All Graphics cards Feeds

Nvidia: Current CPU-GPU balance in PCs is 'obscene'

Jeremy Laird — 1210341913

Just forked out for a new rig with a fast processor on board? Then Nvidia has some very bad news for you. Your PC is "obscenely" imbalanced thanks to an overpriced, underperforming CPU - probably courtesy of Intel.

It's just the latest salvo in the burgeoning war of words between Nvidia and Intel this year. But what exactly is Nvidia getting at? Talking to TechRadar earlier this week, Nvidia's VP of Content Relations Roy Taylor outlined a developing strategy for leveraging Nvidia graphics technology to accelerate a wide range of PC applications. Very soon, the world will discover just how pathetic conventional CPUs really are.

If Taylor is correct, the initiative will deliver a massive, unprecedented boost in PC performance. We're not talking the 2x or 3x boosts in performance that the PC industry delivers on a regular basis. It could promise as much as 20x or even 100x the performance of todays multi-core CPUs. Yikes.

CUDA cometh

The basic premise is the use of Nvidia's CUDA programming platform (itself closely related to the C programming language) to unlock the increasingly programmable architecture of the latest graphics chips.

On paper, it's extremely plausible. In terms of raw parallel compute power, 3D chips put CPUs to shame. A good recent example is the new room-sized, high density computing cluster installed by Reading University.

Designed to tackle the impossibly complex task of climate modelling, it weighs in at no less than 20 TeraFlops. That sounds impressive until you realise that just a single example of Nvidia's next big GPU, due this summer, could deliver as much 1TFlop. So, a few four-way Nvidia GPU nodes will soon offer the same raw compute power as a supercomputer built using scores of CPU-based racks.

General purpose GPU

A little bit closer to home, one of the early applications Nvidia is promoting as a demonstration of the general purpose prowess of its GPUs is a video encoding application known as Elemental HD.

Downsizing a typical HD movie for an iPod using a conventional PC processor can take up to eight hours or more, even with a decent dual-core Intel chip. Nvidia says the same job can be done in just over 20 minutes on an 8800 series Nvidia graphics board.

"When you look at the question of whether you should transcode video on a GPU or CPU, when you consider it in performance-per-buck terms, it's currently obscenely the wrong way round," Taylor says.

And the solution is simple enough. Don't spend any more money overall. Just spend a little less money on your Intel CPU and a little more on your Nvidia GPU.

Hardware PhysX

What's more, Taylor says plans to support the recently acquired PhysX physics-simulation engine on Nvidia's GPUs are also nearing launch. Before the end of May, a total of eight games with GPU-based PhysX are due to announced. 30 to 40 such titles will be available this time next year.

So, that's it then. The game is up for the CPU and Intel alike? Not so fast. For starters, there's a good reason why CPUs don't deliver the raw compute power of contemporary GPUs.

CPU cores are big, complex beasts, designed to turn their hands to almost any task and make a decent fist of it while not excelling in any one area. GPUs, even the most recent and programmable examples, are still a lot less flexible. When they're good, they're great. When they're not, well, they simply won't do the job at all.

"At the moment general-purpose GPU applications are admittedly very high end. But increasingly people are asking why are scientific research industries including medicine and climate modelling are using GPUs," Taylor says.

The answer is the unbeatable bang-for-buck performance ratio that GPUs deliver. Taylor reckons Nvidia has a large number of partners with consumer-level applications lining up to key into its GPU technology. Several are due to be revealed later this summer.

Waiting game

Until then, however, it's impossible to say whether the benefits will be as spectacular as Nvidia claims. Likewise, we'll have to wait and see just how smoothly it all works. The only non-3D consumer application for GPUs that has been widely tested on the market so far is video decode assist. And that has been a distinctly hit and miss affair.

But even if Nvidia can deliver reliable, transparent hardware acceleration for a wide range of applications with its GPUs, it will still have a huge fight on its hands from Intel.

Intel's intriguing new GPU, known as Larrabee, is due out in late 2009 or early 2010. Apart from the fact that it will be based on an array of cut-down X86 processor cores, little is known about its detailed architecture. But as Intel's first serious effort to compete in the GPU market, it's a game-changing product.

For Taylor, of course, the Larrabee project merely confirms that the GPU is where the action is. “Why does Larrabee exist? Why is Intel coming for us? They're coming for us because they can see the performance advantage of our GPUs,” Taylor says.

