NVidia GT300, Larrabee, and ATi

[BDUAres]

Moving the technical discussion about the GT300 to the tech forum.

i do not find his analysis acceptable that the GT300 is equivalent to Larrabee because he is not a computer engineer, presents no references for his arguments...if you have a source other than Charlie that confirms this, i'll accept it as an argument to consider.

From many IT web sites:

The GT300 Cores rely on MIMD-similar functions [Multiple-Instruction Multiple Data] - all the units work in MPMD mode, executing simple and complex shader and computing operations on-the-go

There are literally dozens of IT sites which repeat this same information, which you can look up yourself, and no one is arguing this point. The key point is that the GT300 is a MIMD architecture card.

Here is retired IBM Distinguished Engineer, U. of Colorado Research Faculty member and parallel cluster expert (the system used by Larrabee and GT300) who literally wrote the book on clusters talking about the GT200 chip compared to Larrabee, Greg Pfister:

Computing expert Greg Pfister, who’s worked in parallel computing for
30 years, has a good blog about the differences between MIMD and SIMD
architectures...which is well worth a read if you want to find out
more information. Pfister makes the case that a major difference
between Intel’s Larrabee and an Nvidia GPU (edit: pre-GT300) running CUDA is that the
former will use a MIMD architecture, while the latter uses a SIMD
architecture.

This is what Charlie meant when he spoke about the GT300 being like Larrabee: they use the same type of system, which is MIMD architecture. (Larrabee is not dead BTW, they only canceled one part of the Larrabee Project, the Larrabee Prime. The rest of the project is humming along nicely)

Pfister goes on to explain part of the software side controls on a MIMD architecture:

Pfister says
that “Larrabee should have a big advantage in flexibility, and also
familiarity. You can write code for it just like SMP code, in C++ or
whatever your favorite language is.”

http://www.bit-tech.net/news/hardwar...architecture/1

NVidia confirms this with their press releases which state that the GT300 will run on C++ code. Specs given so far give not just memory and gpu core clock speeds but now shader clock speeds as well. This is because the shaders are simple generic multi purpose cores that take the C++ and other software commands to present the shader information.

Previous articles about MIMD architecture compare it, and the CGPU chip running it, to presenting graphics on a computer without a graphics card at all, exactly as in software DirectX mode. This is basically what the GT300 does, now that cpu and memory hardware technology is at the point where the software mode is able to handle the amount of information fast enough to present the graphics in a way comparable, or even better than, the hardware based graphics cards.

This is the big 'revolution' in parallel MIMD cards like the GT300 and now defunct Larrabee card- they can compute both graphics processes and regular cpu processes in one piece of hardware. Not using heavy duty graphics? It can even help your computer's regular cpu do it's job. Some experts are even saying the line is now blurred between the need for a GPU and CPU seperation at all, that speeds are fast enough that we only need the one piece of hardware. As a matter of fact people have pointed out that these types of graphics cards can run an OS all by themselves without any need for a mainboard cpu.

Grasp the concept now?

[rEDkILLER]

[BDUAres]

NVidia released the architecture specs of the Fermi, originally called the GT300 card, but now it seems it will be called the GF100.

Tarinder Sandhu at Hexus.net went into great detail about this card, published 18 Jan 2010.

On top of presently supported C, via CUDA, there will also be support for standards-based OpenCL, DirectCompute, and DX11, as well as C++, the latter made possible by the use of a new, low-level instruction set, known as Parallel Thread eXecution 2.0

Jen-Hsun Huang, NVIDIA's CEO, has spoken about Fermi in a parlance that's more familiar to the dissemination of CPU architecture. With constant references to cores, caches, and hierarchy, one could be forgiven (for thinking) he was describing an Intel Larrabee-esque design.

Radeon HD 5870 is out...etailing for £300. The higher-end Fermi cards will cost at least as much, we imagine, because fitting 3bn transistors - and yes, NVIDIA and AMD count them differently - must mean a die-size that's appreciably larger than HD 5870's 334mm². Yields are inextricably linked to die-sizes, so producing a 50 per cent larger die (than HD 5870) will, ceteris paribus, lead to a greater number of per-wafer flaws. Going big(ger) is bad for business.

AMD's hit the ground running with its high-end GPUs and will see healthy sales before Fermi ever gets packaged inside a retail box. Time to market is telling, and NVIDIA's procrastination may well cause financial pain - in the short term at least. We expect the card to hit the shelves no earlier than February 2010 (now estimated at March 2, 2010), giving AMD a clear four-month-plus run with DX11 parts in the channel.

Ultimately, NVIDIA has detailed Fermi as a forward-looking, programmable architecture that's aimed at enlarging the company's footprint in the lucrative HPC space - somewhat pre-empting Intel's (defunct) Larrabee - whilst keeping desktop and mobile 'gaming' customers happy.

