually means greater headroom for faster clock speeds.

Indeed, the G0 stepping Q6600 has been found to reach speeds well in excess of 3.3GHz with just air cooling. But we haven't yet seen any reports of a Phenom X4 9850 getting much beyond 3.1GHz. Since the Phenom can't beat the Core 2 even with a 100MHz clock advantage, it's not going to be breaking any benchmarking records for the foreseeable future.

Budget consciousness

So the new Phenoms may be a small step in the right direction for AMD, particularly as the X4 parts are the new B3 stepping, and therefore aren't prone to the infamous Translation Lookaside Buffer (TLB) bug. But since the fastest, enthusiast-oriented X4 still can't keep up with Intel's entry-level offerings, it's nowhere near enough to put AMD back on track with enthusiasts, or even the mainstream.

The budget end of the market is the only battleground left. Intel seems to have left this low-margin section of business open to AMD. After all, there may be volume there, but it's not exactly good for image building, nor tremendously profitable. Non-Intel processors have targeted the lowest prices, and most of the companies involved have not done well out of it. IDT Winchip, anyone?

Coming hot on the heals of reports that AMD will be culling 5 per cent of its workforce, the new processors have the smell of desperation about them. AMD obviously needs a lean operation to keep the ball rolling until it can get some more competitive products onto the market.

Nvidia also seems to be desperate to get into the x86 business, so AMD could well have even more competition to contend with in the near future. In other words, AMD's latest round of processor releases isn't going to help its cause much at all.

When a new memory type arrives, there is always controversy over whether it is actually any faster than what it is supposed to replace. Sometimes, the controversy continues long enough to end the new RAM's hopes of dominance - as in the case of RAMBUS.

But after a while, the inflated price at introduction drops, speed increases, and the new memory type succeeds the old. Is it now that time for DDR3?

Memory remembered

When DDR2 memory arrived, it really wasn't faster than DDR. In fact, it was slower in lots of tests. This seemed a bit odd, when the bandwidth available appeared to be greater. But that's not the only factor in memory performance. Another important consideration is its timings - represented by those four numbers in the memory specification, like 2-2-2-5 or 7-7-7-18. The first of the quartet is the most significant, as it refers to the CAS Latency.

Although memory throughput is a major contributor to system speed, the time modules take to get up to speed will also have an effect. A longer CAS Latency - represented by a larger timing number - will mean the modules take a few more nanoseconds to deliver their maximum throughput.

So a module with a CAS Latency of 4 theoretically takes half the time to fire up data delivery than one with a CAS Latency of 8 (although all the other timing values means it's really much more complicated than this).

When DDR2 first arrived, it boasted 50 per cent or greater throughput than PC3200 DDR. But it usually had twice the latency to go with it. So, in true swings and roundabouts fashion, it wasn't really any faster.

The problem memory regularly faces is that the speed the DRAM cells can achieve internally reaches a ceiling, which prevents faster modules being created. So each successive generation of DDR adds clock-doubling technology so the cells themselves can actually run slower - and consume less power. The latter will be particularly important as we sit on the cusp of 64-bit operating systems finally becoming the norm, so that more than 4GB of memory are worth having.

We're now on the third clock-doubling generation. Where DDR is 2x SDRAM, DDR2 is 4x, and DDR3 8x. At the same time, voltages have dropped from 2.5/2.6V to 1.8V to 1.5V respectively (and SDRAM operates at 3.3V). But each time, increased latency has had to be introduced to make things work.

Eventually, as the memory technology is improved, the latency is reduced a little whilst the full throughput possibilities are realised. For example, the fastest DDR standard is PC3200, but 1,066MHz DDR2 is PC2-8500 - ie 2.65 times the bandwidth. But the fastest CAS Latencies are 2 and 5 respectively (although 4 is possible at 800MHz). So the extra bandwidth just about outweighs the increased latency.

DDR3 versus DDR2

The question is, now that Nvidia has come out with chipsets supporting DDR3, and Intel has had them for over a year, is this happening to DDR3 yet? So I thought I'd do a little spot testing of my own to see how DDR3 is faring. Helpfully, MSI makes a motherboard which accommodates either DDR3 or DDR2. And it's not based on some reduced entry-level chipset, either. The X48c Platinum uses Intel's latest X48 chipset, so will even support 1,600MHz FSB processors.

I put together a PC with an Intel Core 2 Extreme QX9650 and BFG Technology NVIDIA GeForce 8800GTS 512MB graphics. Using two 1GB sticks of Crucial Ballistix DDR2 PC2-6400 running at 800MHz with 5-5-5-18 timings, the PC achieved 13,650 3DMarks in Futuremark's 3DMark06 v110.