He's probably right. It will be a fascinating contest.

Nvidia: Only Brits care about ATI

Jeremy Laird — 1210332780

No one cares about AMD or its graphics subsidiary ATI anymore. Well, apart from we Brits, that is. Then again, nobody loves rooting for a hapless chump quite like the British, either.

At least, that's what Nvidia spinmeister extraordinaire Roy Taylor reckons. Taylor is actually Nvidia's main man for Content Relations Stateside. But he has been kind enough to hop the pond to provide some insight into Nvidia's plans and prospects.

Insight such as this: "The UK is the only place in the world where anyone talks about AMD or ATI". To prove his point, Taylor went as far as showing us a graph representing the entire market for GPUs last year, both discreet and integrated. Astonishingly, ATI did not feature at all.

Where was ATI among the 366 million graphics chips which had apparently been split exclusively between Nvidia and Intel in 2007? "No one cares," Taylor says.

ATI who?

Well, to be more accurate, Nvidia didn't care enough to include ATI in that slide - Taylor did have another showing ATI with 18 per cent of the market, a figure apparently so small as to render it effectively non existent.

Indeed, Nvidia may care just a little more than it lets on, since Taylor was keen to press home the point by quoting two independent surveys (Valve's Steam users and the YouGamers community) which put ATI's market share for DirectX 10 compliant GPUs in the 12-13 per cent range. Ouch.

With what is shaping up to be a killer new GPU due out this summer and ATI still struggling to land an significant punches, there's a real swagger to Nvidia's gait of late. And yet its merciless dismissal of the old enemy ATI is only half of the story.

For Nvidia is now stalking a much bigger beast. The real threat over the next five years will almost definitely come from Intel. Especially now that Intel itself has staked a claim to Nvidia's graphics turf with the upcoming Larrabee chip.

And wouldn't you know it, Nvidia - viaTaylor - has that angle covered too. But that's a whole different ball game. In fact, it's one you can read about later on TechRadar.

Nvidia subtly slides out 9600 GSO deets

Dan Grabham — 1209398988

Offering none of the usual buzz surrounding its announcements, Nvidia has subly outed a new low-end card by just posting the details on its website. The GeForce 9600 GSO has been rumoured for a little while, and its stats reveal its lowly position at the base of the series underneath the 9600 GT.

So what have we got then? 384MB of onboard memory on a 192-bit bus, for starters. The GPU is clocked at 550MHz, the memory at 800MHz and the 96 shaders at 1375MHz. Because of those stats it looks like the card is based on the same wound-down G92 chip as the 8800 GS.

The GS sufflix may well have been replaced by GSO to make things seem, ahem, a little simpler. After all, GT and GS are quite similar yes?

Mind you, the cheapo pixel cruncher hardly skimps; there’s still SLI capability as well as support for PCI Express 2.0, Nvidia’s PureVideo HD for Blu-ray playback and DirectX 10. Not bad in a card that will sell for around the £100 mark.

The GSO will go head to head with the forthcoming Radeon HD 3830 low-end series from ATI. The GSO looks like it will be on shelves in a couple of weeks – keep ‘em peeled.

AMD preps graphics bomb for Nvidia

Jeremy Laird — 1209139198

Rumours involving the next 3D chip from AMD's graphics division ATI are reaching critical mass. The new GPU, codenamed RV770 and expected to be marketed as the Radeon 4800 series, promises to be the fastest GPU ever.

It could be unveiled in a matter of weeks

If true, it will be a long overdue victory for ATI and a timely blow for its increasingly smug rival, Nvidia. The weight of opinion on RV770 now points to a chip with 480 stream processors, 32 texture samplers and 16 render output units. All told, RV770 is expected to weigh in at over 800 million transistors.

The most complex GPU currently offered by ATI, the Radeon 3870, squeezes 320 stream processors, 16 textures and 16 render output units into a 666 million transistor budget. It most other regards, RV770 is expected to carry over the DirectX 10.1-compliant architecture of the existing Radeon 3800 series.

AMD's mean 55nm machine

It will likewise be bashed out on the same 55nm silicon node by production partner TSMC in Taiwan. What's more, the rumours suggest it will still use the 3800's 256-bit memory bus. However, it's expected to up the effective data rate to nearly 4GHz courtesy of some fancy new GDDR5 memory chips.

