AMD Introduces the FirePro S10000 Server Graphics Card

[HumanSmoke]

Originally Posted by Xzibit

From the Press Release it looked to be aimed at Quadro.

Well, firstly, since there is no longer a FireStream product line, FirePro is now aimed at HPC (hence being unveiled at SC12) , as well as workstation. I'll let Dave Baumann (AMD Product Manager) make the distinction:

"S" pretty much stands for server, and these are targetted towards number crunching workloads not CAD workloads; Visualisation and sim are more maths problems. So these are more inline with prior "Firestream" offering than Quadro competitors

Secondly, there will be a GK110 Quadro. Bank it. There has never been a Tesla part that didn't have a Quadro counterpart

Originally Posted by Xzibit

The market i think they are aiming for is SP users with no DP or minimal DP workload who want to save $ & space.

Which is what I meant by "niche". SP performance in a relatively compact form factor ( a desktop ATX specification) where noise and performance per watt are irrelevant. What make it even more niche is the fact that the 7970X2/7990 offer the same functionality for single precision and cost a fifth of the price.

Of course, if the application requires single precision performance, a K20 isn't required. A K10 (4.58TFlop/s FP32) would suffice, and even using Amazon's/Sabre PC inflated prices ($3200) the numbers are a lot closer than what you're painting.

S10000...$3599...5.91TFlop....1.64 GFlop/$....15.76 GFlop/watt
K10........$3200...4.85TFlop....1.51 GFlop/$....21.56 GFlop/watt

[Xzibit]

Like I said I think its a happy medium in the price for those with no or minimal DP workload.

S10000
(SP) 5.91 TFLOPS
(DP) 1.48 TFLOPS

K10
(SP) 4.58 TFLOPS
(DP) 0.19 TFLOPS

[repman244]

Originally Posted by HumanSmoke

Secondly, there will be a GK110 Quadro. Bank it. There has never been a Tesla part that didn't have a Quadro counterpart

But another thing to consider is that this is the first time Tesla uses a chip that wasn't in the consumer counterpart first, which can be an indication that it was built for only Tesla in mind.

But that's just my guessing, only time will tell what exactly is the chip meant for

[HumanSmoke]

Originally Posted by Xzibit

Like I said I think its a happy medium in the price for those with no or minimal DP workload.

True enough to a degree, but then there is a reason that Nvidia command 80+% of the pro market- namely a better and more evolved professional driver and software environment. The toolkits offered by AMD aren't even close to that offered Tesla and Quadro (SceniX, CompleX and OptiX , or the CUDA SDK for example in relation to AMD's APP (ex-Stream) SDK). If hardware were the only criteria involved then AMD wouldn't be in the present situation regarding workstation and HPC GPGPU.

Originally Posted by repman244

But another thing to consider is that this is the first time Tesla uses a chip that wasn't in the consumer counterpart first, which can be an indication that it was built for only Tesla in mind.
But that's just my guessing, only time will tell what exactly is the chip meant for

Ryan Smith (Anandtech) seems to think a Quadro variant will eventuate (2nd page of the comments in the link below). Bearing in mind that ORNL were the lead customer and Nvidia seem to have their hands full satisfying Tesla demand...

Interestingly NVIDIA tells us that their yields are terrific – a statement backed up in their latest financial statement – so the problem NVIDIA is facing appears to be demand and allocation rather than manufacturing.

...it would seem that Nvidia are fulfilling HPC contracts first.
What would you expect Nvidia to do with the full 15SMX GPU's and other GPU's that are likely to fall outside of server power budget? It would seem you either fuse off perfectly good blocks for 15SMX parts, or throw away high leakage GPU's that could be utilized in a 250+W consumer card- and gain some PR into the bargain. Looking at Nvidia's past record, I'm pretty certain which course of action they would likely take.

[[H]@RD5TUFF]

Seems nice, and prolly makes a joke of most CAD processes.

[KooKKiK]

Originally Posted by HumanSmoke

375W TDP for a server part!...that should go down well.


S10000...1.48 TFlops FP64 @ 375W
K20X......1.31 TFlops FP64 @ 235W

Must be one hell of a niche market

AMD and nVidia doesn't measure TDP in the same way.

[3870x2]

Originally Posted by repman244

Less than Radeon cards, drivers are optimized for OpenGL (CAD etc.), some also have ECC GDDR which has it's impact.

They are usually on par, performing identical to their desktop counterparts. While they have additional parts to them, the GPU that would run graphics is very much the same.

