NVIDIA and AMD Launch New High-End Workstation, Virtualization, and HPC GPUs

MojoKid writes "Nvidia is taking the wraps off a new GPU targeted at HPC and as expected, it's a monster. The Nvidia K20, based on the GK110 GPU, weighs in at 7.1B transistors, double the previous gen GK104's 3.54B. The GK110 is capable of pairing double-precision operations with other instructions (Fermi and GK104 couldn't) and the number of registers each thread can access has been quadrupled, from 63 to 255. Threads within a warp are now capable of sharing data. K20 also supports a greater number of atomic operations and brings new features to the table including Dynamic Parallelism. Meanwhile, AMD has announced a new FirePro graphics card at SC12 today, and it's aimed at server workloads and data center deployment. Rumors of a dual-core Radeon 7990 have floated around since before the HD 7000 series debuted, but this is the first time we've seen such a card in the wild. On paper, AMD's new FirePro S10000 is a serious beast. Single and double-precision rates at 5.9 TFLOPS and 1.48 TFLOPS respectively are higher than anything from Intel or Nvidia, as is the card's memory bandwidth. The flip side to these figures, however, is the eye-popping power draw. At 375W, the S10000 needs a pair of eight-pin PSU connectors. The S10000 is aimed at the virtualization market with its dual-GPUs on a single-card offering a good way to improve GPU virtualization density inside a single server." My entire computer uses less power than one of these cards.

Comments

Intel's Xeon Phi is also out

[symbolset]

Here is the spin from El Reg.

And they have an 8-board FirePro system running

[mozumder]

3.6 kilowatts, 16 GPUs:
http://fireuser.com/blog/8_amd_firepro_s10000s_16_gpus_achieve_8_tflops_real_world_double_precision_/

and to think all this comes from video games.

Re:And they have an 8-board FirePro system running

[knarf]

Those 8 TFLOPS would have landed it somewhere at the top of the #500 supercomputer performance list in November, 2011. ASCI White used 8192 375MHz Power3 cores to achieve this performance. It took up a fair bit of space and used 3 MW to run the machine with a further 3 MW needed for cooling. It had a theoretical processing speed of 12.3 teraflops.

Re:And they have an 8-board FirePro system running

[knarf]

Of course that should read 'November 2001', not 'November 2011'...

How long is the wait ?

[Taco Cowboy]

Right now they are all too expensive, and consume too much juice.

How long would the wait be before these things get to have pricetag that average Joe (well, advance version of average Joe) can afford ?

Re:How long is the wait ?

[hairyfeet]

I've found if you stay a couple of generations behind you can get excellent bang for the buck without breaking your wallet. For example I've had the HD4850 for a couple of years now, paid a grand total of $75 for a 256bit card that was nearly $300 when first released. Come Feb I'll be moving to an HD6850, it'll be around $90-$100 by then and again 256bit only with GDDR 5 instead of 3 and I have no doubt it'll be good for a couple of years if not longer.

Desktop Virtualization does graphics on a server

[billstewart]

Server virtualization doesn't really need this - running web servers or databases or name servers, which are all essentially fancy timesharing.

But "Desktop Virtualization" emulates your entire desktop as a virtual machine on a shared server, graphics and all, and just ships the rendered screens back to your desktop, accessible from anywhere, with RDP or VNC or whatever, kind of like a clumsy version of X Windows except you get to do full-scale graphics acceleration at the server farm instead of at your desktop. The mainframe IT crowd like it, because the PC on your desk can be dumb and low-powered, and the server back in the server farm they get to maintain can be big and fancy, and they can have better control over it than over your desktop, don't need to keep every bit of software up to date on everybody's remote PC, and it's generally easier to manage. And if you're logging into your work desktop from Starbucks, they don't have to protect it as thoroughly from everybody else there, and you can access your work Windows desktop from your personal iPad or your kid's gaming machine or whatever, and the company data's not very vulnerable because there's really nothing running on the remote machine.

And using this chip, they've got a lot more graphics horsepower available for rendering desktops, so for instance they can provide you with adequate performance for video editing, not just for email and word processing .

Re:My entire house uses less power than one of the

[Anonymous Coward]

so 375 watts at 12 volts is the same as 375 watts at 120 volts?

Yes, unless you're talking 375 amps at 12 volts vs 375 amps at 120 volts (which is quite a lot)

Re:My entire house uses less power than one of the

[epyT-R]

just wait until you wet your noodle on capacitive/inductive AC where watts is watts except when it's volt-amps, and efficiency is measured as a ratio between watts and volt-amps.

Re:Quick!

[Wonko the Sane]

Beowolf these mofos.

That's so last century.

These days it's, "imagine a Bitcoin mining rig of these"

Doesn't quite have the same cadence, though,

Re:I am confused.

[Martin Blank]

I'm hoping for a consumer-level version so I can run Linux as my primary and Windows as a secondary gaming OS and still get full GPU performance on both platforms.

Re:Desktop Virtualization does graphics on a serve

[Anonymous Coward]

I think you missed the point here. Even if the card is capable of doing super high res rendering, the network traffic would be the weak point. The graphics card is much much better off being put in the client than rendering to a bitmap that has to be transferred. (And remote desktop protocols try to compare what has actually changed in order to send the least amount of data, which will cause high processing load when you run something at a high FPS).

I think is for using GPUs for processing (Password cracking, protein folding, etc.), as many specialized computational software packages do.

Re:Thread sharing data?

[darkain]

SHARING IS CARING!

