Icrontic: Dissecting Fermi; NVIDIA'S New GPGPU Architecture

Icrontic takes a look at NVIDIA's newly-announced Fermi core. They're calling it a GPGPU (General Purpose Graphic Processing Unit), and it's gonna do more than just play games...

[primesuspect]

Now the question is: how will it stack up against current king-of-the-hill Radeon 5870?

[Kakkoii]

Well know that in a month or 2 when test cards start going out to reviewers.

Also, AnandTech has done a much more in depth review of the architecture:
http://www.n4g.com/industry...

[MNicholas]

Cell (full version): 230 GFlops/234 Million Transistors
Performance Density: 1
Fermi: 1500 Gflops/3000 Million Transistors
Performance Density: 0.5

Transistor for transistor, Cell (full IBM version) has literally twice the 32-bit floating-point performance.

A board with 8 cells has 1.6 Billion Gigaflops from the 64 SPEs alone yet has only a little over half the Fermi's transistor count and cheaper to produce. The nice bonus is 8 PPEs to do whatever else you could want. The rendering possibilities are endless.

The biggest benefit of Fermi, of couse, is that it's compatible with traditional PC based APIs like DirectX 11.

From a purely performance standpoint, however, there are better GPGPU-type solutions out there.

[Kakkoii]

So, your comparing a single purpose RISC processor, to a GPU that can do some CPU calculations.. Ya know the 360 and PS3 still need a GPU. So you should be factoring in that also lol.

http://en.wikipedia.org/wik...

"comparing only floating point abilities of a system is a one-dimensional and application-specific metric. Unlike a Cell processor, such desktop CPUs are more suited to the general purpose software usually run on personal computers. In addition to executing multiple instructions per clock, processors from Intel and AMD feature branch predictors. The Cell is designed to compensate for this with compiler assistance, in which prepare-to-branch instructions are created. For double-precision floating point operations, as sometimes used in personal computers and often used in scientific computing, Cell performance drops by an order of magnitude, but still reaches 14 GFLOPS (the PowerXCell 8i variant, which was specifically designed for double-precision, reaches 102.4 GFLOPS in double-precision calculations [35])."

[MNicholas]

The idea was to look at GPGPU solutions since that's what the article header alluded to.

I have no idea why you brought Intel and AMD desktop CPUs into this. Those represent an architecture that is outdated and increasingly irrelevant. Similarly, your quote from Wikipedia is meaningless in the context of this discussion.

There is something in that quote that's relevant to this discussion: A simple "gigaflop" comparison can be misleading. Tests using optimized code have showed the Cell to provide results that are up to an order of magnitude faster than numbers would suggest.

Check out the Fermi ray-tracing thread for more information.

[siliticx]

A flabidy bajillion dollahs! for maybe 10 mroe performance but 100% the price of the 5870. Give up nvidia.

[Kakkoii]

You obviously have no idea what your talking about, and lack the ability to understand any of the data put forth in these articles.

The raw processing power of this card is a larger fold increase than ATI's 5870 was. And they also stated at the event they were wrong in their pricing of the GT200.

[primesuspect]

"Yo, Icrontic, I know you like Fermi, and Imma letchu finish, but Anandtech had the best Fermi breakdown of ALL TIME. All time"

[Kakkoii]

Sorry to break the news, but this ain't no awards show. I don't know if you work for Icrontic, but I assumed not. I was just sharing the link because I thought you'd be interested to read an even more in depth review of the architecture.

And you went and added yours as an alt on mine too, don't see me b*tchin. So don't be a hypocrite.

[hoops]

Lets see the pricing for this beast. If it hits in the $500+ range when it's released then its dead in the water before it hits because you know AMD's pricing strategies. They will discount the 5870 when this arrives and release the 5890 to compete with the 5870x2 in tow. Have to love competition. Keep them prices low!
As for time-line for Nvidia's GPU?

Jensen stated, "Nobody likes when the competition has a product. I don't like keeping our enthusiast waiting on our next generation processors. I would rather be shipping today, but we will ship when the product is ready to ship."

"Jensen just had this to say live at the NVIDIA GPU Conference in regards to the next-gen GPU: "You will have to wait just a little longer."

"If you are waiting for NVIDIA to jump out of the GPU closet with a 5800 killer and put the fear into you for making a 5800 series purchase for Halloween, we suggest paper dragons are not that scary. We feel as though it will be mid-to-late Q1’10 before we see anything pop out of NVIDIA’s sleeve besides its arm. We are seeing rumors of a Q4’09 soft launch of next-gen parts, but no hardware till next year and NVIDIA has given us no reason to believe otherwise."
-HARDOCP.COM-

[lagoonalight]

This is the series I see launching in the PS4. At that time you will see an improved Cell and a better bluray drive and one of these cards coming out for something like 399. There will not be another 600 dollar console because Sony has all that tech already invested. The PS4 is going to be a Monster.