Using two 1GB sticks of Corsair TwinX DDR3 PC3-10,600 running at 1,333MHz with 9-9-9-24 timings, the PC achieved 13,749 3DMarks. So that's 0.7 per cent extra performance.

Bearing in mind that the DDR2 costs around £40 and the DDR3 more like £70, you're paying 75 per cent more for 0.7 per cent extra speed. From this rather unscientific single test, it would appear DDR3 still isn't very good value.

Thank heavens at least MSI has had enough sense to release a motherboard where you can still enjoy the low cost of DDR2 still whilst reaping the benefits of the fastest Intel processors this year. Then you can move over to DDR3 when it really does make a difference.

But in the long run, the lack of difference between DDR2 and DDR3 will be a moot point, since performance features and overclocking enhancements like XMP will increasingly only be available with the premium chipsets, which will primarily be focused on DDR3. So even if the performance benefits aren't really that significant, we'll all have to move over to DDR3 soon anyway.

As m'learned colleague Mr Morris pointed out the other week, Nvidia has been making some intriguing product decisions recently. Decisions that might just leave the door open for AMD to retake pole position in the PC graphics market.

All the indications are that Nvidia's new high end graphics card, to be known as the GeForce 9800 GTX, is a frankly feeble revision of its existing G92 GPU. Slightly higher clocks and few detail tweaks is about the size of it. 

G92, of course, is itself little more than a mildly modified die shrink of the beastly G80 chip. All of which means the green-tinged graphics goliath has remained essentially static, in terms of high end GPU performance, since the launch of G80 in autumn '06.

And please, let us not speak of the dual-GPU irrelevancy that is soon to be known as the Geforce 9800 GX2. Like every other dual-GPU board ever made, it's an inefficient and deeply unattractive device that's likely to deliver patchy, unreliable performance.

18 months is a long time in graphics

Anyway, the 18 months since the launch of G80 is an incredibly long time in the cut throat graphics game. Moreover, the impending release of the 9800 GTX indicates Nvidia has no immediate plans for a more powerful and complex GPU. 

It's likely we won't see such a chip from Nvidia until at least this autumn. Is that complacency? Has Nvidia been too dominant for too long? Or is there an unseen crisis bubbling behind the scenes, a troubled new GPU that has been killed off even before it was born?

It's impossible to say. But either way, I'm convinced that Nvidia will shortly be wishing things had gone differently. Because even a company with a recent track record as poor as AMD's becomes a threat given enough time.

Specifically, that threat takes the form of AMD's upcoming RV770 high end graphics chip, very likely to be sold under the Radeon HD 4800 brand. It might just be a corker. 

Based on the latest 55nm production process, the chip is expected to address most of the weaknesses of AMD's current Radeon HD 3800 family. For starters, I hear that the GPU's texture unit count has been doubled from 16 to 32 units, solving the existing HD 3800's most obvious flaw.

Rumour round-up

Then there's the rumour doing the rounds that AMD has nearly tripled the chip's shader grunt. The HD 3800 architecture currently packs 64 shader units arranged in four arrays (ignore AMD's claims of 320 stream processors, it's mostly marketing). The most popular theory for RV770 is no less than 160 shaders in five arrays (or 800 stream processors in marketing parlance).

If the details aren't confirmed, what is almost guaranteed is that RV770 will at least offer some additional functional units - unlike the disappointing GeForce 9800 GTX from Nvidia. Simply put, it's a significantly larger chip built on the same process as the HD 3800. The extra transistors must have been spent on something.

Best of all, it looks like it won't be long before we find out for sure. Early engineering samples of RV770 has been spotted in the wild. It could be on sale before summer is in full swing.

Nvidia is grabbing all the attention at the moment. After confusing us by making the GeForce 9800 GTX the same chip as the GeForce 8800 GTS, and the GeForce 9600 GT a cut-down version of it too, Nvidia has put off their successor until nearly the end of 2008. According to the rumours, the company's next-generation chip is ready to go right now, but there is simply no need for it at the moment.

The long reign of the GeForce 8000

The graphics business hasn't been quite as one-sided as the processor business over the last couple of years. But Nvidia has mostly had the upper hand since it launched the GeForce 8000 series. ATI's Radeon HD 2900 XT could not take the top performance slot, even if it did represent decent value.

For the next month or so, ATI does have the fastest card in many benchmarks, the Radeon HD 3870 X2. But it only just beats Nvidia's GeForce 8800 GTS (the 512MB version, just in case you're confused...), which is significantly cheaper. So the soon-to-be-released GeForce 9800 GTX and GX2 should take the top slots, even if they are merely a bit faster than the GTS.

