VizWorld.com

In response to the big news about the breakdown of NVidia’s Fermi presentation at GTC, Nvidia has issued a format response confirming what I suspected:  Fermi is real, and was running at GTC, but it’s still very much an “engineering prototype”:

Nvidia confirms Fermi that was running PhysX at Jensen’s GTC keynote was real but the one that we all took pictures of was a mock-up.

The real engineering sample card is full of dangling wires. Basically, it looks like an octopus of modules and wires and the top brass at Nvidia didn’t want to give that thing to Jensen. This was confirmed today by a top Nvidia VP president class chap, but the same person did confirm that the card is real and that it does exist.

via Fudzilla.

Hardware fermi, gt300, gtc, nvidia

Be sure to read the followup to this, including comments from NVidia.

NVidia showed the GT300 (Fermi) board up on stage during the keynote of the GTC this week, but was it a real Fermi board? Or merely a prop?  They were pretty positive about release and production schedules, but SemiAccurate takes a close-up look at the photos and PR material Nvidia gave out and paints a less rosy picture of the situation.

Some things to note on it. First is that the second set of digits on the first line says 0935A1. A1 is for first silicon, something that, when coupled with a direct quote from Jen-Hsun of, “You are currently looking at a few day old silicon” (from here select the video “See video of Jen-Hsun Huang announcing Fermi”) kind of blows the whole ‘silicon in Santa Clara last spring’ story out of the water. I wonder if anyone will retract that, or just change the article retroactively?

To make matters worse, the other digits are a date code, 0935 means 2009, work week 35. If Nvidia starts its work week on the first full week of the year like everyone else, that would put WW35 at August 28 to September 5, 2009. Where have I heard that date before? Oh yeah, here. (edit: Referring to the 2% Yield number rumors).

While it’s no surprise that the boards were engineering prototypes (in reference to the author’s shock of finding common wood-screws holding it together), it is a bit disturbing if his analysis of the chip production numbers is correct.  He also gets into amazing depth on some of the odd solder design and PCB layout issues, like this:

The 6-pin connector, on the other hand, lines up with, umm, nothing. There is a potential 4-pin floppy/sound/jumper block below it, but you can clearly see there is nothing in the vias, not even solder. The 6-pin connector connects to nothing, and nothing is holding it in. Except glue. Notice the connector is black and the hole below it shows white. The only real question now is, Elmers or glue stick?

To make matters worse, the mounting holes for the 8-pin connector, which should be between the 6-pin and 8-pin fakes if the card was real, are empty. Piss-poor fake job guys. Go read your fanboi forums, they do a better job, and work for much cheaper than your ‘geniuses’.

Unlike the original author, I do believe that Fermi was alive and well at the conference, but he does raise some interesting questions.  Perhaps that was a “defective board” held up to show, or a very early engineering prototype (as all the functioning boards were in-use).  Perhaps that was Huang’s personal “lucky presentation board” that he likes to use, but has long-since been obsoleted.  Who knows.

Thoughts?

via SemiAccurate :: Nvidia fakes Fermi boards at GPU Technology Conference.

Hardware conference, fermi, gt300, gtc, nvidia

Nvidia has been talking about Fermi all week and keeps talking about the power of GPGPU on it, but they haven’t really said anything about the graphics capabilities.  Why is that?  Here’s a response on the hpc-focus from NVidia CEO Bill Dally:

He also explained why the chip was billed as a supercomputer chip initially and not a gaming chip. “It’s a zero-sum game. You have a certain amount of die (chip) area, a certain power budget. It is the case that we put a bunch of die area into double-precision floating point, a bunch of die area into ECC. And for gaming graphics applications, those give less returns than they do for the scientific applications,” he said. Double-precision floating point operations are used heavily in scientific computing. ECC, or error correcting code, is a technology that can correct data errors on the fly.

Of course, it’s not really for HPC-only, it will run graphics as well.  However, give a user the choice of a minor increase in graphics quality (bump mapping or better shadows) and better physics (via GPU-accelerated systems like PhysX) and I bet they’ll take the physics, and that’s exactly what Fermi brings.

via Nvidia ‘Fermi’ chip for Mac, Windows too | Nanotech – The Circuits Blog – CNET News.

Hardware gpgpu, gt300, nvidia

Just how powerful is “Fermi”, Nvidia’s newest GPU?  A new video available on Youtube shows a real-time fluid simulation of 128,000 particles rendered and simulated on the GPU with full surface tension effects.

See the video after the break… wow.

Read more…

Science cfd, fermi, gt300, hpc, nvidia, simulation

Photo courtesy of John Leidel of InsideHPC.

If you are one of those people that hates all the extra power connectors that several modern video cards require just to sustain basic function, then your not gonna like the new GT300 (Fermi) From Nvidia. Seems that 6GB of GDDR5 memory pushes things a bit too close to the line, resulting in two additional power connectors.

In case of upcoming high-memory configurations nVidia Tesla, Quadro and GeForce cards, the company had to install a 6-pin and an 8-pin connector, getting 300W of power to play with. However, this was a precautionary measure. According to information we have at hand, the GT300 board [yeah, featuring "Fermi" CUDA architecture] barely missed 225W cut-off for the 6+6 pin if the board comes with 6GB of GDDR5 memory.

nVidia could have gotten 6+6-pin configuration and still ship 6GB version, but the margin of efficiency would be just too low [even with digital PWM, boards cannot be 100% power efficient] to qualify inside OEM systems. The decision was thus made and the 6GB cards come with 300W of available power.