375 W

[timeOday]

Here is a dumb thing to say:

My entire computer uses less power than one of these cards.

Does the person who wrote this know how much a TFLOP actually is, let alone 5.9 TFLOPS (single precision) and 1.48 TFLOPS (double)? As an example, an Intel Core i7 980 XE does 109 GFLOPS double-precision. This is over 13 times that! It is really exciting to see the power of GPUs broadened to scientific computing in general. I doubt these cards would be cost-effective or are really intended for gaming.

Re:375 W

[serviscope_minor]

Does the person who wrote this know how much a TFLOP actually is, let alone 5.9 TFLOPS (single precision) and 1.48 TFLOPS (double)? As an example, an Intel Core i7 980 XE does 109 GFLOPS double-precision. This is over 13 times that!

By way of comparison, the Opteron 6174 can hit a bit over 180GFlops in LINPACK. On a shared memory machine multi socket machine, the efficiency is very high, so a quad socket Opteron 6100 box would probably manage in the region of 700GFlops.

Anyway, plugging the numbers, the 6174 has a TDP of 175W, making the GPU about 4x better interms of FLOPs / Watt.

That's actually less than I was expecting.

Still can't beat the grunt and efficiency and density of GPUs yet, as long as you can program the little buggers. They're still very limited by access to memory.

Re:375 W

[jeffb (2.718)]

This. Remember the Cray-1 supercomputer? $5-10 million, 100KW power draw. The standard against which other systems were compared through the early 80's.

This card provides more than ten thousand times the computational power, using less than one-half percent the electrical power.

The future really is a cool place.

Re:My entire house uses less power than one of the

[Anonymous Coward]

Volts = Difference in height between two water tanks
Amps = Flowrate in the hose connecting both tanks
Watts = Number of full water buckets * height difference you'd have to carry _per second_ to do the same
Energy = Total volume of water transfered * height difference

There is no "realistic HPC workload"

[S3D]

Specific algorithmic implementation and limitation imposed on it by hardware are wildly different. Some algorithms don't need If's and branch prediction, other do. Different algorithms have different memory access pattern, different complexity of the kernels(for GPGPUU) and different requirement to memory bandwidth. Even on GPGPU algorithms doing the same thing in CUDA and in OpenCL can have several times performance difference. And some algo consist of mostly matrix multiplication, and quite a number of useful methods reach peak GPGPU performance (for example PDE solution). You can't find consistent "average" HPC algorithm.

Re:Desktop Virtualization does graphics on a serve

[Anonymous Coward]

I've done several Virtual Desktop projects for "scientific users" using GPU based acceleration on the physical server hosting the VDIs. The benefits of a powerful GPU are many:

1) You can provide GPU accelerated applications to all or some VDI users (I'm talking about GPU for calculation, not necessarily rendering graphics).
2) You can provide accelerated rendering by pairing server-GPU with whatever GPU is on the endpoint desktop for a better user experience (especially over limited network bandwidth), using technologies like HDX-3D Pro.
3) You can share out that GPU in the server to tens or potentially hundreds of users at once, so getter a better return on your capital investment in such a high-price item.

Since I can have 3 or 4 GPUs in a single server, VDI projects tend to be limited more by memory (we use 1TB VMCo nodes, each can hold hundreds of desktops), or IOPS (rather, storage latency to be accurate), where we use a tier of solid state storage.

Whether its a good idea or not, there are still loads of scientific users out there who are graduating from using local tools on their local PC (rather than a dedicated compute/rendering farm/cluster) and they like to continue working that way. VDI+GPU+Big memory appliance+Solid State means they dont have to "re-train".

Re:My entire house uses less power than one of the

[serviscope_minor]

and efficiency is measured as a ratio between watts and volt-amps.

No, efficiency is still measured as useful work out / Joule in.

You are referring to power factor, not efficiency.

The two are related, since electricity companies bill you on the number of Volt-Amps * Time used, not number of Joules, since it's much harder to do the latter, and the former determines the actual current which determines most of the expensive things, i.e. transmission losses, wire guage (infrastructure costs), etc.

If you're seriously screwey on your power factor, you can buy a power factor corrector and it will probably pay off quite quickly if you're a big user.

Re:My entire house uses less power than one of the

[Anonymous Coward]

Ugh...

Repeat after me, Watt/hour = Amps/hours * Voltage

Amperage = Current over time
Wattage = Power over time

Current(Amp) != Power(Watt)

Btw, if you don't know, normal electrical devices are measured in hours, meaning X watt usage really means X watts over a period of an hour.

So 375 watts = total power
Higher voltage and lower current can equal the same power as lower current and higher voltage. Realistically, certain applications require certain voltages to work (high voltages for ovens as they are basically giant resistors). End result though, 375watts used in a device at 110v is the same as 375w used in say a 12v device. All that means is that the current draw is different.

Sincerely, a different AC who never bothered to create an account

Re:My entire house uses less power than one of the

[jones_supa]

so 375 watts at 12 volts is the same as 375 watts at 120 volts?

Yes, unless you're talking 375 amps at 12 volts vs 375 amps at 120 volts (which is quite a lot)

I think it's better to say 31 amps at 12 volts and 3 amps at 120 volts, both totaling to 375 watts.

Export restrictions

[hjf]

These things are regulated. Argentina tried to buy a few (5!) of their previous line, I think it was Tesla. The US government wouldn't allow it. Guess you have to be a NATO member to buy these.

Re:Desktop Virtualization does graphics on a serve

[Luckyo]

"Can use" =! "works well".