With the GeForce 8000 series arriving at the end of 2006, Nvidia has been dining out for quite some time on the same architecture - far longer than usual. Graphics chips have generally had an annual cycle in the last few years, with a refresh halfway through. But it could be close to two years before G80 and its developments are radically upgraded. The new chip is now expected in the third, or even fourth quarter of 2008.

The GT200 secret

There's considerable confusion what Nvidia's new chip is actually called. It was referred to as G100 for a bit, at least by industry commentators. Now it's being known as GT200. Nvidia itself has switched over from the Gxx nomenclature (and the NVxx before it) to yet another GPU reference system. So the GT200 also goes by D10E (desktop, 10th generation, enthusiast market).

There's still no definitive idea what G100/GT200/D10E will actually be, either, or how it will perform. Rumours fly about the web with abandon, though, and recently a VR-Zone forum post claimed Nvidia's new chip was capable of over 16,000 in 3DMark06 when running at 2,560 x 1,600. But it turned out to be a fake.

Other GT200 rumours worth mentioning for the sheer speculative fun of it are that the GPU has 64-bit floating point operations (FP64), 64 raster operations (ROPs), 256 texture address units, and 1,024 shader units. It has been suggested that the core will run at 1.2GHz, the shaders at 2.8GHz, and the memory at 3.4GHz. So you really might need that 1000W power supply after all...

If only half of this is true, the GT200 will be a major leap forward in performance. Most interestingly, it looks like it won't be multi-GPU as standard - it will still be a single monolithic core, albeit with heavy parallelism internally.

ATI's RV770 is allegedly due in May, and will only perform up to 50 per cent faster than current single-GPU cards. So this could be why ATI has jumped on multi-GPU in the shape of the Radeon HD 3870 X2 faster than Nvidia - it needs the extra horsepower. The company is also widely expected to be putting more than two GPUs on a single card in the near future.

The graphics business has become a game of poker, with Nvidia keeping its hand hidden until it sees what cards ATI is holding. The GT200 could be a straight flush, or it could be a raw deal. But we will probably have to wait until November to find out for sure.

As I suspected last week, Nvidia's new high-end graphics card, the GeForce 9800 GTX, is essentially just a GeForce 8800 GTS with faster clocks.

GeForce 9800 GTX uncovered

Details of the 9800 GTX have been leaked by VR-Zone, and a few other sites. The card's GPU core is clocked at 673MHz with 1,683MHz shaders, and although VR-Zone doesn't know the memory clock, another site, Expreview, managed to get its hands on one for benchmarking. It puts the memory at 1.1GHz (2.2GHz effective due to dual-channel configuration).

Compare this to the GeForce 8800 GTS (the most recent 512MB-equipped version, that is). This has a 650MHz core clock, with 128 shaders running at 1,625MHz and an effective memory clock of 1.94GHz. So unless NVIDIA has somehow managed to unlock some more shaders in the 9800 GTX, it's shaping up to be a mere baby step ahead of GeForce 8.

The 3DMark06 score that Expreview quotes of just over 14,000 is only about 25 per cent ahead of the GeForce 8800 GTX when it was first launched. And that would have been with a dual-core processor rather than quad-core. So the difference on the same platform will be even less.

No wonder Nvidia chose to focus on the 9600 GT as the vanguard of the GeForce 9 series. Thanks to its faster clocks and doubling of the shader allocation from 32 to 64, the 9600 GT really is a leap forward from its 8600 GT predecessor.

Is Nvidia saving the big stuff for another day?

Perhaps this explains the strange strategy with the launch of GeForce 9 series. GeForce 7 wasn't a huge increment over GeForce 6 either, at least not in terms of technology. GeForce 7 had a few more vertex and pixel shaders and texturing units, plus a mildly faster core clock. But it was a development of the same basic theme.

The GeForce 8, with its adoption of unified shaders, was the sea change. After that, it's not surprising we're back to incremental steps for a while. The G90 series chips, which first arrived in the 512MB GeForce 8800 GTS, are essentially the G80 of the GeForce 8 series, but made with a 65nm process rather than 90nm. Apart from some major improvements to the PureVideo features, there are few significant architectural changes.

The one thing the 9800 GTX will have going for it over its direct predecessor, the 8800 GTX, if not the 8800 GTS, will be power consumption. Thanks to the reduction in chip size, the GeForce 9800 GTX can run faster and consume the same or less Watts.

Even if you don't care about saving the planet, complete speed freaks will be able to create a more stable Tri-SLI setup. When your graphics cards on their own require a peak power close to 600W, losing 100W or more will really help bring PSU requirements back into the realms of sanity.

But otherwise the Nvidia GeForce 9800 GTX doesn't look like it's going to set the world on fire, although the dual-chip 9800 GX2 version might be more to write home about. It looks like Nvidia is milking the technology it developed for G80 for all it can, and holding back its truly new developments for another time. With Intel's graphics now allegedly waiting in the wings, perhaps Nvidia is keeping its trump cards close to its chest for a future battle.

It's playing the numbers game again. This time, it's Nvidia's turn to bring confusion to the computing world. In what the manufacturer has called "one of the largest generational leaps in company history", the launch of the GeForce 9600 GT sounds like it should be the next benchmark in performance, showcasing a new graphics processing architecture.

But it isn't.

The GeForce 9600 GT is actually a midrange card, which is no bad thing in itself. Bringing performance to the masses is very commendable. And there's nothing particularly wrong with launching a new GPU architecture into the middle of the market rather than the high end, either.

However, the GeForce 9600 GT isn't a new architecture. In fact, other than the price, I'm struggling to see why it should be the first GeForce 9. The new card uses a GPU called the G94, which is essentially half of the G92 launched in November 2007.

Where the G92 has 128 stream processors, the G94 has 64. Both operate at the same core and shaders clocks of 650MHz and 1,625MHz respectively, although the G94 has a slightly slower memory clock (1.8GHz versus 1.94GHz).

Déjà vu all over again

What makes this all so confusing is that when the G92 arrived, it wasn't the first incarnation of GeForce 9, it was another GeForce 8. Its launch was one of the most bizarre events in computing of 2007, which still makes me wonder if someone in Nvidia's marketing department either has a very dry sense of humour or a strange penchant for unusual forms of mushroom.

The company first launched a card called the GeForce 8800 GTS at the end of 2006 as the vanguard of the 8000 series. This was the 'affordable' version of the flagship GeForce 8800 GTX. A few months later, another version with half the memory arrived, and ended up being the really big seller.

But then Nvidia launched a product with the same GeForce 8800 GTS name again at the end of 2007. This time, however, it was using the newer G92 GPU rather than the G80 of the original version. Why Nvidia wanted to launch a significantly faster card with the same model name as its predecessor still remains a mystery.

Now, with the introduction of the GeForce 9600 GT, we have the GeForce 9 launched in a cut-down version of the chip inside the GeForce 8800 GTS. Okay, so car manufacturers often put an older engine in a supposedly new car design. But this is the first time I've seen a feature-reduced version of an existing chip marketed as the new latest and greatest in the computer market.

It is going to be a lot faster than its 8600 GT predecessor, just as the marketing alleges. After all, it has twice the number of stream processors. That is quite a leap forward. But in reality Nvidia should have launched the G92 update of the 8800 GTS as the first 9 series - because that is effectively what it is. Then things would have made sense and logic would have remained in the universe.

Maybe it should have been the 9800 GTS instead, although the various flavours of the 9800 name are now widely expected to be based around faster versions of the G92, and a dual-GPU configuration. These are due in the next month or so.

Nvidia has changed the internal GPU code names it uses as well, moving from the Gxx nomenclature to ones like D8P or D9E (which are G92 and, er, well, a faster version of the G92 respectively...). So maybe someone at Nvidia has just gotten confused about which chip is in which model name. I don't blame them. I certainly am.

The AMD Phenom has hardly had an auspicious start. We were shocked when we learned the release clock frequency would be a lowly 2.3GHz. But then it turned out this was partly caused by a specific problem with the revision of the processor design available at launch.

Now a fixed version looks set for April. Will this provide the boost that AMD's Phenom so desperately needs?

TLB problem - TLC needed

According to AMD, its reasons for holding back the expected 2.4GHz spin of the Phenom at launch was due to a problem with the Translation Lookaside Buffer (TLB), although there are numerous conflicting statements from the company about this.

The TLB is part of the Level 3 cache logic - a key element of the Phenom's design. The TLB deficiency wasn't expected to cause instability very often, but clearly it would be frequent enough for some concern.

Originally, AMD stated that this problem only affected Opterons, but now it has become clear that all Barcelona-core processors up to and including stepping B2 have the issue, also known as 'Errata 298'. All processors have errata, which BIOS tweaks, microcode updates or even operating system software patches can counteract.

In the case of AMD's TLB problem, a BIOS update has been able to prevent the instability. However, the BIOS fix also has a downside. It causes a substantial performance hit, with AMD itself quoting this as 'around 10 per cent', and some putting the deficit at as much as 20 per cent with some applications.

AMD's B3 stepping

This is where the B3 stepping comes in. The new revision of the Barcelona core includes a microcode update to fix the TLB issue. So these new chips won't incur the performance hit of earlier versions. When AMD needs every per cent of extra speed it can get, this is very good news.