My NV 4600 (8800GTX) can max out CS:GO and play anything out there currently. These cards still need to have the ability to run DirectX applications just as well as OpenGL.

[HumanSmoke]

Originally Posted by KooKKiK

AMD and nVidia doesn't measure TDP in the same way.

Who knows? Looking at non-boost cards from both AMD and Nvidia shows that each tends to use around 80-85% of TDP in average workloads, and ~95% for peak workload. A quick look at W1zzards power consumption charts for graphics cards reviews should bear that out.

If the S10000 isn't specced for 375W use;
1. Why does AMD specify 375W board power for a part which has no capacity for boost/overclocking?, and,
2. The "S" series are all passively cooled with the exception of the S10000 which obviously requires three fans. Why would that be?

[KooKKiK]

Originally Posted by HumanSmoke

Who knows? Looking at non-boost cards from both AMD and Nvidia shows that each tends to use around 80-85% of TDP in average workloads, and ~95% for peak workload. A quick look at W1zzards power consumption charts for graphics cards reviews should bear that out.

If the S10000 isn't specced for 375W use;
1. Why does AMD specify 375W board power for a part which has no capacity for boost/overclocking?, and,
2. The "S" series are all passively cooled with the exception of the S10000 which obviously requires three fans. Why would that be?

No. You don't understand.

for AMD, TDP means the maximum power that can be delivered to graphic board.

but for nVidia, TDP means power that limited by their power limiter.


take a look at HD6970 that has 250W TDP VS GTX580 244W TDP

how on earth that 6970 has higher power consumption than the beast 580 ???


then look at the 'real' power consumption tested by Wizz





even the next gen graphics, 7870 ( 175W TDP ) VS 660 ( 130W TDP ) also follow this trend.




So direct TDP comparison between both companies doesn't make any sense at all.



PS. the 6990 also has 375W of TDP.

[eidairaman1]

Its the same for Intel, there is no set standard of Measuring TDP, just like Monitor response times.

its just Denial some people are in

Originally Posted by KooKKiK

No. You don't understand.

for AMD, TDP means the maximum power that can be delivered to graphic board.

but for nVidia, TDP means power that limited by their power limiter.


take a look at HD6970 that has 250W TDP VS GTX580 244W TDP

how on earth that 6970 has higher power consumption than the beast 580 ???


then look at the 'real' power consumption tested by Wizz

http://tpucdn.com/reviews/HIS/Radeon...er_average.gif



even the next gen graphics, 7870 ( 175W TDP ) VS 660 ( 130W TDP ) also follow this trend.

http://tpucdn.com/reviews/Zotac/GeFo...er_average.gif


So direct TDP comparison between both companies doesn't make any sense at all.



PS. the 6990 also has 375W of TDP.

[HumanSmoke]

Originally Posted by KooKKiK

No. You don't understand.
for AMD, TDP means the maximum power that can be delivered to graphic board.
but for nVidia, TDP means power that limited by their power limiter.

True enough, but if you're measuring TDP when there are a couple of caveats. Firstly, power consumption measurement should be maximum load. We are talking about GPGPU here, and these boards don't spend much (if any) time at rest or under light load. Average load for gaming would would represent a heavily underutilized co-processor.
Secondly, the Fermi cards are blatantly fudged by Nvidia for PR purposes. I wouldn't argue that Fermi cards TDP aren't based on wishful thinking, but the issue here is Tahiti (S10000) and Kepler (K20)
Your analogy would be HD 7970 ( 250W board power) vs GTX 680 (195W TDP) - although that an apples-to-apples comparison either since Tesla lacks the boost facility. Closer would be a non-boost Kepler vs non-boost Southern Islands. But since W1zz hasn't tested any stock cards, maybe check out another site...
GTX 650Ti (110W TDP).......191W system load (173.6% of TDP)
HD 7850 (130W board).....216W system load (166.2% of TDP)
Not a huge difference.

Originally Posted by KooKKiK

PS. the 6990 also has 375W of TDP.

True. And it's maximum power draw is 404 watts.

TH put the single GPU FirePro W9000 through it's paces earlier (the S9000 is a passive version of the same card). The card is basically a HD 7970 non-GHz edition with 6GB VRAM and 225W board power. Under GPGPU the card clocked 275W

[KooKKiK]

you can't have nVidia card at its maximum power because power limiter will cut off anything before reaching that point.


look at the Furmark test of GTX680 by Wizz




and compare to the gaming test




you see a difference between Furmark and gaming test of 680 compared to 7970 ???

