AMD FirePro W8000, W9000 Challenge Nvidia's Quadro - HotHardware
AMD FirePro W8000, W9000 Challenge Nvidia's Quadro

AMD FirePro W8000, W9000 Challenge Nvidia's Quadro

As of today, AMD has a new set of high-end professional GPUs available -- and a new shot at the professional GPU crown. It's a move the company hopes will allow it to take a greater share of the professional 3D business. AMD's battle against Nvidia in the professional graphics market is similar to the company's struggles against Intel ten years ago. In both cases, initial AMD gains have been partly reversed; Jon Peddie Research estimates that AMD now holds just 15% of the professional market, down from 18.4% six months ago.

The new W9000 and W8000 cards we're reviewing today are replacements for AMD's nearly two-year-old FirePro 3D V9800 and V8800. Memory bandwidth has increased by 80% and double-precision floating point performance is up to 1TFlop. On paper, these cards appear to be dynamite...

AMD FirePro W8000, W9000 Challenge Nvidia's Quadro

0
+ -

Im really not surprised honestly. AMDs last run of firepro GPUs was a joke.

0
+ -

Ask yourself why.

Seriously. Why?

Because as the reviewer, I don't have a satisfactory explanation for this behavior. It bucks every trend I'm aware of. "They're just a joke" is not a valid answer when examining performance that runs counter to hardware performance in every other field.

0
+ -

I have always wondered why this game (pro graphics cards) is either play very strong in the game or don't show up for it. They keep showing up though! I have used ATI/AMD now cards for quite some time. Yes i have had Nvidia cards as well but my first self built PC contained a (PII 400 pre-release by a week and 3 days) All in wonder pro. I still have it downstairs on a shelf and it is operational as well. In professional cards I do not ever recall them even having a chance though really.

0
+ -

AMD may be releasing the workstation cards now (at SIGGraph), rather than waiting until the drivers and ISV packages are optimized for GCN, if there's a significant opportunity for customers who use them as reliable 24x7 engines in render farms and media workstations. These customers may not be relying on the certified engineering packages tested in SPECPerf. AMD can show benchmarks like the ray-tracing in CLBenchmark.com instead.

(I think a previous generation initially exhibited a large performance regression in 2D performance. I can imagine the driver developers optimize the newer APIs first, perhaps assuming the latest most demanding software will use the newest APIs. So software using older APIs may initially suffer.)

(As I understand, even with standards like OpenGL or OpenCL there is still work to do to tune the kernels to the parameters of the hardware available, such as local memory size. I speculate that there may also be parameters in how the data is organized (tile size?), which limit the number of compute units, or lanes within a compute unit, an algorithm can use at once, analogous to how a 4-thread game can't make use of all of a 6- or 8-thread CPU. AMD may not have the resources or market share clout to convince the engineering package software companies to adapt their software to Tahiti/GCN on AMD's schedule.)

0
+ -

If we look at these benchmarks

http://hothardware.com/Reviews/AMD-FirePro-V7900-and-V5900-Professional-Graphics-/?page=6

we will see big improvements int Quadro 6000 perfomance. It is drivers impovements, or diferent system setup?

And every time in differents setups we will get very different results... As for Spec, they are so lazy and using very old versions of programs. I hope AMD continue close work with softwre developers.

0
+ -

Spec just released a new version of Maya; I'll be testing it.

 

Presumably that's driver performance improvements. SPEC workloads are designed to only stress the GPU; CPU changes should produce a minimal effect.

0
+ -

@HotHardware: some suggestions:

1. please include GPU driver version numbers alongside hardware model numbers on the test configuration table/page for any benchmarks, especially engineering software benchmarks.

(It will help readers compare numbers from different review benchmark runs. It will help users who come back to reviews trying to diagnose performance differences, and determine whether an driver upgrade may be responsible.)

2. please include output settings such as rendered resolution, color bit depth, AA settings, etc.