Better start shopping around for power supplies.

via nVidia GT300 real power consumption revealed: Not 300W! – Bright Side Of News*.

Hardware fermi, gt300, nvidia

The age of GPU-computing got another huge boost today as a press release from NVidia shows that their new GT300 “Fermi” chip is more than vaporware and will, in fact, be the foundation of Oak Ridge’s newest supercomputer.

Jeff Nichols, ORNL associate lab director for Computing and Computational Sciences, joined NVIDIA co-founder and CEO Jen-Hsun Huang on stage during his keynote at NVIDIA’s GPU Technology Conference.  (…)

“This would be the first co-processing architecture that Oak Ridge has deployed for open science, and we are extremely excited about the opportunities it creates to solve huge scientific challenges,” Nichols said. “With the help of NVIDIA technology, Oak Ridge proposes to create a computing platform that will deliver exascale computing within ten years.”

I don’t think this has anything to do with ORNL & UT’s latest announcement of their “Nautilus” NSF-funded machine.  There’s currently no real specs for what this new hybrid supercomputer will be, but they promise exascale compute-capabilities.  Read the full announcement after the break.

Read more…

Hardware, Science fermi, gpgpu, gt300, hpc, nvidia, ornl

It seems the GT300 news was actually under embargo until the GTC keynote was finished (guess BSN decided to ignore it, oh well), and now that it’s over the information is pouring out.  One great resource is an article from John West over at InsideHPC touting the new HPC-centric features of the design.

The new design features a dedicated 64KB L1 cache per Streaming Multiprocessor (GPU cores are organized hierarchically into “Streaming Multiprocessors,” or SMs; 32 cores form an SM, and there are 16 SMs on a board), and a 768KB L2 cache shared among all SMs. NVIDIA calls this the “Parallel DataCache Hierarchy,” and Sumit Gupta, senior manager in the Tesla GPU Computing group, says that this feature is very important not only to sparse matrix and physics calculations (for gaming), but also for traditional graphics applications like ray tracing. Application engineers should now see a much more familiar programming environment when porting code from CPUs.

via NVIDIA’s next generation GPU architecture has a lot for HPC to love | insideHPC.com.

Hardware gpgpu, gt300, gtc, hpc, nvidia

BSN has some information about the upcoming GT300 chip from NVidia, and the hard silicon numbers alone are staggering:

• 3.0 billion transistors
• 40nm TSMC
• 384-bit memory interface
• 512 shader cores [renamed into CUDA Cores]
• 32 CUDA cores per Shader Cluster
• 1MB L1 cache memory [divided into 16KB Cache - Shared Memory]
• 768KB L2 unified cache memory
• Up to 6GB GDDR5 memory
• Half Speed IEEE 754 Double Precision

But the awe doesn’t end there, check out this list of supported languages:

Ferni architecture natively supports C [CUDA], C++, DirectCompute, DirectX 11, Fortran, OpenCL, OpenGL 3.1 and OpenGL 3.2. Now, you’ve read that correctly – Ferni comes with a support for native execution of C++. For the first time in history, a GPU can run C++ code with no major issues or performance penalties and when you add Fortran or C to that, it is easy to see that GPGPU-wise, nVidia did a huge job.

If NVidia was aiming to revolutionize the world of GPGPU programming, then native support for C++ would do it for sure.  Especially if the speed boosts are anything close to CUDA/OpenCL.  Most C++ code won’t parallelize easily, but if it simply pushes raw instructions through faster then that would be a huge improvement.

Update:Read more about this architecture in the following stories:

• ORNL Looking to use GT300/Fermi for a new hybrid supercomputer
• Post GTC Keynote news about the GT300

via nVidia GT300′s Fermi architecture unveiled: 512 cores, up to 6GB GDDR5 – Bright Side Of News*.

Hardware gpgpu, gt300, nvidia

In response to the recent rumors of the abysmal 2% yield of the GT300 silicon, NVidia has come out swinging with reports that the GT300 will begin production next month (October) with a launch near the end of November.

A recent post at Fudzilla suggests that GT300 has been taped out, and the mass production may start in the middle of October. It’s very likely that NVIDIA will launch the new GPU around November 27th, and start shipping in December.

The company will probably show the card in the next few weeks to make the fanboys more comfortable when waiting for their DirectX 11 offering, as the positive reviews of Radeon HD 5870 have spreaded quickly around the internet.

via NVIDIA: Our 40nm Yields Are Fine – Expreview.com.

Hardware gt300, nvidia

If you’ve been saving your pennies and holding your breath for the new NVidia super-card, the GT300, then we’ve got some bad news for you.  NVidia has gotten back the first silicon dies of the new GT300 chipset and the news is amazingly bad.

How many worked out of the (4 x 104) 416 candidates? Try 7. Yes, Northwood was hopelessly optimistic – Nvidia got only 7 chips back. Let me repeat that, out of 416 tries, it got 7 ‘good’ chips back from the fab. Oh how it must yearn for the low estimate of 20%, talk about botched execution. To save you from having to find a calculator, that is (7 / 416 = .01682), rounded up, 1.7% yield.

So 416 chips were made.  7 were “good”.  And good doesn’t mean “functional”, just “as designed” (a design flaw would carry through).  If they can’t iron the kinks out of the process, we may never see the GT300 as advertised.

via SemiAccurate :: Nvidia GT300 yields are under 2%.

Hardware gt300, nvidia