It's a bigger, beefier version of the Radeon 3800 series rather than an all-new chip, in other words. Based purely on the raw architectural layout, RV770 ought to be around 50 per cent quicker than the best single-chip Radeon 3870 boards available today. However, the 4800 will also bring a boost in GPU core clockspeeds.

It's thought the headline frequency will increase from 775MHz to around 850MHz. It's also just possible ATI will copy Nvidia's split-frequency-domain approach and clock the RV770's stream processors even higher. If so, look out for stream processor clocks somewhere north of 1GHz.

Teraflop chip

All in all, it's predicted RV770 will just breach the magical 1TFlop barrier for raw compute performance. It should also be easily a big enough bomb to flatten Nvidia's best single-chip graphics card, the GeForce 9800 GTX.

And that in itself will be awfully big news. ATI hasn't led the market since the introduction of the Radeon X1950 series way back in August 2006. Nearly two years is an awfully long time to be second best.

What's less clear is whether single-chip RV770 boards will have enough fire power to beat the best existing dual-GPU boards from both ATI itself and Nvidia. However, that's a somewhat moot point. For starters, ATI is also prepping a dual-chip variant of the 4800 series, codenamed R700.

Moreover, dual-chip graphics cards tend to be rather unreliable beasts. We've little doubt enthusiasts will welcome a new single-chip card that lifts the bar for performance.

What about the green team?

All of which just leaves the question of how long it takes Nvidia to respond. The configuration of Nvidia's next heavyweight GPU, codenamed GT200, is still the subject of some debate.

However, a German language website recently claimed that Nvidia has confirmed the chip will pack one billion transistors. If so, it will clearly be more complex than ATI's RV770.

But when will GT200 appear? Given the recent roll out of the 9800 GTX, it's hard to see Nvidia releasing another new high end graphics chipset before late summer.

For now, AMD will only say that RV770 is out soon. We'd put good money on a late May launch date for RV770. But even a few months at the top would do wonders for both the image of the ATI brand and morale at AMD and ATI in general.

Our verdict: Nvidia's 9800 GTX and GX2 on test

Dave James and James Rivington — 1208270155

Just over eighteen months ago the much-heralded age of the DirectX10-capable graphics card dawned, as the supreme G80-powered GeForce 8800GTX dropped into the TechRadar office. Six months later came the updated 8800 Ultra, a card that has remained Nvidia's top end offering... until now.

We've had to wait 12 long months for the refresh, during which time we have been treated to a mass of mid-range cards, including the excellent 8800GT - Nvidia's first card with a 65nm core.

But still, it's been a long time coming for the 9800 GTX and GX2.

Long time passing

Both new cards are powered by the same 65nm G92, a core that is itself now six months old. And it's the first time that Nvidia has released a brand new family of top-end cards based on old architecture. Replacing the 8800GTX and Ultra is a necessity as far as furthering the Nvidia brand is concerned, competition-wise though it's less of an issue. AMD still hasn't managed to create anything to seriously outperform these year-old cards, so is the lack of a new core an acknowledgment that Nvidia only has to turn up at the track to win the race?

The GTX version of the 9800 card is a straight, beefed up version of the G92 with higher clock speeds across the board. While it shares the number of Raster Operators (ROPs) with the 8800GT, it does have the old GTX's complement of shader units at 128, giving it the necessary speed boost.

The GX2 follows the example of the old 7950GX2, strapping two G92-stuffed PCBs together. But this time both PCBs face into the same heatsink and are housed in a vaguely coffin-like surround. The clock speeds are slightly slower than the GTX, but a fair bit of optimising has gone into making this single-card SLI offering an impressive piece of engineering.

Swiss cheese memory

The first difference you'll notice when comparing the two new cards with their predecessors is the change in memory capacity. Both the 8800 GTX and Ultra had a 384-bit memory bus with 768MB of GDDR3, while the 9800s make do with the same 256-bit 512MB of memory that resides on the GTS and GT iterations of the G92-based 8800s.

Due to its two cores, the GX2 comes out tops in the memory bandwidth stakes at 128GB/s compared with the Ultra's 103.7GB/s, but the 9800 GTX lags well behind both of the previous cards. What this all means in real terms is that at the higher resolutions, and most especially with full screen anti-aliasing turned on, the new cards take quite a hit at the levels we were hoping these big-panel pixel pushers would excel at.