(It will help readers compare numbers from different review benchmark runs. It will help diagnose if a strong GPU has enough work to exhibit scaling over weak GPU. Top gaming card reviews often don't show scaling benefits unless there are large screens or multiple screens. The HH V7900 V5900 article tested at 2560x1600. Officially submitted results on the SPEC site seem to use 1920x1200 or 1920x1080, but I expect workstations with such high end hardware would often be connected to higher resolution screens, sometimes even with 10-bits-per-pixel IPS screens rather than normal 6 or 8 bits per pixel. Thanks for discussing vsync, but no other settings are listed. Maybe the article is missing a table?)

(3 nit: please add a legend to the first bar chart, catia and ensight. It was difficult to read the text the first time while trying to keep multiple possible interpretations in mind.)

@GKvarta: you may be confusing two different benchmarks. SPECviewperf Maya runs a trace of the graphics commands sent to the GPU, focusing on the GPU performance. It can be run without a copy of the application. SPECapc Maya runs the actual application, so it may stress the CPU more to generate the model and emit the graphics commands.

On the other hand, in the linked article, the SPECviewperf 11 Maya scores

Quadro6000: 115fps-at-2560x1600

Quadro5000: 99fps-at-2560x1600

are much more impressive than the top officially submitted scores on the SPEC site

Quadro6000: 115fps-at-1920x1080

Quadro6000: 107fps-at-1920x1200

Quadro5000: 95fps-at-1920x1200

because the 2560x1600 implies rendering near twice as many pixels per frame as 1920x1080.

The linked article used a i7 980X overclocked to 4.27GHz,

while the officially submitted results use safer stock-clocked Xeons (3.46GHz-3.73GHz)

which might account for a part of the difference, but maybe not all of it. (I wonder if that article's benchmark runs rendered at a full 2560x1600 as the chart titles say, or at a smaller resolution on a 2560x1600 screen.)

0
+ -

@gc9: good highlights. I like to read HotHardware, but sometimes benchmarking articles is too superficial. As for new FirePro card, I think it's more of preview, not the review, drivers is not ready (as was with 7970 at launch day). Feels like V8 engine with not enough feeding with petrol...

0
+ -

GC9,

The drivers used were AMD's provided (8.982.3) and Nvidia's Quadro 296.70. If AA had been enabled, I would've said so. 

2)  All tests were done using default settings. That means the monitor was set to 1920x1080 and 32-bit color. 

3)  "Top gaming card reviews often don't show scaling benefits unless there are large screens or multiple screens." This can be true, but evidence suggests it's not the case. Given their relative performance and data from the V3900 review (which I went back and eyeballed), it's incredibly unlikely that scaling is somehow limited to Quadro due to benchmark design. 

The HH V7900/V5900 article was written by someone with a 2560x1600 monitor. If you would like to A) Donate or B) Annoy my boss into buying me one, I would be happy to use a 2560x1600 display in future professional GPU reviews. Until such time as this occurs, one does what one can with what one has. 

4) To the best of my knowledge, none of the SPEC tests are configured to utilize 10-bit color workloads. In any event, I do not own a 10-bit monitor. 

0
+ -

@JH: Thank you for the driver version numbers, resolution, and color settings. 1920x1200 or 1920x1080 may be the resolution for official submissions,so that will help comparisons.  (My intent was not to lay any fault on you for rendering at the official submission resolution.  I was trying to come up with ideas for other causes of the unexplained lack of scaling.  I was also trying to explain why it may be reasonable for the prior review to maybe render larger than the official submission resolution.  Sorry if I raised your hackles.)

0
+ -

GC9,

No, not at all. Good feedback is how we create better reviews. If I sounded a bit tetchy, blame it on my own frustration at being unable to thus far explain these performance figures. 

I am reluctant to blame ROP output without more data. Consumer-side tests have indicated that GCN makes substantially better use of its 32 ROPS, even if it doesn't increase the number over V7900. 

0
+ -

I like this site for being only independent reviewer of professional graphics. And in CAD forums I always posting links to Your site. Keep that path and dig deeper. Cheers!

0
+ -

GKVarta,

Where do you post? Because I'd love to get some input from any knowledgeable CAD/CAM people on this.