Quite how the TLB problem caused the delay to the 2.4GHz Phenom, which is still on pre-order on most websites, remains a little unclear. It isn't because the TLB issue directly prevents higher frequencies, but appears to be because AMD is waiting for the B3 stepping before release. So everything from Phenom 9700 upwards will be the new B3, and the current chips will be replaced with it too.

However, unlike Intel's stealthy introduction of its G0 stepping last year, AMD will make it obvious you're getting a B3 chip by adding 50 to the number, so the Phenom 9500 will be replaced by the 9550, and the 9600 by the 9650. This isn't going to help the number confusion we have already noted. Just bought an Intel QX9650? Soon you will be able to get an AMD Phenom 9650 to go with it. But at least you can easily avoid the faulty earlier stepping.

Another reason not to buy a Phenom

Only the most unrepentant Intel fanboy wouldn't want AMD to be producing processors which are at least competitive with what Intel has to offer. The rise of the Athlon 64 injected a considerable amount of enjoyment into hardware enthusiasm. After all, the computer industry thrives on competition.

The problem for AMD is that, even if you were thinking of getting a Phenom, you've now got a good reason to put off the purchase that little bit longer. You might as well wait for the B3. It probably won't get AMD that much closer to Intel just yet, but at least it's what the Phenom was supposed to be at launch.

It had to happen sooner or later. Futuremark is finally going to be producing its own games. This is hardly a surprise. In fact, it's more surprising that it has taken so long. Futuremark's new Games Studio plans to combine 'highest quality game play' with 'blow-your-socks-off visuals'. But can it really make a difference in what is now a very competitive games market?

Peeling back the Onion

Founded in 1997, Futuremark has been producing cutting-edge games engines ever since it released 3DMark99. This was the first really serious benchmarking tool for Direct3D, when it still wasn't certain this would be the de facto standard for PC gaming.

Indeed, 3DMark has always been ahead of the market. After all, if a new benchmark is to remain a relevant test of 3D performance for a year or two, it needs to access the features of games and hardware which haven't become mainstream yet.

For example, 3DMark99 tested fixed function vertex transform and lighting, and led the way with multi-texturing. 3DMark2000 added hardware-accelerated transform and lighting before any game took advantage of it. 3DMark2001 debuted shaders, with extensive use of Vertex Shader 1.1 and the first use of Pixel Shader 1.1 in its last test routine.

Although 3DMark03 stuck mostly with Shader Model 1.1, it contained the first use of Shader Model 2, again in the last test. And so on, until we get to 3DMark Vantage, due later this year. Although this will stick with DirectX 10, Vantage will exclusively require Windows Vista - a bold move considering how many gamers are sticking with XP for now.

Looks aren't everything

So Futuremark already has a clear strength in getting to grips with the latest DirectX technologies. You would expect its Games Studio to produce titles showcasing what each new generation of the API can do, just as is already the case with 3DMark.

But leading the way with visuals is not all there is to a successful game. Just look at the failure of Crysis and contrast that to the success of the Wii. Crysis may look great, but the hardware required to achieve this feat has actually been a disadvantage.

In contrast, the Wii is hardly a cutting-edge console in terms of graphics performance. In many ways it's a generation behind the Xbox360 and PlayStation 3. Nevertheless, Nintendo had the runaway hit of the Christmas period. Why? Because its console design successfully pinpointed a family gaming market, which doesn't care so much about the visuals looking more like movies.

Another classic example is the Lego Star Wars game franchise. Not exactly the most visually amazing series in the world, but the most amazing fun. Its balance of problem solving, nostalgia and humour has made it one of the most successful games ever on any platform.

To be a hit game developer, Futuremark's Games Studio needs not only to produce titles which look great - they need to play great too. It's obvious from the company's benchmarks that they know how to do the first of these. Just look at the recently released screenshots of the new 3DMark Vantage.

But whether they can actually produce games which are fun to play is another matter. For that, each title will need an enticing premise, a good plot, pacing so that it's neither too hard nor too easy... all things which have very little to do with mastery of DirectX technology.

A curious thing is happening in the processor business. Although AMD's Phenom hasn't exactly taken the fight back to Intel in the benchmarks, its numbering system is strangely easy to mix up with the latest generation of Intel quad-cores. Both are called 9000 series.

Just to obscure things even more, it looks like the triple-core Phenom will be the 8000 series - the same as the 45nm Wolfdale Core 2 Duo. Those of us who keep a close eye on the processor business will know the difference between an Intel Core 2 Duo E8xxx and an AMD Phenom 8xxx. But less savvy consumers are likely to be misled by what these model numbers actually mean, and shop assistants in retail stores probably won't help much either.