The differences between the GTX and GX2, and indeed the 8800GT, are slight; the GX2 simply relying on the brute force effect of the single card SLI factor. Where the difference between the two new G92 parts is most obvious though is the number of ROPs. The GTX is still hobbling along with 16, less than both the 8800GTX and Ultra at 24, but due to the doubling up, the 9800 GX2 has 32 ROPs. The difficulty is in knowing how much of a benefit this multi-GPU's extra ROPs gives us as opposed to the single card with 24.

Bigger, faster, stronger

So where do we find ourselves with the two new top-end cards? Well, mostly in the same place we were before to be honest. There's very little difference between this new set and the old, with the 9800 GTX being the biggest disappointment.

It struggles to find any space between itself and the 8800 GTX (which it's supposed to be replacing), and there's also the fact that you can still pick up the older card - with the extra memory, bandwidth and ROPs - for less than £200. In some places you can save yourself around £50 and come out with an equivalent, and in some cases, faster card. The march of progress seems to have stomped right past this iteration of the 9800, and here at TechRadar we might just have to plump for the original DX10 monster.

With regards to the GX2, Nvidia had to go down the multi-GPU route, not just to prove it could produce a functional version like AMD, but also to create a card that it could legitimately call the fastest graphics card around.

The final verdict...

Still, the memory constraints hold the GX2 back from being the superlative, stand out, top-end card de jour. On lower-res panels without silicon-melting anti-aliasing, it speeds ahead of the competition. Yet with all the bells and whistles cranked up to a deafening roar it struggles to break-even with the old 8800 Ultra. Again, if you shop around you can pick up an Ultra for around £350, and be fairly sure that your card will have drivers mature enough to cater for whatever you throw down its graphics tubes.

The long and short of it is that if you've got yourself an 8800GTX or Ultra, and felt that twinge of envy at the announcement of this new generation of top-end cards, then you can stop worrying. In fact, you can probably be down-right smug as your slightly geriatric cards are still more than capable of holding their own against these youngbloods. Til the GT200 comes out that is...

The full version of this review will appear in PC Format magazine issue 214 and will go on sale on 4 May.

Intel's UMPC is still a crap idea

Jeremy Laird — 1207269208

The smorgasbord of technology, slightly crazed enthusiasm and side-stepped questions that is the Intel Developer Forum is over. Here's our take on the key highlights of IDF: Shanghai and what they mean for computing's future.

It's a long way to go...

At this IDF, Intel fleshed out speeds and feeds for the first ultra mobile variant of its Atom processor. But that was about it regarding significant announcements relating to Intel's core computer chip activities.

For that, thank a conference call a few weeks prior to IDF Shanghai that dished out details on both the upcoming Nehalem CPU architecture and the Larrabee graphics chip. Not enough news, perhaps, in return for the typically monumental IDF carbon footprint.

The UMPC is still a crap idea

Intel may have given the UMPC concept a new name - the Mobile Internet Device, or MID for short. But it's still the same lame idea. Indeed, when pressed to discuss the pocket-busting proportions of the MIDs shown at IDF, the typical reaction of Intel suits was to get a little shirty. We've a feeling Intel knows that MIDs are not the answer.

Splitting the Atom

In the long run, we've little doubt that future revisions of Intel's Atom Centrino mobile platform and CPU promise revolutionary performance and software compatibility improvements for mobile digital devices.

But if the oversized MIDs on show are anything to go by, the first iteration is simply too bulky and power hungry to change the world. We therefore suspect Atom will actually make more impact on traditional form factors. Expect to see some astonishingly cheap Atom-based desktops and notebook systems before the year is out.

Larrabee remains a mystery

Intel may have come clean with its intentions for the mystical Larrabee processor. We now know it is unambiguously targeted at Nvidia and AMD's graphics boards. In fact, Intel reckons it will kill those cards stone dead.

And there's little doubt Intel is hoping Larrabee will drag the entire graphics industry over to ray-tracing tech. But we remain clueless as to Larrabee's true potential in traditional raster-based 3D rendering scenarios. And that is what will make or break it, in its initial iteration at least.

Critical mass for solid state drives

Intel is nothing if not the absolute bomb when it comes to economies of scale in silicon chip production. So, confirmation that it is planning the imminent roll out of a full family of solid state drives (SSDs) looks awfully promising in terms of driving prices down.

In partnership with STM Microelectronics, Intel will shortly unleash SSDs spanning a range of capacities from 32 to 160GB. With any luck, the spinning hard disk will be dead for all but seriously high capacity bulk storage by mid 2009.

Political posturing

Is Intel shameless enough to go on stage with a straight face and talk about an energy saving partnership with China... all the while with the Chinese reportedly commissioning a new coal fired power station every week? You betcha!

Intel was also happy to give stage time to the jocular but somehow slightly sinister looking Chinese Minister for Railways. We're not sure about his presentation skills, but he certainly looked the kind of guy who can make the trains run on time. So, it was no doubt a great week for Sino-Intel relations. Just don't mention capital punishment or widespread human rights abuse, umm-kay?

Intel puts TechRadar on a par with The Economist...

...And The Wall Street Journal, The New York Times and BusinessWeek. At least that was the august company we shared (in our outgoing Tech.co.uk guise) in the form of a foil festooned with quotes from leading publications during the opening keynote on day one. No other technology publications made the grade.

Shopping in Shanghai

...And finally, don't go shopping in China unless you are prepared to haggle for Britain – don't be surprised to see the Chinese shoehorn in haggling as a demonstration sport at Beijing this summer. And if you do manage to beat your opponent down to your target price, for goodness sake don't then change your mind and flatly refuse any purchase at all. You'll be lucky to get out of his market stall alive.

If that makes shopping in Shanghai sound like a chore, try this for size. You can still have a suit made to measure and delivered to your hotel room within 24 hours. And all for around £70. Long live the Chinese experiment in capital-socialist fusion!

Intel predicts death of conventional PC graphics

Jeremy Laird — 1207213956

PC graphics technology as we know it is dying, according to Intel. And its replacement will be a new age of visual computing enabled by Intel's Larrabee chip.

At least, that's what Intel bigwig Pat Gelsinger told attendees during his opening keynote at the Shanghai instalment of the Intel Developer Forum today.

It's a bold claim. But what does it actually mean? Gelsinger's specific beef is with the raster-based 3D graphics hardware that has dominated the PC industry for over a decade. That includes video chips from the two heavyweights of PC graphics, Nvidia and ATI, as well as Intel's own integrated graphics solutions.

Rethinking rasterisation

Raster chips focus on converting vector data into bitmaps that can be displayed on computer screens. It's been an extremely successful strategy to date, enabling ever more sophisticated graphics as raster technology has become more sophisticated.

But Intel reckons they're too rigid and lack scalability. The solution is a new "programmable" approach to graphics that majors on ray-tracing and model-based computing. And that solution will take the form of the near-mythical Larrabee co-processor.

Intel has been slowly trickling out info on Larrabee for over a year. But this is the first time it has unambiguously claimed it will usurp the prevailing graphics technology in PCs.

Visual computing

We already know that it is based on an array of approximately 16 small X86 execution cores, and is largely instruction-set compatible with Intel's existing PC processors.

To that Intel says Larrabee adds a new high-performance vector engine. The result is teraflops of processing power on a single chip, photorealistic graphics and a more immersive user interface. All of which adds up to what Intel is calling "visual computing".

Bigging up the Larrabee chip, Gelsinger said that in his 30-year career he had never been involved in a chip development programme that had generated so much enthusiasm from software developers.

What about software support?

He also emphasised that Intel is working hard to support developers by creating various tools and tuning libraries. Nevertheless, one of the big questions surrounding Larrabee remains software support.

For starters, Intel hardly has a track record for great video drivers. Moreover, raster chips from the likes of ATI and Nvidia have a decade of development and fine tuning behind them. That's a major challenge for Larrabee, whatever its theoretical advantages.

What's more, in a recent interview with TechRadar, Nvidia's chief scientist David Kirk was dismissive of the threat posed by the Larrabee chip. Ray tracing, Kirk says, will merely add to the box of tricks Nvidia has at its disposal.