0
+ -

Joel, one of my favorite is SolidWorks forum, or some discussions take place in linked in. Some link for starting search.

https://forum.solidworks.com/message/267903#267903

http://www.linkedin.com/groupItem?view=&gid=82201&type=member&item=136409652&qid=7f8540da-af1a-49db-9800-8a4c2285d5da&trk=group_items_see_more-0-b-ttl

0
+ -

Hm very solid site of CAD guru Anna, with benchmarks and another useful information.

http://www.solidmuse.com/2011/06/core-i7-2600-vs-dual-socket-xeon-x5690/

0
+ -

Tom's Hardware got some different results...

http://www.tomshardware.com/reviews/firepro-w8000-w9000-benchmark,3265.html

0
+ -

THG's results are odd. I'm guessing at least some of their tests had Vsync enabled, although it's supposed to be deactivated for a CADalyst run. I have no idea why they show CATIA performance on AMD running 10 fps below me or what the explanation for the other variances is.

0
+ -

I built a new PC last month but I held off buying a firepro v7900 as I thought I'd wait for the W series to be released. The results seem to be a little disappointing. I'm sure they'll get better with a few driver updates but still...

Do you know when they'll actually be available to buy? Apparently they've 'launched', whatever that means, but no where (including AMD's website) is actually selling them yet.

0
+ -

My firm buys workstation hardware, and I am tasked with testing them before placing orders.

What I have seen is that viewperf scales with engine clock, and what seems to be some vendor tuning that is specific to the benchmark (because I'm not seeing the gains in our applications).

It is a useless benchmark IMO, so I am always surprised when you reviewers use it. Based on the time I have spent with it, I am amused and pleased to see that you are perplexed by it (because I have had similar hardship). APC is better, but not great. In the end, we run our own models (we are spending hundreds of thousands of $). I'm guessing that you don't have that luxury! :)

We use Quadro and FirePro - I don't think that the engineers know the difference (if they do, they are not complaining about it - and engineers love to complain about their tools). On spec, the W8000 looks like a good product for us, but W9000, that's pretty expensive. Makes way more sense to equip two engineers with W8000s.

Hopefully you can figure workstation products out, and I can reduce my workload! :) It's only every eight months or so that we test cards, but it is still a pain.

Good luck!

0
+ -

MStowe,

You could possibly help with that, if you could get permission to share your own tests (or create equivalents that could be shared). We'd absolutely be open to some collaboration and independent testing, and you'd have access to data that would slim your own workload.

If that's impossible, there's not much more we can do. The problem with workstation product testing is that these suites are capable of an enormous array of tasks. It's extremely difficult to duplicate "typical" workloads when so much of the information on what those workloads are is restricted.

If there's a way we can work together, let me know. If you can share data on what applications your firm uses and what you do with them -- background data, not for publication -- feel free to email me.

0
+ -

Couple of problems for me there:

1. Our models have our IP in them, so they can't go out the door

2. I don't actually design in CAD (I use a macro of a workload) so I'm not in a position to create a model for you

A potential problem for you is that you don't have a seat for the applications you want to test. I'm guessing not since they are expensive.

I mentioned APC earlier - it also requires the apps, but has tests. It is not great because it uses static images. So, as an example, you can't see what the performance would be if you are rotating a model, which is obviously important - perhaps most important.

An option might be getting APC and some applications on trial period (not sure if they still do that).

0
+ -

1)  That's the problem I was worried about. Don't know if you have friends on the design teams who could whip up something that would create an equivalent load using simple shapes or stock / freely available scenes. 

 

2)  There's plenty of stock scenery around for most applications if the workload can be made to match it. Heck, the SPEC_APC tests themselves might be a suitable launch point. 

Application availability is a non-issue. Free trials are ubiquitous and typically last 30 days -- more than enough for testing. In the few cases where this has been an issue, we've been able to work with vendors to secure access.

0
+ -

I bought a V7900 anyway (I got it cheaply off ebay,new but OEM) and installed it today. For my current project I'm using Modo which, as I had seen from a review elsewhere, does not work too well with the firepro series, presumably because ATI didn't bother to optimize the drivers for it.

I think their lack of support for Modo sums up the difference between Nvidia and ATI pretty well. ATI (usually) offer more bang for your buck, certainly from a hardware point of view but Nvidia bother to put in the work and support to make sure their hardware is doing all it can do.

Login or Register to Comment
Post a Comment
Username:   Password: