Nvidia gives first look at next-gen Fermi GPU

(Credit: Nvidia)

We wish we could provide you with information like clock speeds, shipping dates, and prices for 3D cards using Nvidia's new graphics architecture, code-named "Fermi." Instead, all we've been able to garner from the various reports around the Web from Nvidia's preview event is that Nvidia is pushing the parallel computing capabilities of its new chip harder than ever.

If you really want to get into the dirty architectural details, Anandtech, PC Perspective, and the Tech Report each have multipage stories that dig into the information Nvidia unveiled so far. From a gaming perspective, the most significant features Nvidia mentioned are that Fermi will indeed support DirectX 11, and that it will use GDDR5 memory. Those features answer two of AMD's most obvious advantages with its new Radeon HD 5800-series cards, but Nvidia hasn't provided information on availability, which remains AMD's most important edge.

Gaming was not the primary topic of the day with Fermi, however. Instead Nvidia focused most heavily on its CUDA GPU computing technology as it relates to its Tesla, enterprise-class product family. AnandTech reports that Nvidia cited one bragging point about a company using its previous generation GT200 chips to migrate "a cluster of 2000 servers to 32 Tesla S1070s, bringing total costs down from $8M to $400K, and total power from 1200kW down to 45kW." Nvidia hasn't mentioned clock speed figures for Fermi, so we can't predict its performance just yet, but as PC Perspective reports, Fermi "is made up of 3.0 billion transistors and features 512 CUDA processing cores organized into 16 streaming multiprocessors of 32 cores each." That's more than twice the core count in the 240-core GT200, so expectations are reasonably high.

In addition to more cores, Nvidia has also added support for the C++ programming language to Fermi. That should increase its appeal to programmers, many of whom have found GPU-targeted software development difficult. And of course in addition to CUDA, Fermi will also support Microsoft's DirectCompute and the open standard OpenCL GPU computing standards.

Other features for Fermi abound, and we encourage those interested to delve in with the enthusiasts sites that attended Nvidia's preview event. We expect information on the consumer-level products that emerge from Fermi won't be too far off either, so stay tuned for more in the coming weeks and months.

[tipoo_]

One of the Nvidia white papers (yes, I'm a geek who actually reads them) says 1.5 Terraflops, which is dissapointing considering the Radeon 5870's 2.72, but of course thats a far from a perfect indicator of performance.

The other architectural changes look incredible! Maybe Nvidia was really on to something when they said the importance of CPU's is deminishing.

[rhbrown]

I don't see AMD or Intel (especially) having trouble moving their faster multicore CPUs any time soon, but it's become obvious that the lines between CPU and GPU will only get blurrier. Anyone want to give odds on the distinction going away by the end of the next decade?

[Saltiva]

Looks like AMD is making good on their winnings from suing Intel. And I agree with RHBROWN that the distinction between GPU and CPU is getting blurred. Maybe eventually the CPU with all its wonderful cores (soon to be a norm of 6) will incorporate the functionality of the GPU and the demise of the GPU altogether? Seems almost pointless to asign a seperate chip for video tasks?

[amirault0]

by the end of the decade the CPU and GPU will be together on one chip. Kinda like what happened to the north and south bridge on the P55 chipsets.

[shellcodes_coder]

So when are they releasing a DirectX 11 compatible graphics card?

[rhbrown]

I've read speculation ("December," "Q1 2010"), but haven't heard or seen anything concrete.

[EdmondDantes]

Fermi? Does it require an attached nuclear power plant to run it?

[cnet-og]

"Fermi" is probably in reference to the fact that electrons are Fermions... remember the Fermi level from physics class???

[C433Z]

Will it have something in competition with the 'Eyefinity' ability of the ATI cards?

[rhbrown]

I'd expect multimonitor support from Nvidia's new cards, but it's anyone's guess as to the maximum number.

[douggdangger]

I used to judge video cards on numbers and specs. Not so.

I've seen ATI cards with hundreds of "stream processors" and NVidia having a small fraction of that and the NVidia still beats its brains out.


I was an AMD fan during the Pentium4 vs Athlon days, but Intel's Core and i7 are just running circles around AMD's best. The same goes for ATI vs Nvidia.

[tipoo_]

The number of stream processors (or CUDA cores) has nothing to do with it - AMD and Nvidia just have different approaches. AMD's is a higher number of simpler stream processors, Nvidia's is a lower number of more complex ones. One to one, an Nvidia stream processor (or whatever they want you to call it) can do more than an AMD one in the same ammount of time, but AMD simply puts more of them on thier chips to compete. Its a good strategy, it allows them to be highy modular and create an entire range of cards from high-end to low-end very quickly.

[tipoo_]

p.s, Nvidia's current cards are not "running circles" around AMD's cards in any way shape or form. Looking at GPU sales (as well as benchmarks) AMD is clearly doing better than Nvidia. Their price to performance ratio is superior on the 4k series cards, and with the launch of the 5870 they have the performance crown as well.

[ArsFragica]

Nvidia has no GPUs that can match ATI's new 5870 GPU.

[krosafcheg]

As a researcher investigating highly non-linear, transient phenomena in Computational Solid Mechanics, I need all the computational power I can get and more. There's never enough of it. Colleagues working in areas such as Fluid-Structure Interactions and Computational Aerothermodynamics are just as excited.

Why? The kind of compute power that FERMI would bring to the workstation would save me (and many others) a lot of the usual bother involved in HPC. Submit a job today, wait one week for it to move to the head of the queue, write proposals for CPU time that involve competing with 100 other people etc. But a box with 2 of these babies would give me 2 TFlops of DP power - on a workstation. The reduction in turnaround time provided by such a system would be incredible.

[_sub_zero]

<snip>
And of course in addition to CUDA, Fermi will also support Microsoft's DirectCompute and the open standard OpenGL GPU computing standards.
</snip>

I believe you meant OpenCL instead of OpenGL?

[rhbrown]

Ha, of course you're correct. Updated.

[Ryan_R]

I'm glad the AMD 5800 card came out first, since it probably means that nVidia have a bit more time to make sure their next series of cards are better. I'm still running a 9800 GX2 which still runs great (hot, but great). If I don't get a new nVidia card from the upcoming series, I'll likely get one from the series after that.

[intru3d]

haha, ati hd 5870 in some benchmarks lost to GTX295?
though ati won in the price point, hope nvidia will return with a nice price tag and
more powerful gpus. and also multi monitor capabilities.

[sqw3rty]

Well, comparing a Card with 2 GPUs in it with a newer card with only one isn't exactly a fair comparison (the margin isn't evem very big)
The 5870x2 will beat up the GTX295 for it's lunch money, lets just wait and see if nvidia's next internal SLI Card card can stand up to it.

[kamyar07]

Make these things quieter and cooler

[tipoo_]

With more than double the transistor count of the current (already hot running) Nvidia GPU's, that isnt looking likely.