From clocks to model numbers

Once upon a time, processor names were easy to understand. The brand and model name told you the architecture (Pentium, Athlon...) and the number was the frequency of the processor. More gigahertz meant more speed.

Then AMD realised that this wasn't doing it any favours. Clock-for-clock, its processors were punching above their weight. The result was the equivalency numbers like 2500+ which have adorned AMD's consumer processors since the launch of the Athlon XP.

Intel also switched to model numbers in mid 2004. This was the final death knell for marketing processors by their clock frequency. It was also Intel's first admission that the Netburst Pentium 4 wasn't going to fulfil its original goal of 25GHz - particularly as Intel's own Pentium M was also outperforming it clock-for-clock as much as AMD's Athlon 64.

Model behaviour

When Intel and AMD were using very different model numbering systems, this wasn't a problem. After all, a Pentium 4 670 and an Athlon 64 3000+ are obviously incomparable, just as only a fool would confuse a BMW 530 with a Rover 75.

But when faced with a choice between a quad-core AMD Phenom 9500 or an Intel Core 2 Quad Q9450, a PC novice is sure to be scratching their head. They could easily assume the Intel processor is overpriced and the AMD one better value. But we know a 2.66GHz quad-core processor from Intel will run seven rings round a 2.2GHz one from AMD. For many, that's well worth the extra £100.

When the Phenom 8000-series triple-core CPUs arrive in the middle of the year, there will potentially be Intel Core 2 Duo E8500s up against AMD Phenom 8500s. Surely, you'd want the Phenom, because you get an extra core - right? You might, as it could have some advantages. But most people would probably be better off with a faster clocked dual-core from Intel. The similar model names will only cause punters to think there is some similarity when there isn't.

Or is this precisely what AMD hopes? Its graphics division is giving as good as it gets at the moment, with the new Radeon HD 3870 X2 arriving in reviewers' hands before Nvidia's GeForce 9800 GX2. But the processor division is going to be an exercise in damage limitation and saving face for most of 2008, at least until AMD can get its own 45nm production process off the ground, or even until it brings out Bulldozer.

So maybe AMD is hoping a little confusion won't hurt its sales. Either way, the truth will be in the numbers.

Ray tracing has long been the Holy Grail of gaming 3D. It promises much more photorealistic output than current real-time 3D engines, and is favoured by the film world. But it's very processor intensive, too much for today's hardware. Fortunately, a recent move by Nvidia could make GPU-based ray tracing a reality - and bring it to a game near you in just a few years.

Famous rays

This isn't the first time ray tracing has hit the news. Back at the Intel Developer Forum in April, Intel demonstrated real-time ray-traced Quake 4. But that was running on a pair of quad-core CPUs, and led a few pundits to argue that the end of the GPU was nigh - when processors had enough cores.

But just before Christmas, Nvidia quietly purchased a company called Mental Images. This has stoked interest amongst professional 3D designers, because Mental Images produces Mental Ray, the renderer used by a legion of professional 3D animation applications including Autodesk 3dsmax and Maya, and Softimage XSI. So now Nvidia owns the most widely used professional ray-traced rendering engine.

It has another implication as well, though. Nvidia hasn't explicitly confirmed this, but it has hinted that hardware acceleration for Mental Images' rendering technology is a future goal, and a primary reason for the purchase. In other words, running ray tracing on GPUs.

The first aim of this will be accelerated rendering for professionals. But once that has been achieved, it's not exactly a leap for ray-traced games to run on consumer graphics cards as well. After all, the GPUs inside Nvidia's professional Quadro FX cards are basically the same as those in the GeForces - just optimised for OpenGL, with tighter quality control and much longer warranties.

Even more importantly, the next generation of consoles are expected to arrive in 2011 or 2012. With HDTVs likely to be well bedded in by then, the next generation of consoles need the photorealism to match the HD movies everyone will supposedly be watching by then. Real-time ray-traced graphics would provide that.

End of the line for the z-buffer

For most of the history of GPU acceleration, a system called z-buffering has been used to calculate which objects are visible. Each pixel is given a depth value, which is used to work out which ones are in front of which. Lighting and shadow effects are created by performing active operations on the textures applied to objects - the pixel shaders we are all now familiar with, even if we don't know precisely what they do!

Ray tracing attacks the problem of visibility in an entirely different way. Since you can only see things which have light falling on them, it traces the rays from light sources as they bounce off reflective surfaces. This way, it figures out what an observer can see, and how brightly it is lit. Best of all, ray tracing calculates shadows accurately as part of this process. Hence, the end results look more like reality - because they have been calculated in the same way reality actually works.