Interview: Nvidia's Chief Scientist David Kirk

Jeremy Laird — 1206009586

Nvidia was left looking a little lonely after ATI, its main rival in the graphics market, was gobbled up by AMD. But Nvidia's Chief Scientist David Kirk says the future remains bright for the world's leading producer of PC graphics chips.

For a wider discussion of Kirk's views on the future of Nvidia and the threat posed by ray tracing, fusion processors and Intel's foray into graphics, see our main story: Do new CPUs threaten Nvidia's future?

In the meantime, chew on these key highlights from our discussion with Kirk on hot topics including ray-tracing and the CUDA programming language.

The impact of ray-tracing

TechRadar: Other than simple performance issues, why has ray tracing not been widely adopted for real-time rendering on the PC?

David Kirk: It would be easy enough to "just do everything using RT (ray tracing)," but then you would have to do everything using RT! Doing everything using RT in practice means tracing an enormous amount of rays, more rays for anti-aliasing, more rays for soft shadows, more rays for global illumination, more rays for glossy reflections. And so on.

There are certainly clever ways to avoid each of these, but every clever thing requires more software and more special cases. After all that, RT is not sounding so much like the "simple, elegant, handles-everything-easily" solution.

I mentioned anti-aliasing [removing or avoiding jagged edges] first, because I've from some commentary that people seem to think I'm ignorant of all of these techniques. There are ways to do anti-aliasing and not trace a lot more rays, but they all require more work (clever software) and they all have flaws.

One example is adaptive anti-aliasing. In this technique, you trace fewer rays, and look for edges by comparing adjacent rays to see if they are different. If they are different, you have found an edge and you trace more rays to make it smooth.

This has several problems. First, you may miss small things or small parts of things if they fall between the rays. Second, (I wrote a paper about this about 10 years ago!) this method is flawed because it introduces bias. Bias means that the picture could be arbitrarily wrong. This decision making influences the resulting picture in undesirable ways.

One other issue is that RT is famous for shiny, metallic looking reflections. What if you don't want that? Maybe you want a glossy, soft reflection, like brushed metal, or something more like fabric? You require a more complex shader, that either looks a lot like the shaders that people write in a rasterisation pipeline, or ... (here it comes again) ... you need to trace a lot more rays.

Parallel graphics processing

TechRadar: You've suggested the idea of a hybrid approach to the introduction of ray-tracing rather than the wholesale replacement of raster hardware. How do you see this happening? Can ray-tracing taking place simultaneously with other methods such as raster in future game engines?

David Kirk: Yes, RT and rasterisation can (gasp!) coexist. I don't understand why people find this remarkable. A game engine could rasterise the environment (using hierarchy, to make the complexity log, not linear, as it touted with RT), and find out what object is in each pixel. This is much faster than RT.

Then, for each pixel, the shading could either be done using conventional (and hardware-accelerated) pixel shaders, or by tracing some rays to find reflections, shadows, or ambient occlusion / light inter-reflection, or any combination of the two techniques.

This is totally doable on current GPUs, since you can rasterise and shade with OpenGL or DirectX, and trace rays with a program written in CUDA (Nvidia's parallelised version of the 'C' programming language), running on the GPU. Not only is this doable, I believe that this is the preferred way for using RT.

Why trace rays for cases when rasterisation simply is better and faster? In short, use RT for the features that it can do best.

New Nvidia GPUs?

TechRadar: with that in mind, is Nvidia doing any specific work to optimise future architectures for ray-tracing? Do you think chips optimised for "hybrid" rendering would look substantially different?

David Kirk: As I said, GPUs can do this now. It is certainly possible that we could provide special hardware that would make RT better or faster, but I think that today's hardware is pretty good.

The combination of current APIs and CUDA allows developers to write any program they want, anyway. Some programs are faster and more efficient than others, though, and I expect we will work to optimise the hardware to run these better. RT is certainly one such program, but there are many others.

I think that chips optimised for hybrid rendering will look substantially the same as GPUs do now. They would have hardware for accelerating special features in the APIs, such as texture, rasterisation, and programmable shaders, and they would have a general purpose interface for running parallel C programs, like CUDA. We'll continue to expand CUDA to make it better for a larger class of programming problems, but I don't see any need for substantial changes yet.

TechRadar: Regards CUDA and our discussion about the possibility that it might be adopted by other vendors of graphics hardware and your suggestion that NVIDIA positively welcomes this - what's in it for Nvidia to have CUDA supported for competing hardware? How would this actually work - would licenses need to be acquired / paid for?