[eidairaman1]

ya NV forced any Voltage Mods out (EVBot being the biggest example of this)

Originally Posted by KooKKiK

you can't have nVidia card at its maximum power because power limiter will cut off anything before reaching that point.


look at the Furmark test of GTX680 by Wizz

http://tpucdn.com/reviews/KFA2/GTX_6...er_maximum.gif


and compare to the gaming test

http://tpucdn.com/reviews/KFA2/GTX_6...er_average.gif


you see a difference between Furmark and gaming test of 680 compared to 7970 ???

[HumanSmoke]

Originally Posted by KooKKiK

you can't have nVidia card at its maximum power because power limiter will cut off anything before reaching that point.

Makes no difference. The point is performance/watt, or in the case of servers/HPC, staying within the rack specification (more often than not) of 225W per board.

If a power limiter affected stated performance you'd have an argument, but as the case stands, you are making excuses not a valid point. And just for the record, the gaming charts don't have a direct bearing on server/WS/HPC parts- as I mentioned before, you can't get a true apples-to-apples comparison between gaming and pro parts- all they can do is provide an inkling into the efficiency of the GPU. If you want to use a gaming environment argument, why don't you take it to a gaming card thread, because it is nonsensical to apply it to co-processors.

Originally Posted by eidairaman1

ya NV forced any Voltage Mods out (EVBot being the biggest example of this)

Because volt modding is (of course) the first requirement for server co-processors [/sarcasm]
Take your bs to a gaming thread.

[Steevo]

Its true, when the GCN is in use under full load like GPGPU it is hungry, no parts of the chip are powered down and the marching instructions fit in the caches meaning transistors are busy.

But thats also why they manage a huge score in most GPGPU applications. And compare the power efficiency per FLOP even if it draws a few extra watts....

http://www.tomshardware.com/charts/2...mark,2971.html

http://www.tomshardware.com/charts/2...ing,2970.htmlb

[Xzibit]

Originally Posted by HumanSmoke

Makes no difference. The point is performance/watt, or in the case of servers/HPC, staying within the rack specification (more often than not) of 225W per board.

Maybe I dont understand what your saying but 225w rack specification ?

How old is that PSU in there. Swap it for a new one with correct connections.

[HumanSmoke]

@Steevo
Fix yo links.

True enough that Tahiti/GCN is optimized for GPGPU, but then GK104 is just the opposite....and if the S10000 were a desktop gaming card I certainly wouldn't disagree with the premise, but if server GPGPU is the point of the discussion- and it should be for this thread- shouldn't the comparison be between server parts? Seems a little pointless making a case for the S10000 using desktop cards running at higher clocks using desktop drivers, while comparing them to deliberately compute hobbled Nvidia counterparts.

Wouldn't a more apropos comparison be gained by testing server parts to server parts ?
(BTW: The W/S9000 is a Tahiti part (3.23TFlop), the Quadro 6000 is a Fermi GF100 (1.03TFlop) based on the GTX 470 ).

Originally Posted by Steevo

But thats also why they manage a huge score in most GPGPU applications. And compare the power efficiency per FLOP even if it draws a few extra watts....

Single precision......................................... ..............Double precision
W/S9000.(225W)...3.23FTlop....14.36 GFlop/watt.......0.81 TFlop.....3.58 GFlop/watt
S10000...(375W)...5.91TFlop....15.76 GFlop/watt........1.48 TFlop....3.95 GFlop/watt
K10........(225W)...4.85TFlop....21.56 GFlop/watt........0.19 TFlop.... Negligable
K20........(225W)...3.52TFlop....15.64 GFlop/watt........1.17 TFlop.....5.20 GFlop/watt
K20X......(235W)...3.95TFlop....16.81 GFlop/watt........1.31 TFlop.....5.57 GFlop/sec

And, for all the hoo-hah regarding the S10000 powering the SANAM system to number two in the Green500 list, the placement still relies more upon the asymmetric setup of the computer. 420 S10000's vs 4800 Xeon E5-2650's

Originally Posted by Xzibit

Maybe I dont understand what your saying but 225w rack specification ?

Servers and HPC racks in general are built around a 225W per board specification. Example HP , and from Anandtech...

K20X will be NVIDIA’s leading Tesla K20 product, offering the best performance at the highest power consumption (235W). K20 meanwhile will be cheaper, a bit slower, and perhaps most importantly lower power at 225W. On that note, despite the fact that the difference is all of 10W, 225W is a very important cutoff in the HPC space – many servers and chasses are designed around that being their maximum TDP for PCIe cards –

for example. That is why pro co-processors and are invariably rated at 225 watts. Check the specifications for top tier FireStream, FirePro, Quadro and Tesla. All the top SKU's are geared for 225W power envelope.

[Xzibit]

Originally Posted by HumanSmoke

Servers and HPC racks in general are built around a 225W per board specification. Example HP , and from Anandtech...

for example. That is why pro co-processors and are invariably rated at 225 watts. Check the specifications for top tier FireStream, FirePro, Quadro and Tesla. All the top SKU's are geared for 225W power envelope.

Still dont get what your implying

Right-sized HP ProLiant power supplies from 460 Watts at 94% efficiency, 750 Watts at 94% efficiency, to 1200 Watts at 94% efficiency

[Steevo]

Originally Posted by HumanSmoke

@Steevo
Fix yo links.

True enough that Tahiti/GCN is optimized for GPGPU, but then GK104 is just the opposite....and if the S10000 were a desktop gaming card I certainly wouldn't disagree with the premise, but if server GPGPU is the point of the discussion- and it should be for this thread- shouldn't the comparison be between server parts? Seems a little pointless making a case for the S10000 using desktop cards running at higher clocks using desktop drivers, while comparing them to deliberately compute hobbled Nvidia counterparts.

Wouldn't a more apropos comparison be gained by testing server parts to server parts ?
(BTW: The W/S9000 is a Tahiti part (3.23TFlop), the Quadro 6000 is a Fermi GF100 (1.03TFlop) based on the GTX 470 ).


Single precision......................................... ..............Double precision
W/S9000.(225W)...3.23FTlop....14.36 GFlop/watt.......0.81 TFlop.....3.58 GFlop/watt
S10000...(375W)...5.91TFlop....15.76 GFlop/watt........1.48 TFlop....3.95 GFlop/watt
K10........(225W)...4.85TFlop....21.56 GFlop/watt........0.19 TFlop.... Negligable
K20........(225W)...3.52TFlop....15.64 GFlop/watt........1.17 TFlop.....5.20 GFlop/watt
K20X......(235W)...3.95TFlop....16.81 GFlop/watt........1.31 TFlop.....5.57 GFlop/sec

And, for all the hoo-hah regarding the S10000 powering the SANAM system to number two in the Green500 list, the placement still relies more upon the asymmetric setup of the computer. 420 S10000's vs 4800 Xeon E5-2650's

Servers and HPC racks in general are built around a 225W per board specification. Example HP , and from Anandtech...

for example. That is why pro co-processors and are invariably rated at 225 watts. Check the specifications for top tier FireStream, FirePro, Quadro and Tesla. All the top SKU's are geared for 225W power envelope.

http://m.tomshardware.com/reviews/fi...rk,3265-7.html


Funny how much faster the Radeon is.

[HumanSmoke]

Originally Posted by Xzibit

Still dont get what your implying

Well that's rather unfortunate.
Most other people would realize that 225W input power means 225W heat dispersal requirement, as well as power requirement.

Originally Posted by Steevo

http://m.tomshardware.com/reviews/fi...rk,3265-7.html
Funny how much faster the Radeon is.

Wow! Tahiti's faster than a GTX 470 in LightWave, Ensight, SolidWorks, bitmining, Luxmark and CAPS viewer. Colour me surprised. I am truly shocked and stunned!
Isn't it more surprising that the latest generation AMD GPU isn't overly convincing against a two generations old Nvidia GTX 470 in AutoCAD 2013, May 2013 and Siemens freeform modelling?
BTW
You missed out the Maya benches in the same review...
and you missed out the Catia benches in the same review...
and you missed out the Pro/ENGINEER benches in the same review...
and you missed out the Siemens Visualization benches in the same test...

Then of course you've got AMD's forte- OpenCL - which is also a very mixed bunch. AMD is strong in Image processing, but the Video benches and general benchmarks are pretty much a wash. Shouldn't Tahiti be putting up better numbers against a GTX 470 than this?


Not to worry though, I bet the GK104 and GK110 based Tesla and Quadro will be shit at everything- and if they aren't, you can just look at the pages you like- just like the TH review.
(Better not look at the HotHardware review)

[xorbe]

Quadro 6000 is like from July 2010 (about 2.5 years ago.)

[Xzibit]

Originally Posted by HumanSmoke

Well that's rather unfortunate.
Most other people would realize that 225W input power means 225W heat dispersal requirement, as well as power requirement.

I still dont get what you mean ?

Can you link me to the 225w specification. I've never seen it.

