backtop

Print E-mail del.icio.us 43 comment(s) - last by peternelson.. on Jun 25 at 2:13 AM

NVIDIA Tesla D870 GPU card

NVIDIA Tesla C870 quad GPU HPC system

NVIDIA Tesla S870 dual GPU 1U HPC system
NVIDIA takes on AMD's Stream Computing Initiative with its Tesla GPU cards and HPC systems

Late last year AMD and ATI merged into one company, forming not just a bigger entity but creating a very specific roadmap altogether. AMD's grand scheme with ATI is to develop a single chip handling both general purpose computing as well as graphics. The Fusion project as AMD calls it is this very goal.

Today, NVIDIA reveals that it is not behind when it comes to general purpose GPU, or GPGPU, computing. Earlier this year the company announced its complete unified device architecture, or CUDA, Technology, which laid the groundwork for GPGPU programming for NVIDIA GPUs. CUDA Technology directly competes with AMD’s Stream Computing initiative.

The GPGPU product lineup will be known as Tesla. Tesla is a top to bottom product lineup consisting of internal PCIe cards and external high-performance computing, or HPC, systems – Tesla C870, S870 and D870.

The internal PCIe solution consists of an output-less GeForce 8-series based card on a PCIe x16 card. The Tesla D870 is NVIDIA’s only internal GPGPU card for desktops. The GPGPU still requires two external PCIe power connectors and consumes up to 170-watts of power at maximum. NVIDIA claims the Tesla D870 delivers 518 Gigaflops of GPGPU processing power.

Last year, the company announced a highly integrated graphics sub-system named QuadroPlex. Using a number of GPUs in a tightly integrated system, the QuadroPlex family of machines accelerated 3D rendering and graphics work. QuadroPlex became the stepping-stone for the new Tesla C870.

The Tesla C870 GPGPU server packs two GeForce 8-series GPUs in an external system with packaging similar to the QuadroPlex. The GPGPU delivers one Teraflop of GPGPU computing power while consuming up to 550-watts of power.

Finally, the Tesla S870 comes equipped with four GeForce 8-series GPUs and offers up to two Teraflops of computing power. The Tesla S870 consumes up to 800-watts of power and fits into a stackable 1U chassis.

Tesla C870 and S870 systems connect to workstation systems via an external PCIe Gen2 x16 interconnect. The machines contain PCIe switches and can be daisy-chained with more systems. As with the Tesla D870 GPU card, the Tesla C870 and S870 systems lack output capabilities. Theoretically, customers can purchase multiple Tesla GPGPU systems and chain them up for big increases in performance.

NVIDIA designed the new Tesla family for everything from graphics rending to medical research and data farming. At the core level, GPUs are far more efficient at dealing with parallel computing than general-purpose processors. This makes Tesla very powerful for cluster-type applications.

The Tesla S870, D870 and C870 carry an MSRP of $12,000, $1,499 and $7,500, respectively.

Comments     Threshold


This article is over a month old, voting and posting comments is disabled
Do the numbers make sense?
By HercDriver on 6/20/2007 2:18:47 PM , Rating: 4
If one GPU (C870)equals 518 GFLOPS, and 4 GPUs (S870)equals 2 TFLOPS, then how can 2 GPUs (D870)also equal 2 TFLOPS? Shouldn't the specs be 1 TFLOP? I'm no math PHD, but even Forrest Gump could figure that out. Unless there's something I'm missing?




RE: Do the numbers make sense?
By Anh Huynh on 6/20/2007 2:26:25 PM , Rating: 2
No idea. That's what NV documentation said. However, they could have different core clock and memory configurations.


RE: Do the numbers make sense?
By Anh Huynh on 6/20/2007 2:43:13 PM , Rating: 2
I stand corrected. It's supposed to be 1 Teraflop.


RE: Do the numbers make sense?
By sviola on 6/20/2007 4:22:33 PM , Rating: 2
You should review the whole phrase:

quote:
Finally, the Tesla D870 comes equipped with two GeForce 8-series GPUs and offers up to one Teraflop of computing power – similar to the S870.


The D870 offers half the computing power.


RE: Do the numbers make sense?
By nerdtalker on 6/20/2007 7:47:04 PM , Rating: 2
Those numbers are nice, but how is this going to integrate with desktop operating systems?

Does this thing run C code complied with some special Nvidia compiler? Do applications need to be specially written to work on it (besides being parallel enough for stream)?

I want one, but I want plug-and-play added horsepower. It's already complicated enough to get a grid up and working.


RE: Do the numbers make sense?
By defter on 6/21/2007 2:08:53 AM , Rating: 2
Tesla is programmed using CUDA which is a C like language (but isn't exactly C).


RE: Do the numbers make sense?
By PandaBear on 6/22/2007 12:34:34 PM , Rating: 2
This is not for general desktop use, it is for scientific simulation that runs horribly on CPU but good on DSP, you can say it is a compromise on a full scale DSP or FPGA that cost $10k each (yes, the high end cost that much).

They probably won't sell too many of those, but rather as a development platform that eventually will replace the need for certain task like wind tunnel simulation or off line signal processing.


Compared to Intel's 80 core chip...
By Asspollo on 6/20/2007 3:34:06 PM , Rating: 2
IIRC, Intel's 80 core chip got 1 TFlop at 100 W power consumption, or ~5.5x more Flops per Watt than this product. This gives me high hopes for Intel's upcoming Larrabee GPU... if it comes anywhere close to the performance of the 80 core chip, it will be a real killer for GPGPU and should make a great GPU as well.