Once again we have tech reviewers comparing the Fermi to the Larrabee, though this time after NVidia's architecture tech info release. Score one for Charlie.

They confirmed again the much larger size die, and the resulting much higher failure rate of the tape outs of die wafers that we have already heard about. That's two for Charlie.

Parallel thread execution with C++ software. That's three.

Real World Technologies, publishing the 11 pages of the NVidia tech release online, had this to say about it:

Curiously, they are also not discussing the graphical capabilities of this chip and instead focusing only on compute. Hence our discussion is focused primarily on the GPU as a compute device...Perhaps the most significant demonstration of Nvidia's commitment to compute is the fact that a great deal of the new features are not particularly beneficial for graphics. Double precision is not terribly important, and while cleaning up the programming model is attractive, it's hardly required. The real question is whether Nvidia has strayed too far from the path of graphics, which again depends on observing and benchmarking real products throughout AMD's, Nvidia's and Intel's line up; but it seems like the risk is there, particularly with AMD's graphics focus.

Meanwhile the card as shown at CES runs at 280w with a single gpu. That is damn close to the 300w max power supported by the actual PCI-e slot the card plugs into. This means that you can kiss a dual gpu card goodbye unless it is severly powered down. Additionally though the cards will have 512 cores, and over 3 billion transistors (hence the very high power demand) some of NVidia's released documents state that they will only be using 448 cores, and 2 out of their 16 shader blocks will be disabled.

ATi meanwhile has been busy and has had their HD 5870 overclocked to 1+Ghz up from 825mhz. Things look promising for a very good range of HD5xxx cards for the summer with even more speed.

[Jasio]

So what I essentially said in the other thread was spot on... yay for accurate speculation. Looks like I'll be swapping out my 280GTX's for 5870's or 5890's in early-mid 2010.

[BDUAres]

Nvidia has once again changed the name of the Fermi cards. The first two cards based on Nvidia's GF100 chip will now be called the GeForce GTX 480 and the GTX 470, skipping the GT300 series names completely.

It seems that it thinks there is no point to having any high-end 300 series cards, and will push its Fermi products into the 400 series instead. We can't say we blame it. After word got out that Nvidia's 300 series would be associated with old cards that could not run Direct X11, the brand for that range of model names was pretty much poisoned. Besides if its next graphics cards have bigger numbers they must be better and more value for money, right?

This is just the latest in a long string of micro-information releases on Fermi. Nvidia recently showed us the Fermi architecture, sort of, in block diagrams.

But what we don't know yet are more important things like power consumption and clock speeds. Oh well, never mind.

When are they coming out? Pick a number, multiply it by your hat size and then divide it by your shoe size, and you might be close. Nvidia's graphics card manufacturing partners have all stopped talking about dates, but word on the street is that it will be after CeBIT, sometime, maybe.

As it is now February and we still don't have a Fermi card released, I will have to give another hat tip to Charlie for being spot on with his info yet again.

[BDUAres]

NVidia has announced the official launch date of its Fermi-based Geforce GTX 480 and GTX 470 graphics cards. Both are set to be out for sale on March 26th.
NVidia's announcement? Via tweet:

“Apologies for the confusion around our most recent GF100 update. To clarify, launch date for GeForce GTX 480 and GTX 470 is March 26, 2010”

This date coincides with the NVidia PAX 2010 LAN event

I also have a pre release price from the SabrePC hardware company. It is taking pre-orders now for the GTX 480 at a whopping $699 with a $20 discount bringing the price down to $679 USD. Ouch. This seems to be related to the shortage of chips they are able to make to put in these things.

[BDUAres]

Update on the Fermi.

After many long delays the Fermi cards have been launched (an 'in name' paper launch only, not actually shipped or available yet in any store for another 2 1/2 weeks minimum) and the first reviews have been done by a dozen PC review companies or so. While so far it seems that as far as single GPU based cards go the GTX 480 is for now the fastest card on the planet (in some games) until the 2 GB Radeon 5870 comes out in 2 weeks, there are concerns being raised by most of the reviewers regarding things like power consumption, heat, and even noise.

As I wrote here months ago the 2 Fermi cards released have had to have their 512 cores cut down to 448 and 480 for the GT 470 and GT 480 respectively, and the GT480 has lost 1 of the 16 SMs while the GT 470 has lost 2 of them. This is directly related to their still being big issues with the die process and there is still a very high die failure rate at the foundry, with the manufacturer even saying that this particular size of core die is a no-go and they would be working on developing other sizes of dies in the future.

The prices for the cards are now reported to be $349 for the 470 and $499 for the 480, or about $100 more than the top of the line ATi cards.

Frame rate winner for various games vary between NVidia and ATi cards, but most games have a higher framerate on the GTX480, if only by a small fraction:

Battlefield: Bad Company 2 - winner by a large margin = ATi 5870
Crysis: Warhead - winner = GTX480
BattleForge: DX10 and DX11 - winner = GTX480
HAWX - winner = GTX480
Left 4 Dead - winner = ATi 5870
STALKER: Call of Pripyat - winner = GTX480
DIRT 2 - winner = GTX480


Lets look at power consumption.

The GTX 480 is now officially the highest power sucking single GPU card in the world. NVidia is claiming it sucks 250W at load while the GT 470 sucks 215W. At idle they claim 47W for the 480 and 33W for the 470. The hardware reviewers found some discrepancies with this claim but let's take a look.

With a 1200W power supply the figures for power consumption (via anandtech.com):
Idle


Under load



As you can see you would need a minimum of a 600W+ power supply to even think of running one of these cards, and a 1000W+ minimum to run 2 of them, or, looking at it another way, depending on your electric KWh cost, an extra $50-$100 per month in electricity. The final benchmark shows that the single GPU GTX480 also draws nearly as much power under load in Crysis as the dual GPU 4870 x2 (38W less) or even 2 5870s (39W less) combined.


To sum up, overall the GTX480 is 5-10% faster than the Radeon 5870. The GTX470 is 5-10% slower than the 5870. At this moment, though the Fermi's won't be out for sale until about the same time that the 2GB 5870 is released, the GTX480 is the fastest single GPU card. However you can't count out ATi by a long shot. Unfortunately for NVIDIA the Radeon HD 5870 is cheaper to produce, consumes less power, is quieter, and it costs about 25% less, while the Radeon 5970 is still the fastest single graphics card in the world. ATi has factory overclocked their 5870 cards to over 900mhz and those haven't been benchmarked against the Fermis yet, nor have the 2GB cards, which are important as in these game tests there were memory limitations on the ATi's which indicated they would perform even faster with more RAM.

So the really interesting contests will begin in about a month or 2.

[Bad_Conduct]

I'm an Nvidia fanboy, but I dunno. The 5970 seems too come out on top.

[Weatherlight]

Fanboyism is a foreign concept to me. I see two Titans struggling against each other; each blow represents a step forward that I, the consumer, can take advantage of.

[Bad_Conduct]

Nvidia just always had easier drivers, and more games branded themselves with Nvidia compatible hardware. That's always been trump for me.

At the end of the day, I think it comes down to the companies who make the games that decide which cards will be good or not. It's more to do with marketing than the actual numbers.

But 3D mark - 5970 is clearly ahead.

[BDUAres]

Oh the 5970 easily beats the Fermi, that isn't being debated by either company. But the 5970 is a dual GPU card while the Fermi is a single GPU card. The 5970 basically consists of 2 significantly underclocked 5870 GPUs on one card, and the Fermi is really competing against single 5870s.

Yesterday a couple of companies released new versions of the 5970, which run the 5870 GPUs at normal clock rates rather than underclocking them. Those cards are the fastest single card devices in the world, and unfortunately no Fermi will ever be able to compete with them on a single card to single card basis, as the Fermi draws far too much power to ever allow it to operate in a "two GPU on one card in a single slot" type scenario. The PCI-e bus is already at it's power limit with a single Fermi GPU. Now if you have TWO Fermi cards and a monster 1200W power suply and an extra $50-100 per month of disposable cash on hand to pay for the increase in your electric bill, then you could possibly run 2 Fermi cards in tandem in separate slots, though no software graphics have been created yet that would take any advantage of such a broiling hot power drinking system.

[BDUAres]

Had to laugh when I read this headline today:

DESIGNER OF WARM GPUS Nvidia has released its solution for birth control by announcing the Geforce GTX480M for laptops.

This was followed by

The chip is intended for desktop replacement models used for gaming, meaning the machines will be cumbersome and heavy, presumably from all the cooling required to keep the GTX480M from melting. All this might in turn deter users from placing a GTX480M equipped laptop on their laps, or could make the heat of the battle an altogether more immersive experience.

Given that Nvidia didn't publically release any information on the thermal design power (TDP) of the chip, something it did even with its "mobile workstation oriented" Quadro FX chips, this leads one to wonder how high the figure must be

Those who want to own a mobile sauna will have the chance to do so as the graphics chip designer has allowed multi-chip SLI support. Quite how long it will take laptop designers to shove two or three of these GPUs along with the associated industrial strength cooling system into a 'portable' package is anyone's guess.

For Nvidia, putting Fermi into a laptop one way or another was vital to showing that it has been able to rein in, at least by some small margin, the power consumption of its folly in chip design. The remaining question is how many notebook vendors will be willing to put up with all of the associated baggage that Nvidia's latest and largest graphics chip brings.

Those considering purchasing a notebook equipped with a GTX480M GPU chip perhaps might do well to consider some form of protection from the heat.

I can already see the special sales on Hurt Locker brand combo packs of these laptops, complete with oven mitts and insulated lap protectors.