Direct CAD News & Views from Kubotek USA

Kubotek NVIDIA Graphics Card Shootout, Part 2 of 2

Posted by John McCullough on Tue, Nov 09, 2010

Refresher

In my last post, NVidia Graphics Card Shootout, Part 1 (which I’m assuming you read carefully), I detailed frame rate testing done with large models in KeyCreator with both NVIDIA Quadro and GeForce based graphics cards by PNY. I listed the test results found while using the Hardware OpenGL graphics type setting. The $800 Quadro card did out-perform the $150 GeForce card in a majority of tests, but not in a dramatic way. The only noticeable performance advantage of the $800 Quadro card over the $150 GeForce card was with the Wireframe render mode.

The other choice for hardware graphics acceleration in KeyCreator is Direct3D, a Microsoft created competitor to OpenGL. Direct3D has been around for a while, developed extensively for the Microsoft Xbox gaming system, but has been commonly accepted as inferior to OpenGL for CAD. However, Direct3D has now been upgraded and designed into the core of Windows Vista and 7 which provides a performance advantage to Direct3D over OpenGL. The workstation used for these tests is running Windows 7 (Pro 64-bit).

 

Results 2 – Direct3D

The first set of tests used a complex 21 MB solid (53,000 facets) and 4 different PNY cards.  This model was not a challenge for any of the cards tested.  

GPU

Memory

Price

OpenGL

Direct3D

GeForce GT 220

1 GB DDR2

$79.99

64 fps

192 fps

GeForce 9800 GT

1 GB GDDR3

$139.99

62 fps

967 fps

Quadro FX 580

512 MB GDDR3

$147.79

64 fps

384 fps

Quadro FX 3800

1 GB GDDR3

$819.99

62 fps

967 fps

Frames-per-second in KeyCreator 9, Shaded HLR mode

I’ve left the OpenGL results here for comparison. The most obvious issue between Direct3D  and OpenGL results is that frame rates are hard to compare because the Direct3D frame rates are so much higher. In practice the super high Direct3D frame rates are not significant. It’s like the difference between two cars with top end speeds of 220 mph and 600 mph. If 220 is more than you’ll ever need what’s the point of counting past that?

This first set of Direct3D tests show that the same two cards out-perform the others. Again, none of these tests cause a performance problem with any of the cards.

The next set of tests used larger models to show more performance difference between the two best of these four cards.

Model

Model Size

Contents

$800 Quadro

$150 GeForce

2

42 MB

1,595 solids

34.3

29.5

3

221 MB

50 spheres with 500 holes each

76.9

64.1

4

153 MB

149 solids

32

34.3

5

170 MB

10 solids (perforated sheet metal)

175.4

175.4

6

90 MB

1,256 solids

27.8

28.2

Frames-per-second performance of KeyCreator 9, Direct3D, HLR mode

These tests show the same type of mixed results as the OpenGL results. The GeForce took the edge in two of the tests, the Quadro about 15% better in two of the tests, and one test a draw. A key difference between the OpenGL and Direct3D results is that none of the Direct3D tests dropped performance into the 20 frame per second or lower category that start become noticeable to the user. The GeForce performed better in Direct3D than with the same 5 models in OpenGL. The Quadro had better frame rates in Direct3D than OpenGL in 4 of these 5 tests.

The next set of tests focus on models with transparency.

  Model

Model Size

Contents

$800 Quadro

$150 GeForce

1

21 MB

1 solid, 50% transparent

148

148

2

42 MB

1,595 solids, 1 solid with 50% transp.

7.5

7.6

4

153 MB

149 solids, 2 solids with 50% transp., cutting plane

25.9

27

Frames-per-second performance of KeyCreator 9, Direct3D, Shaded HLR mode with portions transparent

The results with model 1 show frame rates dropping much more dramatically compared to the OpenGL test but both cards dropping equally in comparison to each other, just like OpenGL.  Across these three tests the GeForce card beat or tied the Quadro, and performed better than in OpenGL. It is interesting to see that the frame rates of Quadro card in Direct3D are lower with the two larger models compared to the OpenGL results.

Model 4 in OpenGL had shown the largest advantage of the Quadro over the GeForce in these shaded tests. In Direct3D mode the GeForce beats the Quadro. However, it is interesting to note that with this particular model the best frame rates were found with the Quadro in OpenGL mode.

As I discussed in Part 1, the Quadro is designed to handle wireframe data better than its GeForce sister. In OpenGL this was confirmed. The next set of Direct3D tests was done in Wireframe render mode to see if the extra features of the Quadro would kick in.

Model

Size

Contents

$800 Quadro

$150 GeForce

2

42 MB

1,595 solids

53.3

51.9

3

221 MB

50 spheres with 500 holes each

74

50.5

4

153 MB

149 solids

76.9

91.7

5

170 MB

10 solids (perforated sheet metal)

214.3

192.3

6

90 MB

1,256 solids

32

32

Frames-per-second performance of KeyCreator 9, Direct3D, wireframe render mode

The Direct3D wireframe frame rates were better than the OpenGL results in all cases. Again the results show the Quadro provides an advantage to processing wireframe graphics. However, the Quadro advantage in wireframe is much less pronounced in Direct3D than in OpenGL.

Conclusion

Based on the frame rate performance results obtained with the NVIDIA GeForce 9800 GT and Quadro FX 1800 across the large models tested on our Windows 7 workstation, Direct3D seems to be the better choice over OpenGL. When using Direct3D, the GeForce 9800 GT card was the clear winner over the Quadro FX 1800 in this shoot-out. At a $650 savings over the Quadro these results call into question the traditional wisdom of purchasing ‘professional’ grade cards for KeyCreator.

I believe this graphics card shoot out with KeyCreator was the first of its kind ever published. I am pleased that the results answered several questions. First they confirm that the Quadro firmware is improving performance for wireframe data. Second they show that Direct3D is out-performing OpenGL. Last they show that KeyCreator customers can save considerably on graphics cards without sacrificing frame rate performance.

NVIDIA GeForce 9800 GT

The GeForce 9800 GT was far and away the best value of the four NVIDIA cards tested.

Next

As with any successful experiment, more research is needed. The most burning next questions in my mind: is there a better value for KeyCreator users than the 9800 GT in the NVIDIA GeForce line; and would the ATI FireGL professional cards compare similarly?

Topics: RAM, KeyCreator, Graphics, GPU, OpenGL, NVIDIA, Quadro, GeForce

Kubotek NVIDIA Graphics Card Shootout, Part 1

Posted by John McCullough on Fri, Oct 29, 2010

Graphics Overview

The hardware inside a PC which handles video display processing comes in two basic grades; integrated and dedicated. Integrated graphics is what you’ll find in 90% of PCs sold and should be adequate for your Mom. The integrated graphics chip comes built directly into the motherboard and borrows whatever memory the system will loan it.  If you are doing any 3D anything (gaming or modeling) you should have a dedicated Graphics Processing Unit (GPU) with a separate supply of Video RAM (VRAM) to avoid potential performance problems. Dedicated graphics for desktops and workstations is sold on expansion cards that plug into the motherboard. There are around a half dozen common manufacturers of graphics cards such as eVGA, BFG, Matrox, MSI, PNY, and XFX. Today, nearly all of these manufacturers buy the GPUs for their cards exclusively from NVIDIA or ATI (AMD) whose brand names tend to overshadow the card manufacturers.

NVIDIA calls itself the World Leader in Visual Computing Technologies. They evidently do have a large lead in market share for dedicated GPUs. They do over US $3 Billion in sales annually. I’m not an expert on hardware, but from what I’ve read there is every reason to be impressed with their history of consistent innovation in a competitive market. So when NVIDIA tells you that you need to spend an extra $600-$3,000 on your dedicated graphics because you are a professional user of 3D, you might not question that recommendation. If you look at the NVIDIA partner list for these professional level graphics, you’ll see many of the big name corporations in engineering and graphics software. ATI also has a professional-grade set of products at similar price premiums over their consumer-grade products. So the whole industry seems to agree that these professional-grade cards are necessary. What kind of people would dare question the accepted practices of the entire design industry? I think you know the type…

Trouble

For me, it all started very innocently. I actually wanted to learn about the differences between the consumer and professional grades of cards so that I could make clear recommendations to our customers about why they needed the professional cards. The trouble was that when I looked around, I noticed that no one ever makes a direct claim regarding the capabilities of these professional cards over the consumer ‘gaming’ cards. I realized it could just be that the marketing people at NVIDIA, ATI, and all of their manufacturers don’t understand the differences, so they can’t communicate it. So I dug just a little deeper.

The most knowledgeable sounding information I found said that the GPUs are actually the same between these product lines and that the differences lie in the firmware on the card. That seemed a bit hard to believe. It almost sounded like there was something sinister being hidden in those vague marketing claims. More digging uncovered benchmarks (written by publishing companies who are supported by NVIDIA advertising) which confirmed that MCAD software designed to work with Quadro (certified by NVIDIA) perform better with Quadro GPUs. But I found still more underground discussions and benchmarks showing that a ‘soft-mod’ (firmware change) can be applied to the GeForce cards so that it reports itself as a Quadro to the driver. This soft-mod significantly increased performance in tested MCAD software (typically Pro/E, SoildWorks, CATIA, etc.). Apparently the GPUs used between the consumer and professional product lines ARE the same, at least in some cases.

But this is not necessarily sinister. It seems there are special features designed into the shared GPUs which are not useful for gaming performance so they are turned off via firmware for the consumer market. Games are usually dealing with 100% shaded polygon data. CAD data often is or includes curves and lines. Turning on these features for the ‘professional’ card pulls resources from the main shaded polygon display calculations to process the curves and lines with special techniques optimal to those data types. This whole strategy is now over 10 years old and conceived at a time when OpenGL was king.

Testing KeyCreator

So, theoretically, the professional cards should out-perform the gaming cards when asked to handle 3D wireframe data. The only way to know for sure what performance each GPU would get with KeyCreator was to conduct an experiment with several of both types. Instead of a subjective feel for how well the display updated, I wanted to measure actual frame rates. Frames-per-second (fps) is a standard measure of how smoothly the display of a 3D model can be manipulated - higher frame rate indicate better performance. Frame rates above 25-30 are generally acceptable. Rates below 10 are painfully frustrating.

To make these tests mean as much as possible I eliminated as many variables as was practical. I started by purchasing a new workstation with a 900W power supply to conduct the tests on (double the Watts of a typical machine). The extra available power should eliminate potential problems with higher-end cards that demand significant power. Next I decided to stick to a single brand of GPU and card manufacturer. PNY manufactures popular NVIDIA-based cards in the consumer line (GeForce) and the professional line (Quadro) so that’s what I used. This choice also allowed use of one driver version across all tests. To keep costs of this experiment reasonable I picked four PNY cards to represent the two product lines.

During testing I discovered that frame rates almost always climbed 5-10% over several sequential repeats of the test. The results I recorded were the maximum frame rate observed. I also found that closing KeyCreator and restarting it typically boosted frame rate results by 3-10%. Once I figured that out I re-ran all tests with a fresh launch of KeyCreator. However these results may still contain some variation due to such factors and potential heating of hardware over time which I did not attempt to measure or control.

Results 1 - OpenGL

I began by testing each of these cards in KeyCreator 9.0.3 with Graphics Type setting on Hardware OpenGL. Test 1a used a complex 21 MB solid (53,000 facets) in shaded HLR render mode. To make the test a little more challenging the model was made 50% transparent (with 3 layer depth peel) for Test 1b.

GPU

Memory

Price Paid US $

Test 1a

Test 1b

GeForce GT 220

1 GB DDR2

$79.99

64 fps

29.5 fps

GeForce 9800 GT

1 GB GDDR3

$139.99

64 fps

61.9 fps

Quadro FX 580

512 MB GDDR3

$147.79

62 fps

53.5 fps

Quadro FX 3800

1 GB GDDR3

$819.99

62 fps

62 fps

In OpenGL frame rates max out around 60-65 fps which is a nice smooth display no one would complain about. In test 1a, none of the cards were challenged. Test 1b results show that performance with the lowest priced card does start to drop off. However, the more expensive Quadro cards did not provide improved frame rates in KeyCreator over the reasonably priced higher-end gaming card.

From here I decided to look for larger models that might show more performance difference between the two best of these four cards.

NVIDIA GeForce 9800 GT and Quadro FX 1800

Model

Model Size

Contents

$800 Quadro

$150 GeForce

2

42 MB

1,595 solids

20

20

3

221 MB

50 spheres with 500 holes each

60.1

62.1

4

153 MB

149 solids

38.4

21.8

5

170 MB

10 solids (perforated sheet metal)

61.9

60.1

6

90 MB

1,256 solids

19.8

15

Frames-per-second performance of KeyCreator 9, OpenGL, Shaded HLR mode

Clearly a few of these models are taxing the abilities of these cards. This data shows that the Quadro FX 3800 performance basically tied the GeForce 9800 GT card with most of these challenging models. With one model (#4) out of six tested the Quadro card did clearly out-perform the GeForce card. That model (#4) was the only one in this set of tests which contained transparency (2 solids were 50% transparent), so my next test looked into that.

  Model

Model Size

Contents

$800 Quadro

$150 GeForce

1

21 MB

1 solid, 50% transparent

62

61.9

2

42 MB

1,595 solids, 1 solid with 50% transp.

9.9

6.6

4

153 MB

149 solids, 2 solids with 50% transp., cutting plane

32

22.9

Frames-per-second performance of KeyCreator 9, OpenGL, Shaded HLR mode with potions transparent

Neither card had any trouble adding transparency to model #1 which contains just one solid. This same test on model #2 demonstrates that calculating transparency display in relation to other non-transparent solids does present an additional challenge for both cards. The Quadro card handled the addition of transparency to model #2 better than the GeForce card but both dropped into a frame rate range I classify as poor performance. Trimming model #4 in half with a cutting plane hurt the performance of the Quadro but slightly improved the performance of the GeForce.

Theoretically, the Quadro is supposed to be handling wireframe data better than its GeForce sister. The next set of tests was done in Wireframe render mode to see if the extra features of the Quadro would kick in.  

Model

Size

Contents

$800 Quadro

$150 GeForce

2

42 MB

1,595 solids

40.8

30

3

221 MB

50 spheres with 500 holes each

62.1

49.3

4

153 MB

149 solids

60.1

30

5

170 MB

10 solids (perforated sheet metal)

61.9

61.9

6

90 MB

1,256 solids

30

20

Frames-per-second performance of KeyCreator 9, OpenGL, wireframe render mode

This set of results was the most conclusive of all OpenGL tests. Both cards had better frame rates on most of the models with the shaded facets removed from the display and only the wireframe edges shown. However, the performance of the Quadro card in this mode was 25-50% better than the GeForce card with most of the models tested. This makes perfect sense considering the theoretical advantages which the Quadro card should have with this data type.

Conclusion – OpenGL

The $800 Quadro card did out-perform its $150 GeForce sister in 9 of 15 OpenGL tests, most significantly when using wireframe display mode. This advantage may not matter since most users today are not using wireframe display mode frequently. Modern GPUs allow excellent to reasonable performance in shaded modes and models are usually far easier to visualize in shaded mode. In 4 tests the performance was tied and in 2 tests the GeForce out-performed the Quadro.

I think these test conclusively say that the Quadro does deliver some benefit.  Is it $650 worth? If you are running OpenGL, frequently use wireframe render mode, and have a nice budget, OK. If it was my money I’d pick the GeForce and spend the $650 savings on an upgraded processor, 2 GB more RAM, and a liquid CPU cooler - so cool.

Next Week

The same set of tests was done in Direct3D mode. Next week I’ll report on the Direct3D results.

Topics: RAM, KeyCreator, Graphics, GPU, OpenGL, NVIDIA, Quadro, GeForce

Does 64-bit Computing Matter for CAD Users?

Posted by John McCullough on Wed, Sep 08, 2010

john wright mccullough

What’s So Great about 64-bits?

happy intel guy

You’d have to be living under a rock to not have heard that the PC market is in the middle of a transition from 32-bit to 64-bit computing. However it’s very likely you have not heard much about how this change is going to help you. The reality is that it has not delivered much yet. This transition has been happening very slowly because there are three separate 64-bit transition steps which need to occur in series. The first step is the processor and motherboard hardware, second is the operating system and hardware drivers, and last are the programs. The big benefits of the change won’t start to come until all three steps are complete. The good news is, we are just about there.

The advantage of a 64-bit architecture is that it allows the PC to juggle much bigger chucks of data – like real big numbers. If you are not regularly doing math with numbers over 4 billion you might think this is not going to help you. The real advantage for CAD users comes in the area of large memory addresses. 32-bit systems are limited to handling 4 GB of RAM. They just can’t count memory addresses higher than that. 64-bit systems can count so much higher that the ceiling of the amount of RAM possible becomes far out of sight. Back in 1990 when I thought my new 32-bit 486 with 4 MB of RAM was cool I didn’t think I would need a thousand times that much RAM in 20 years. I’m just as sure today that I won’t need a million times as much RAM on my desk in 20 years, but the 64-bit processor architecture will handle it.

The big deal of additional RAM is speed. If you are creating large files or performing complex modeling operations that cause the system to access the hard drive based swap file this is a serious performance bottleneck. If the system has access to enough RAM that it does not need to touch the hard drive operations will complete much faster.  Hardware vendors have been quoted that pulling that first bit out of RAM is 60,000 times faster than pulling it from a hard drive. Secondly, a percentage of CAD users are dealing with very large files (200+ MB) and may have run into a memory ceiling preventing them from increasing the size of a single file. An additional benefit from access to more RAM is that this ceiling can be raised.

There is a possible minor downside to running 64-bit programs. 64-bit versions of programs actually run slightly slower in some cases than 32-bit versions of the same program. Remember that 64-bit is not designed to be faster than 32-bit, it is designed to handle more data. Tasks that are not shuffling large amounts of data are not hitting the 32-bit memory bottleneck and won’t benefit from the change. The largest performance hit I’ve seen benchmarked was 20% and in most simple benchmarks 32-bit to 64-bit performance on the same task was close.

 

What’s the hold up? 

conductor checking time

64-bit processors were first introduced to the PC market in 2003. Now that we are in mid-2010, you likely have experienced a 64-bit processor. Did you notice any difference?  You probably didn’t. Most of the 64-bit hardware sold continues to run in 32-bit mode. Even though 64-bit Windows operating systems have been available to the mainstream for two years now, today’s most popular PC operating system is still Windows XP, which is estimated to be over 99% 32-bit. Windows Vista has about 15% of the market but is still only at 11% 64-bit. In early July Microsoft announced that the 64-bit version of Windows 7 has now reached 50% of the Windows 7 installs. It is estimated that Windows 7 has recently reached 15% market share. So don’t feel bad if you are not on a 64-bit operating system yet, you are comfortably in the majority.

A 64-bit operating system is not complete without the drivers to run all of your hardware. This has been a large part of the delay in getting users moved over to 64-bit operating systems. With Windows 7 64-bit most of those driver issues have been resolved (or the hardware has gotten old enough to forget about). If you have special hardware or utility programs you may need to do some hunting for 64-bit versions or replacements before you can use a 64-bit operating system.

Once you have a 64-bit processor and a 64-bit operating system now you are ready to start running 64-bit programs. However, you don’t need to. Just like 64-bit processors are mostly still running 32-bit versions of Windows, 64-bit Windows versions available today can still run 32-bit programs without much trouble. The Windows 7 64-bit has proven itself reliable and can run 32-bit programs very well. For most users buying a new machine, there is no reason not to go with the 64-bit version of Windows 7. Even if you do stay with a 32-bit program in 64-bit Windows there is a significant RAM benefit. By default 32-bit operating systems limit programs (each process) to access a maximum of 2 GB of RAM. Some users with the maximum 4GB of RAM on 32-bit operating systems have been using a special setting that will allow individual program processes up to 3GB. This same setting in 64-bit versions of Windows now allows processes from 32-bit programs access to the full 4GB of RAM that they can handle.

 

Conclusion

64-bit computing has arrived to the PC market. If you are working with large CAD files (over 20 MB) or performing complex modeling operations you are likely to benefit from speed improvements gained by the access to more RAM which a 64-bit architecture allows. This is a benefit you can begin to realize even with 32-bit programs running in 64-bit Windows. Full 64-bit computing (processor, operating system, and program) will allow access to huge amounts or RAM and eliminate the architecture-based memory ceiling problem with very large files.

The next release of KeyCreator will be available in two versions, 32-bit and 64-bit. The 32-bit version will be the right version for most users since they are still running a 32-bit version of Windows. The 64-bit version can only be used in 64-bit Windows. If you are running 64-bit Windows and have over 3-4 GB of RAM, you should consider installing the 64-bit version. Over the next several months Kubotek will be conducting 32-bit vs. 64-bit performance testing on a workstation with 12 GB RAM. These results will be published around the time of the release and make the advantages of the 64-bit version clear.

Topics: CAD, CAD files, 64-bit, RAM, large files