Originally Posted by HumanSmoke

Servers and HPC racks in general are built around a 225W per board specification. Example HP , and from Anandtech...

for example. That is why pro co-processors and are invariably rated at 225 watts. Check the specifications for top tier FireStream, FirePro, Quadro and Tesla. All the top SKU's are geared for 225W power envelope.

I'd be unfortunate if you were implying that 225w is the limit and kind of halarious

Linking to SL390s G7 implying thats the standard is baffiling.

• 2U half width tray offer GPU density, offering up to 3 GPUs (up to 225 watt) in the equivalent space as a 1U server
• 4U half width tray offer GPU density, offering up to 8 GPUs (up to 225 watt) in the equivalent space as a 2U server

The (up to 225 watt) is configuration where the available PSUs options only provide for (2) 6-pin connecters per slot. Not the 225w your implying per board

PCIe Gen 2 = 75w
(2) 6-pin = 150w (75w each)
Total = 225w
Not a Server Specification

[HumanSmoke]

Originally Posted by Xzibit

I still dont get what you mean ?

What I mean, and Anand for that matter, is that server racks are more often than not optimized for 225W per PCIE unit, both for cooling, power usage, and cabling. What's so hard to understand?

Originally Posted by Xzibit

I'd be unfortunate if you were implying that 225w is the limit and kind of hilarious

You mean:

Originally Posted by HumanSmoke

Servers and HPC racks in general are built around a 225W per board specification

What the above sentence actually says is that server racks in general are designed with a 225W board in mind. How you think that translates into a PCI-SIG specification is beyond me, because if that were the case, the K20X (235W) and S10000 (375W) would exceed it. Moreover, do you really expect large server farms, supercomputing clusters and data centres to use ATX PSU's ?
So, yet another instance of were Xzibit's reading skills don't reach the mark.

Just to reiterate. Most racks are pretty standardized which is why most vendors limit themselves to a 225W add in board. The other thing to consider is upgrades of previous generation systems- a 225W swap out for a 225W board is relatively painless. A swap out for a higher TDP board may require more extensive work.

Originally Posted by Xzibit

The (up to 225 watt) is configuration where the available PSUs options only provide for (2) 6-pin connecters per slot. Not the 225w your implying per board

Of course if I actually said anything like that...but I didn't. 225W is a general standard that server vendors have adopted- don't believe me, check Cisco, HP, Dell, Penguin or any other server manufacturer and see how many are 225W per add-in-board and how many are, say 300W PCI-SIG.

Of course I don't expect you to actually do this, since it require you to:
1. Be able to parse the information correctly, and
2. Require you to actually spend some time doing research, and
3. You'd give up as soon as you saw the number of vendors' models specced for 225W boards.

Given that you can't even work out basic information about a company or its ownership, I'm not confident you'll fare any better with a companies product line- so I'm not expecting anything else but some worthless trolling

Originally Posted by xorbe

Quadro 6000 is like from July 2010 (about 2.5 years ago.)

Yep. GF100 has since been superseded by GF110, which in turn has been superseded by GK104/110

[KooKKiK]

TDP doesn't stand for the 'REAL' power consumption.

and both companies do not measure TDP in the same way.


that is my point.

hope you understand.

[Xzibit]

Originally Posted by HumanSmoke

What I mean, and Anand for that matter, is that server racks are more often than not optimized for 225W per PCIE unit, both for cooling, power usage, and cabling. What's so hard to understand?

You mean:

What the above sentence actually says is that server racks in general are designed with a 225W board in mind. How you think that translates into a PCI-SIG specification is beyond me, because if that were the case, the K20X (235W) and S10000 (375W) would exceed it. Moreover, do you really expect large server farms, supercomputing clusters and data centres to use ATX PSU's ?
So, yet another instance of were Xzibit's reading skills don't reach the mark.

Hey, smarty pants all those cards still use 6-pin and/or 8-pin AuX connectors.

Originally Posted by HumanSmoke

Just to reiterate. Most racks are pretty standardized which is why most vendors limit themselves to a 225W add in board. The other thing to consider is upgrades of previous generation systems- a 225W swap out for a 225W board is relatively painless. A swap out for a higher TDP board may require more extensive work.

Of course if I actually said anything like that...but I didn't. 225W is a general standard that server vendors have adopted- don't believe me, check Cisco, HP, Dell, Penguin or any other server manufacturer and see how many are 225W per add-in-board and how many are, say 300W PCI-SIG.

My point is 225w is an implied specification. Nothing stoping someone from putting a higher TDP card there other than dated hardware.

If you have links i'd like to see them tho, remember..

Originally Posted by HumanSmoke

Verifiable numbers or STFU.