It's bizarre how little has been said about Nvidia's move to buy Mental Images - virtually nothing outside the professional market. Perhaps this is because the purchase happened just before Christmas, and perhaps it's because everyone is assuming it only affects 3D professionals.

But the potential impact of ray-traced rendering on GPUs is enormous for consumers too. It could quite literally change the face of gaming as we know it.

In a week or so, Nvidia will be launching what it is claiming will be a revolution in integrated graphics. It doesn't sound that interesting for those of us who like the fastest hardware available. But Nvidia's new integrated graphics is promising benefits even for the highest-end systems. I'm not so convinced.

Dissing integration

As most notebook users will agree, having graphics acceleration integrated into your motherboard chipset is not a recipe for state-of-the-art gaming performance. Although 3D hardware has been a feature of integrated graphics platforms for some years, and a requirement for running Windows Vista Premium, it is never anywhere near the latest discrete graphics cards. You might be able to play games from a few years ago on the latest integrated graphics, but forget anything recent unless it's a new version of solitaire.

According to Nvidia, however, its new integrated graphics will allow PCs without discrete graphics to play recent games at a comfortable resolution. But it's still hard to believe it will be right up there with the latest 3D adapters. After all, Nvidia wants to carry on selling discrete graphics cards.

Nvidia's new pick'n'mix

Fortunately, Nvidia has some other tricks planned for its new integrated graphics, which are intended to keep you buying separate cards anyway. To push this, it intends to roll out integrated graphics across its entire range of motherboard chipsets - even the high-end ones. The reason for this is Hybrid SLI, something we've reported on in the past. Its potential benefits are two-fold.

One feature of Hybrid SLI is a reduction in power consumption, which Nvidia is calling Hybrid Power. This will be particularly useful on a notebook. For a portable, turning off a greedy discrete graphics chip in favour of an integrated one when 3D grunt isn't required is likely to have a very significant effect on battery life.

On a desktop, the power savings will be even more pronounced. Where a top-end discrete graphics card can consume as much as 150W, the option to turn this off when not required will have an even greater impact on your power bill than for a notebook.

However, the Jeremy Clarksons of hardware enthusiasm will scoff at this, and wonder why any high-end PC user would care about power consumption. If you run a LAN gaming centre, maybe you could save a lot of money, but I can't see individual users being that bothered unless electricity prices skyrocket. And it only works with Nvidia's latest graphics cards, released this year.

Fortunately, there is another potential benefit, which Nvidia is calling GeForce Boost. This is the feature early reports of Hybrid SLI focused on - the idea of running SLI across discrete and integrated graphics. There's no point doing this if the integrated side of things is too weedy. But since Nvidia is claiming its new version is a revolution in performance, and GeForce Boost will work with the likes of a GeForce 8400GS (a £30 card) or 8500GT, a cheap PC could provide half-decent DirectX 10 performance with Hybrid SLI.

So Nvidia's new integrated graphics promise some interesting technology, which could have decent benefits for the low to mid-range. But I'm not so convinced high-end enthusiasts will be too bothered about saving a few hundred watts when they aren't using their PCs for gaming. There is a reason why they are called power users, after all.

You don't expect much that is newsworthy to happen in computing over the Christmas period. But a New Year's Eve treat leaked out regarding ATI's R700. Apparently, components of the next generation of Radeon are already entering production. This could mean we see ATI's new graphics cards by May.

After all the disappointing news from AMD, it's a relief to see something this positive from its bought-in graphics department. ATI has had almost as tough a time against Nvidia as AMD's processor division has been having against Intel for the last couple of years.
The rot set in when R520, the first generation of the Radeon X1000 series, arrived a good six months later than planned. So it met an already mature Nvidia GeForce 6000 series, with drivers optimised from numerous revisions. Launching a new product which doesn't beat your competitor's existing one has never been a recipe for praise.

Because of its tardiness, the Radeon X1800 was superseded by the X1900 just three months after its arrival, which didn't impress ATI fans. And that soon had Nvidia's GeForce 7000 series to contend with. So even though the X1900 put up a fight, its lead was short-lived.

The R600, or Radeon HD 2900XT, was also around six months later than it should have been. This again left the market open for Nvidia, in this case with its GeForce 8000 series. When it did finally arrive, the new Radeon HD could only compete on price - like R520 all over again.

So the last three years, whilst not a complete disaster for ATI, have hardly been according to plan.
But things have started to look up more recently. The die shrink of R600, the RV670, has made the Radeon HD 3000 series exceedingly good value, even if it hasn't taken back the performance high ground just yet.

And with the R700 now possibly on the cards in less than six months, ATI could finally have the high-end competition it needs. From what we know of the new graphics architecture, it appears to be taking advantage of multi-GPU technology by combining up to eight cores on a single card. This could even be the first graphics actually capable of coping with Crysis!