David Kirk: I don't have any comment about licensing - interested parties should enquire! I'm simply saying that in much the same way as C can be compiled for many architectures, whether x86 or PowerPC, CUDA is just parallel C and can be compiled for other parallel or serial architectures.

Broader adoption has the advantage that CUDA code can run in more places, so the investment of writing your application in CUDA becomes more valuable if it runs on other architectures. Write once, run anywhere, perhaps many times.

CUDA already runs on multi-core CPUs in our emulation mode, for debugging, and this could become a higher-performance solution. Not nearly as high-performance as a GPU that is optimised for CUDA (of course!), but faster than more parallel CPU code runs today.

Intel's Larrabee: power efficient performance graphics?

Jeremy Laird — 1205856460

The steady trickle of Intel Larrabee data continues. This week, Intel revealed that the new multi-core media processing chip will be much more power efficient than existing graphics processors. Intel also said Larrabee benefits from “vector-processing” extensions designed to improve graphics processing.

Larrabee is a new many-cores co-processor composed of an array of x86 cores. The first Larrabee chips are expected to pack 16 cores, each broadly compatible with Intel's current PC processors. However, Larrabee's “in-order” cores are significantly smaller and less complex than the heavyweight “out-of-order” found in CPU's like the Core 2 Duo. It's a chip optimised for highly parallel workloads, most notably graphics.

Nvidia and AMD too power hungry

Speaking yesterday, Intel vice president and Director Stephen Smith said existing graphics chips from the likes of Nvidia and AMD were much too power hungry. Peak consumption figures of 150 watts or more are not unheard of for high performance GPUs. Smith said the superior efficiency of the Larrabee architecture would allow Intel to create entire performance PC platforms that consumed less than 140 watts.

That's an extremely welcome development given the relentlessly ballooning power profile of the modern high performance PC. But we still don't know how well Larrabee will actually perform or even if Intel expects it to be fully competitive in today's raster-based graphics market. Smith merely said he expected Larrabee to “do well” in the fight with AMD and Nvidia.

However, Smith did disclose that Intel has added a new vector instruction set to the Larrabee core. Vector instructions are typically used to improve graphics and media processing. Although Smith declined to be drawn on the details, it's like the new instructions are designed to boost Larrabee's rasterisation performance. Rasterisation is currently the dominant form of real-time 3D graphics rendering on the PC.

Of course, Larrabee's key advantage over existing GPUs remains its full compatibility with the x86 instruction set as seen in PC processors for several decades. It's a feature Intel reckons will make Larrabee much easier to program for general purpose use than highly specialised GPUs from Nvidia and AMD.

The first Larrabee-based add-in boards are expected to be available in late 2009 or early 2010.

Meanwhile… Intel also fired off a few new details on its Sandy Bridge processor architecture, due in 2010. It too will receive a new tranche of vector-processing instructions. Known as AVX, the new instructions were described by Intel big wig Pat Gelsinger as “SSE on steroids”. By expanding vector width support from 128-bit to 256-bit, AVX will no doubt deliver another hefty boost in multi-media performance.

Do new CPUs threaten Nvidia's future?

Jeremy Laird — 1205838233

It's worrying times for Nvidia. The graphics industry is apparently poised to transition to ray-tracing technology. Fusion-style processors threaten the existence of traditional CPUs and 3D chips. And Intel is getting serious about PC graphics for the first time.

And yet Nvidia's chief scientist David Kirk reckons it's all going to plan.

There used to be a pleasing symmetry among the four key players in PC technology. AMD took on Intel for processing prowess, while ATI and Nvidia duked it out for graphics grunt. Everyone knew their place. Then AMD snaffled up ATI at the end of 2006 for a few billion greenbacks and the balance in the universe was upset.

Since then, shockwaves from the AMD/ATI deal have spread far and wide. It completely torched what had hitherto been an extremely productive alliance between AMD and Nvidia. According to one Nvidia insider, the daily contact he had with AMD dried up literally overnight following the ATI purchase.

Fusion reaction

More recently, the strategic implications of the AMD/ATI tie up have become rather ominous. Intel and AMD both have plans for CPUs with integrated graphics, sometimes known as fusion processors. In the case of Intel, such a chip should be on sale before the end of the year in the form of Nehalem.