By kattanna on 6/20/2007 4:07:23 PM , Rating: 3
that was a test processor and the cores, all 80, couldnt really do anything usefull, like actual work.

but it was a proof of concept chip for making the interconnects for so many cores.


By killerroach on 6/20/2007 4:42:42 PM , Rating: 2
Also, keep in mind that Larrabee is an x86-based processor, while Terascale was something completely different entirely. I'm not convinced of the performance profile of Larrabee, but that's not Intel's concern... they're willing to give up theoretical performance in exchange for a product that developers have more experience coaxing performance out of (e.g. x86).


By arturnowp on 6/21/2007 3:23:12 AM , Rating: 2
but it's not out yet and it's 45nm, G80 is 90nm...


By crystal clear on 6/21/2007 5:43:17 AM , Rating: 2


The 80-core mystery



Ever since Intel showed off its 80-core prototype processor, people have asked, "Why 80 cores?"

There's actually nothing magical about the number, Bautista and others have said. Intel wanted to make a chip that could perform 1 trillion floating-point operations per second, known as a teraflop. Eighty cores did the trick. The chip does not contain x86 cores, the kind of cores inside Intel's PC chips, but cores optimized for floating point (or decimal) math.

Other sources at Intel pointed out that 80 cores also allowed the company to maximize the room inside the reticle, the mask used to direct light from a lithography machine to a photo-resistant silicon wafer. Light shining through the reticle creates a pattern on the wafer, and the pattern then serves as a blueprint for the circuits of a chip. More cores, and Intel would have needed a larger reticle.


http://news.com.com/Intel+readies+massive+multicor...


my head hurts
By colonelclaw on 6/20/2007 5:28:00 PM , Rating: 2
disclaimer: i find this product very confusing and i'm probably a bit thick

ok this is confusing the hell out of me. gpgpu computers rely on using graphics cards to perform general computing tasks. so does that mean gpus are now better than cpus? if not why would this product exist? if so, why not replace all cpus with gpus? and why an earth does nvidia charge so much for this product? as far as i can tell they are just a couple of graphics cards in a fancy box, so why does that make them so much more valuable?

i'm not trying to be clever or jokey here, i really don't understand this product. thanks in advance to someone who can explain it all to me




RE: my head hurts
By tehliberal on 6/20/2007 6:21:04 PM , Rating: 5
Try google and wikipedia


RE: my head hurts
By peternelson on 6/25/2007 2:13:45 AM , Rating: 2
I'll explain it in simple terms:

It's a graphics card that you can't plug a monitor into.

But there's an upside: you can pay more money for it too ;-)

Beyond that, if you were to read the NVIDIA books GPU GEMS 1 and GPU GEMS 2, you may get some idea of what is possible on the last generation of GPUs (DirectX9, Shader model 3), let alone current (DirectX10, Shader model 4) or future ones.


They are still single precision units!
By Lord Banshee on 6/20/2007 8:02:45 PM , Rating: 3
From what i know "MOST" R&D houses will not use this unless they come with double precision. All GPUs today are single precision as games do not require so detail in decimals. But for work on volcanic eruptions, impact analysis, and 90% of all HPC software works with double precision.

So the question is this just prototypes for nvidia and ati to show what is coming or are HPC houses really buying these and are they ok with single precision?

On another note, i heard that Nvidia's next GPU will be double precision maybe for this reason alone. If this is the case they see some large amounts of cash from such a new market.




RE: They are still single precision units!
By Ajax9000 on 6/20/2007 10:04:13 PM , Rating: 2
NVIDIA CUDA Release Notes V0.8:
Q: Does CUDA support Double Precision Floating Point arithmetic?
A: CUDA supports the C "double" data type. However on G80 (e.g. GeForce 8800) GPUs, these types will get demoted to 32-bit floats. NVIDIA GPUs supporting double precision in hardware will become available in late 2007.
http://developer.download.nvidia.com/compute/cuda/...

Unless they have quietly added 64bit to a new revision of the G80 (unlikely, and even more unlikely that they'ed NOT shout about it), you'd be mad to shell out $12k now for 32b if 64b is less than 6 months away.


By peternelson on 6/21/2007 12:59:22 AM , Rating: 2
Precisely! I'm in the market for this kit but full double precision would be much more useful. If the real big iron will be out by year end, that would make me inclined to hold off until then. Meanwhile, some report are that performance of directx10 games and demos give poor frame rates, therefore demand will exist for faster dx10 capable cards. Anandtech please find out any roadmap, schedule or product details for the cards with double precision gpus, and when they will be available in these form factors. Any rumours would be more than what we know now.


Tesla
By fic2 on 6/20/2007 4:31:08 PM , Rating: 2
So is Tesla the new marketing keyword? Is everything going to be blahblah Tesla or Tesla blahblah? Just want to be sure and add it to my resume if so.




RE: Tesla
By mars777 on 6/20/2007 6:47:43 PM , Rating: 2
The sad thing is they are using the surname of the Croatian inventor, more than surely without his permission since he is dead ...

It like me making a product called "Gates B. 2007"...
It is sad... but legal.


RE: Tesla
By spluurfg on 6/20/2007 6:57:40 PM , Rating: 2
I'm just happy they decided to use something resembling a real word or name. Somehow I find it hard to associate anything with Duron or Radeon. I guess Itanium and Geforce I can understand, but sometimes I think the made-up-buzzwords drift into the shameless territory.


PCIe Gen 2?