In contrast, there is virtually no word on what Nvidia's next generation will consist of, although there are rumours that it is likely to hit the market in March. Nvidia is traditionally better than ATI at keeping loose lips from sinking ships. So the G100, or D9E (standing for ninth generation enthusiast), could well pack in some technology we haven't yet heard about.

Before then, the D8E will allegedly shoehorn two G92s - the GPU in the new version of the GeForce 8800GTS - onto a single card rather like the 7950GX2 of a few years ago. So, with unified shaders the graphics technology of the foreseeable future, it's safe to predict that Nvidia's D9E will either pack more stream processors or multiple GPUs, and probably both.

Even though D9E is currently expected before R700, things still look hopeful for ATI - if it can get it's new hardware out the door in May. So let's hope R700's production is successful, and the product really does arrive when expected. It may be a couple of months later than the current predictions for Nvidia's latest and greatest. But if it really does pack up to eight GPUs at least as powerful as the RV670 onto one card, it could be the point when ATI returns to the high ground once more.

For over a year, virtually any mention of Crysis would get awebsite attention. If you happened to have exclusive game footage previews, youcould expect your servers to be hammered to high heaven. But now the game hasbeen out for a month, it's not proving to be the smash hit that was expected.After all the hype, why has Crysis sold so badly?

In-console-able

There's nothing intrinsically wrong with Crysis. In fact, ithas been winning critical acclaim left, right and centre. IGN gave it 9.4, PCZONE gave it 9.2,and that seems to be the theme. Sure, once you get beyond the novelty of thenanosuit, it's basically a game about sneaking around to find the leastexpected place from which to kill people, rather like Far Cry. But then so arelots of successful first person shooters.

So what has gone wrong? One obvious problem is thephenomenal hardware requirements. When even three graphics cards running inparallel can't cope,you know a game is a little ahead of its time. Imagine spending the £1,200 pluson Tri-SLI and still not being able to play Crysis as intended?

Even more ominous, however, is the fact that for theforeseeable future Crysis is a PC game only. This is also linked to itshardware requirements, as consoles traditionally lag behind the abilities ofPCs.

So whilst it was big news for PC gamers, the much larger consolemarket is not interested. They're scrabbling to find that last remaining copyof Guitar Hero III. Crysis may have been hyped amongst PC users, but it hasn'thad the mainstream media coverage a hit console game receives.

The Unreal deal

Which brings us to the other surprising loser at the retailoutlet - Unreal Tournament 3 - although its lack of success can be put down torather different factors than Crysis. UT3 is not a major graphics resource hog.A single high-end 3D card should be perfectly sufficient to play it at 1,920 x1,200 with high quality settings on your prized 24in TFT.

You do have to question its entire existence, though. Theoriginal Unreal Tournament was an absolute classic. It was the first game toembrace the new era of network play wholeheartedly, and jettison any pretenceof having a plot for single players. Sure, there was still an extensivesingle-player campaign - but it felt more like training to keep you sharp forthe next time you were on a LAN with your mates.

Unfortunately, UT3 hasn't significantly moved on from there,other than graphically.

I'd place Unreal Tournament second only to Counter Strike inthe classic LAN top ten of all time. So the same game with better graphics andphysics doesn't sound like such a bad idea. After all, CS: Source could hardlybe called unsuccessful. Those who have been gaming less than ten years mightnot have played the original Unreal Tournament anyway.

But compared to the humorous gameplay of Team Fortress 2 orthe life-sapping immersiveness of a MMORPG like World of Warcraft, UT3 mustseem decidedly old school. No wonder it's not selling in buckets.

At least UT3 is already available on PS3 and will be comingout for Xbox360. But a console has never been the best thing to play deathmatchFPS games on - there just isn't the control or responsiveness. I can't see itcompeting with more story-led titles like Assassin's Creed.

So Crysis aimed so high it has managed to shoot itself inthe foot, and UT3 is like the remake of the Texas Chainsaw Massacre - bettereffects, but ultimately pointless.

In a market which is dominated by consoles,focusing on what's best for PC gamers alone might not make the most successfulPC game. After all, the number two PC game in November was Call of Duty 4 - butit was the number one console game as well.

What are we to make of a game that is too intensive for eventhe fastest graphics platforms on the planet? German website PC Games Hardwarehas managed to get its hands on an early sample of Nvidia's Tri-SLI.And, amazingly, even a trio of Nvidia Geforce 8800 Ultras isn't enough to playCrysis at maximum quality settings at HD resolution. Is Crysis too demandingfor any hardware currently available?

The Crysis of framerates

Before taking a look at PC Games Hardware's tests,