Granted, AMD has suffered all manner of woes following its acquisition of ATI, not least slow and buggy processors. But it's actually Nvidia that now looks most vulnerable.

In part, that's because Nvidia has no x86 kit currently on its books. What's more, even if it wanted to make PC-compatible CPUs, it lacks the necessary x86 license. And isn't the general trend supposed to be a gradual convergence of CPU and GPU technology towards a single chip containing a massively multi-core array of floating point fun? That would surely leave Nvidia out in the cold.

But even if Nvidia can carve out a long term strategy that doesn't include CPU production, its core graphics competency is under threat.

Momentum appears to be building for a new approach to graphics rendering known as ray tracing. It's proponents claim it produces more accurate and realistic graphics than the raster-based technology that currently dominates the graphics chip industry.

Ray tracing revolution

The key worry for Nvidia is that ray tracing might just present an industry-wide inflection point that allows Intel to enter the graphics market on an equal footing. That's certainly what Intel seems to be banking on with Larrabee, a multi-core chip that's thought to be highly optimised for ray tracing.

Suffice to say, therefore, that the easy domination Nvidia currently enjoys in the graphics market is hardly a given for the future. To find out exactly how the green-tinged graphics goliath plans to face up to these challenges, TechRadar crossed swords with none other than Dr David Kirk, Nvidia's chief scientist since 1997. If there's a man alive who has a better grasp of where Nvidia is heading, well, he wasn't available for interview!

Kirk is immediately dismissive about the danger posed by upcoming integrated CPU-GPU chips.

"Integrated graphics has traditionally been a low cost play," Kirk told us. Intel first began integrating graphics into its motherboard chipsets because it could be done almost for free. As Kirk says, the integrated option is essentially "the best graphics that no money can buy".

Not exactly a money spinning market segment, therefore. Yes, AMD's Fusion CPU is likely to raise the bar for integrated graphics performance, as Kirk concedes. But it nevertheless won't come close to what one might describe as acceptable gaming graphics performance. And if there's one thing Kirk is confident about, it's that consumers continue to demand high performance graphics.

What price performance?

For proof, he points to the contrasting fortunes of CPU and GPU pricing in recent years. The current CPU price war proves consumers are not sold on the latest high performance multi-core chips. And yet buyers continue to pay a stiff premium for high end 3D chips.

There's also no doubting the enormous difference between a high end graphics card and an entry level item, in terms of the end user experience. It's much, much larger than the typical gap between budget and premium CPUs. If you want decent graphics performance, discrete will be the way to go for years to come.

But what about the threat from Larrabee, Intel's first real effort to crack the discrete graphics market and due out late next year? Is the assumption that ray-tracing will be the next big thing in 3D graphics accurate? Not exactly, according to Kirk, who says, "there's nothing new about ray-tracing".

Historically, ray-tracing hasn't been used for real-time rendering because it is extraordinarily computationally expensive. That remains the case today. For many operations rasterisation does a very good job and does it 100 times faster than ray tracing. However, there are areas where ray tracing can be used efficiently to increase realism.

Hybrid rendering

The future according to Kirk is therefore much more likely to involve a hybrid approach to rendering. "Ray-tracing will not replace rasterisation. But it will add to our bag of tricks." In any case, Kirks says, there's no reason to assume that Nvidia's GPUs won't be extremely good at ray-tracing. Either way, the implication is that Intel's Larrabee will have an extremely tough fight on its hands.

As for the suggestion that CPUs and GPUs are converging towards a single, floating-point solution, Kirk simply isn't having it. "Even the latest multi-core CPUs only offer a small fraction of the floating point power of Nvidia's fastest GPUs," he says. If anything, this performance advantage will mean so-called general purpose applications for Nvidia's GPUs (known as GPGPU for short) are likely to win an increasing share of the market for really intensive computational solutions.

"We've gained lots of traction in the scientific community. Molecular modelling, astrophysics, climate modelling - all of these are highly parallel tasks that demand much more performance than is currently available."

Not that Kirk thinks that GPUs will replace CPUs. He accepts the need for truly general purpose processors will remain for the foreseeable future. But so will the demand for the increasingly flexible and powerful co-processor that is the modern GPU - preferably Nvidia's.