AMD's FirePro w9000/w8000 review thread.

[3DVagabond]

Remember this?

Well, it was apparently Mark Papermaster having a "wood screw moment". As many speculated, it's not the FirePro W9000. (Likely the 7990 incognito)




This is!

6x mini DP, as we all said it should be. :thumbsup:

AMD has announced their entire lineup of GCN FirePro cards consisting of 4 models. Not that too many here will really care, but it's interesting none the less.

News release at Hot Hardware

Edit: Hot Hardware has a review up of the W9000/W8000. At least in the apps they ran, it's not pretty.

Also Tom's and Legit Reviews have reviews. There are some mixed results.

Well, now we have a real review site to look at. Anandtech has posted a review as well. :thumbsup:

 

[BallaTheFeared]

Not intersted in owning, but I am intersted in seeing how this all plays out.

With Intel entering into the fray with Knights Corner it's clear they're aware of the value of GPGPU and are attempting to break in to the market themselves. It will be interesting to see what they can do with GPGPU while forgoing the gaming graphics portion within their design.

This is AMDs first real attempt at this market segment which played into the laughable performance increase over last gen for them, but marks a key move for them as a business, having $4000 cards on the market with substance at this point is good for them.

Meanwhile the fabled GK110 that everyone scoffs at is nowhere to be found. With so much pressure on the current GPGPU leader who've had this marketed cornered since G80 the big question now becomes, where is GK110 and will it perform.

 

[Arkadrel]

So.... fastest card for professional line, atleast until nvidia releases new cards,
maybe even after that since AMD focused so much on GPGPU this round.

Well, it was apparently Mark Papermaster having a "wood screw moment".

Just a non refernece/early cooling design. Thats alot differnt than holding up a "wood box" with real "wood screws" used on it, made to look like a graphics card, like nvidia did

 

[3DVagabond]

So.... fastest card for professional line, atleast until nvidia releases new cards,
maybe even after that since AMD focused so much on GPGPU this round.



Just a non refernece/early cooling design. Thats alot differnt than holding up a "wood box" with real "wood screws" used on it, made to look like a graphics card, like nvidia did

That was said tongue in cheek. I've seen the back of that card and it's definitely a dual GPU design. Maybe it was done as a jab at nVidia having to use 2x GPU's to make a competent compute capable card? It's obvious they had a single GPU design on the chart behind him. Who knows, but my comment was meant to be humorous.

Interesting they are introducing the whole lineup in one shot. That might be to send a message also that they are much more capable of delivering on 28nm than nVidia is.

 

[Arzachel]

Not intersted in owning, but I am intersted in seeing how this all plays out.

With Intel entering into the fray with Knights Corner it's clear they're aware of the value of GPGPU and are attempting to break in to the market themselves. It will be interesting to see what they can do with GPGPU while forgoing the gaming graphics portion within their design.

This is AMDs first real attempt at this market segment which played into the laughable performance increase over last gen for them, but marks a key move for them as a business, having $4000 cards on the market with substance at this point is good for them.

Meanwhile the fabled GK110 that everyone scoffs at is nowhere to be found. With so much pressure on the current GPGPU leader who've had this marketed cornered since G80 the big question now becomes, where is GK110 and will it perform.

I think you're looking at this from a different perspective. People working with CUDA will need a pretty huge iniciative to move everything to OpenCL. While more power for cheaper is nice, development tools, documentation, support and api penetration are far larger considerations, so Big Kepler can be late, underwhelming and expensive, but people and companies tied to CUDA will still get it. AMD (and Intel) are fighting an uphill battle here until more of the big players pick up OpenCL. Although, that seems more a question of "when" instead of "if".

 

[3DVagabond]

I think you're looking at this from the wrong perspective. People working with CUDA will need a pretty huge iniciative to move everything to OpenCL. While more power for cheaper is nice, development tools, documentation, support and api penetration are far larger considerations, so Big Kepler can be late, underwhelming and expensive, but people and companies tied to CUDA will still get it. AMD (and Intel) are fighting an uphill battle here until more of the big players pick up OpenCL. Although, that seems more a question of "when" instead of "if".

That's all very true. They aren't going to stick with a lower performance arch for very long though. Cost is a very strong determining factor to. The TDP of 274w is nothing for AMD to be proud of. Pretty expensive to run. The productivity has to be able to offset that.

The 6x mini DP would be nice to see carried over to consumer cards. So much more elegant a way to attach multiple monitors.

 

[lopri]

I think some are slightly confused on workstation cards with GPGPU cards. Not that one cannot perform both duties, but the usages are very different. Professional apps still use the same basic stuff that have been around on PC graphics area - i.e., DirectX & OpenGL. GPGPU, on the other hand, a totally different animal that is still at its infancy.

Reading the press release, these cards are meant for professional 3D graphics. Not exactly for GPGPU stuff.

 

[Cookie Monster]

I still think professional cards are ALL about the drivers period because in terms of hardware theres little to no difference between the gaming and workstation versions.

I guess nVIDIA has a habit of limiting the DP performance on its gaming line and unlocking them for the workstation cards but from experience its all about the drivers.

 

[Genx87]

I think some are slightly confused on workstation cards with GPGPU cards. Not that one cannot perform both duties, but the usages are very different. Professional apps still use the same basic stuff that have been around on PC graphics area - i.e., DirectX & OpenGL. GPGPU, on the other hand, a totally different animal that is still at its infancy.

Reading the press release, these cards are meant for professional 3D graphics. Not exactly for GPGPU stuff.

Yup this competes against the Quadro line, not Tesla.

 

[Skurge]

Yup this competes against the Quadro line, not Tesla.

Yeah, K10 and K20 are tesla cards right? What does AMD have to compete with that?

 

[Arkadrel]

I still think professional cards are ALL about the drivers period because in terms of hardware theres little to no difference between the gaming and workstation versions.

I guess nVIDIA has a habit of limiting the DP performance on its gaming line and unlocking them for the workstation cards but from experience its all about the drivers.

So the performance of a card doesnt matter for professionals ?
Nor does the price they pay?

Also first you make a statement about it all being drivers you pay for, then under that you claim that performance is much lower because artifical limiting (but that cant be why they pay more?). Doesnt that defeat your own argument?

"but from experience"? your a professional user, that does this type of work for a liveing?

 

[HurleyBird]

I think some are slightly confused on workstation cards with GPGPU cards. Not that one cannot perform both duties, but the usages are very different.

Reading the press release, these cards are meant for professional 3D graphics. Not exactly for GPGPU stuff.

Yup this competes against the Quadro line, not Tesla.

You're both wrong.

These cards have ECC and aren't castrated in compute, so yes, they definitely compete with Tesla. Simply having mini-DP connectors on the back doesn't make them worse for HPC. In the future I can see AMD selling a stripped down version without display outputs only for compute, but this is *definitely* a compute card. From AMD's website:

"It’s up to 3.9 times faster than competitive solutions in single precision compute performance, due in part to its innovative all-new Graphics Core Next (GCN) architecture that allows it to effortlessly process challenging compute workloads... ...With high performance computing (HPC) features like error correcting code (ECC) memory it’s the ideal choice for HPC workflows."

 

[AtenRa]

The same NVIDIA GF100/110 was used for desktop (GTX480-580), Workstation (Quadro) and HPC with Tesla cards.

Same here with AMDs Tahiti GCN chips.

 

[tincart]

the laughable performance increase over last gen for them

I'd hate to hear what you have to say about the increase from 580 -> 680.

 

[Genx87]

You're both wrong.

These cards have ECC and aren't castrated in compute, so yes, they definitely compete with Tesla. Simply having mini-DP connectors on the back doesn't make them worse for HPC. In the future I can see AMD selling a stripped down version without display outputs only for compute, but this is *definitely* a compute card. From AMD's website:

"It’s up to 3.9 times faster than competitive solutions in single precision compute performance, due in part to its innovative all-new Graphics Core Next (GCN) architecture that allows it to effortlessly process challenging compute workloads... ...With high performance computing (HPC) features like error correcting code (ECC) memory it’s the ideal choice for HPC workflows."

Yes, they can do compute. But nobody doing any serious HPC will be building their clusters with Quadro or Firepro cards.

 

[f1sherman]

You're both wrong.

These cards have ECC and aren't castrated in compute, so yes, they definitely compete with Tesla. Simply having mini-DP connectors on the back doesn't make them worse for HPC. In the future I can see AMD selling a stripped down version without display outputs only for compute, but this is *definitely* a compute card. From AMD's website:

"It’s up to 3.9 times faster than competitive solutions in single precision compute performance, due in part to its innovative all-new Graphics Core Next (GCN) architecture that allows it to effortlessly process challenging compute workloads... ...With high performance computing (HPC) features like error correcting code (ECC) memory it’s the ideal choice for HPC workflows."

Not that SP bears any importance
(it's the workflow performance that matters, not theoretical - see 7970vs680)
but this card is slower SP wise than Tesla K10.

AMD themselves compare it with Quadro(Fermi arch, bolded quote),
so NO, it's not meant to compete with Tesla.

 

[f1sherman]

Meanwhile

http://www.marketwire.com/press-rel...h-Kepler-Architecture-NASDAQ-NVDA-1687677.htm

 

[HurleyBird]

Not that SP bears any importance
(it's the workflow performance that matters, not theoretical - see 7970vs680)
but this card is slower SP wise than Tesla K10.

All I'm doing is quoting AMD. I'm quite aware of the difference between actual and theoretical performance. SP does in fact bear importantance for compute workloads that do not require high precision. Despite a 15% theoretical SP performance deficit vs. the K10, I imagine you'd be hard pressed to find SP workloads that favor the K10 thanks to GCN's many advantages in the compute arena.

AMD themselves compare it with Quadro(Fermi arch, bolded quote),
so NO, it's not meant to compete with Tesla.

Don't be dense. According to AMD they are targeting HPC with these cards. That means they're positioning it against Tesla. End of story.

As far as far as only comparing specs against Fermi-based Quadro cards, why wouldn't they make the comparison that shows their product in the best possible light?

 

[f1sherman]

First sentence in both of your paragraphs has a ring of truth about it.
Everything else is so easily squashable I can't even bother

 

[3DVagabond]

There's a review @ Hot Hardware. I've updated the OP with a link.

Looking at the benches they ran the performance is pretty dismal.

 

[HurleyBird]

Not surprised. AMD's Firepro drivers are terrible and have a history of holding back otherwise superior hardware.

 

[Jaydip]

Terrible performance,it should have beaten my old Quadro 6000 Amd really needs to hire some talented guys in their driver team.

 

[Ancalagon44]

Holy cow, why do they even bother?

They should have spent the money they spent on developing these cards on sending a nice letter to Nvidia saying "The market is all yours, kthxbye". Would have been better use of the money.

I've read in a lot of places that graphics professionals (ie modellers, animators) prefer Nvidia hardware, now I know why.

Its actually kinda embarrassing.

 

[boxleitnerb]

At the time they get their drivers together, Nvidia will stomp them with the K20. AMD has so much potential and they waste it time and again.

 

[Jaydip]

This is actually going against nv's quadro ,nv is preparing the new quadro k lineup