I recently wrote a blog in response to a CAD blog highlighting the value in professional graphics cards over consumer gaming cards. In particular I tried to communicate how the software and drivers associated with these cards are a mainstay of the product you get; along with the investment in ensuring they are optimized, tested, certified for demanding applications such as Catia, Solidworks, etc. I felt I should have included more details of what you get with a professional graphics card such as our NVIDIA Quadro or GRID products. There is an awful lot of testing, development and support to staff with these products but I feel that is actually reflected in the end product.
What you get with professional graphics above consumer gaming cards
Here’s a quick list of what are usually listed as the benefits…
Performance and features:
- Application specific feature development & tuning (Consumer cards cannot run SolidWorks RealView features)
- Ongoing Driver optimizations to maximize GPU features
- IT tools for easy deployment & management
- Unique features to support pro workflows: Quadro cards feature an enormous range of professional graphics features that a gamer will never need: Warp&Blend, Mosaic, Bezel Correction and Overlap, Iray Server streaming, Quad-buffered stereo, sync, PBR (Physical Based Renderers)… and a whole load of others
- The software model of GRID 2.0 has enabled new features to be added without buying new hardware GPUs adding features such as H.264 hardware encode for VMware Blast Extreme
- Teams of staff dedicated to working with application ISVs to help them optimize their software – see my post yesterday on our work with Esri on ArcGIS alone: https://virtuallyvisual.wordpress.com/2016/03/17/esri-arcgis-and-nvidia-grid-an-awesome-list-of-blogs-to-find-out-more/
- Application specific testing by NVIDIA
- 100+ professional application certifications by ISVs
- Certified across workstation and server platforms by leading OEMs
- Unified rock-solid driver with deterministic release schedule for partners to QA against and IT to plan
- Designed and built by NVIDIA to a single specification for 24×7 reliability & stability
- Deep workflow experience across vertical industries such as AEC/Architecture, Manufacturing and Education
- Long-Life Cycle availability and support
- Bulk availability for large Enterprises
- Global technical pre-sales and post-sales support
- With the GRID 2.0 products this has introduced enterprise level SUMS support to align with that offered by virtualization partners such as VMware and Citrix
Gaming/Consumer cards are built under license to NVIDIA’s designs and manufacturers have leeway in the quality of components they use and adapt components to fit price-points and the target lifetime of the cards. Many gamers expect to replace their cards at least annually and manufacturers often design and test for short-life times.
HP have done a rather good job, explaining some of the differences this actually results in, here: http://h20195.www2.hp.com/V2/GetPDF.aspx/4AA5-0490ENW.pdf
My original blog was in response to one by Ed Lopategui at GrabCAD that questioned the value of professional GPUs and used some evidence of a benchmark that showed a Quadro card being outperformed by a Consumer card with an Autodesk Benchmark.
Did the evidence show that gaming cards performed better
Ed presented some evidence on certain benchmarks that showed some CAD applications performing better on similar gaming cards to professional cards. The data was quite old (2013) but I think Ed raised valid points as unless you work very closely with driver and GPU development, knowing what is involved in professional graphics GPUs isn’t top of your list to find out and also when we do charge a premium which I do think offers good value we need to justify it. Ed referenced some Cadalyst benchmarks that included Autodesk, where a consumer card “outperformed” a professional card. I have a long nemesis with Cadalyst and also AutoCAD results being extrapolated as representative of most CAD software.
- I don’t like Cadalyst alone used as a benchmark as it collects raw frame rates etc., but doesn’t examine frame content. I got caught out by a similar benchmark (and blogged about it). OpenGL and DirectX drivers often have fallback paths so if sophisticated graphics effects aren’t available they fallback to less intensive ones. Under Cadalyst how can you really tell if soft-shadow etc. are being executed or falling back to simpler, less computationally intensive but less visually impressive graphics?
- Cadalyst is now very old and I don’t particularly like the workloads or workflows it uses, they aren’t in my opinion very representative of a CAD user and don’t exercise many graphical APIs.
- https://www.youtube.com/watch?v=RLFEFYymk1M. There is a huge amount of very simple analytic geometry on the parts, lots of 2-D lines etc that simply doesn’t reflect the tolerant or 10e-11 NURBs geometry you’ll find in a CAD product such as Solidworks or Siemens NX.
- AutoCAD price-wise is an entry-level CAD product (a few £100s a license (?haven’t kept up) vs. £3000+ for Solidworks); the use cases, workflows and complexity of geometry are very different to most CAD applications. AutoCAD is also very CPU intensive rather than GPU intensive, we’ve frequently been able to virtualize 20+ AutoCAD users on a single card where that card could only support 4 heavy CATIA users.
- Consumer cards are frequently shipped overclocked. In a professional engineering department, reliability and purchasing cycles require an extremely low failure rate over a long-life time. On a CPU intensive workload like Autodesk with simpler GPU demands you may see this have a “performance benefit” in the short term – how long before your card burns out though?.
- The benchmarks referenced, yet again used wireframe and “realistic view mode”, I’ve long blogged about how these are precisely the tests used by sales/demo folk that mislead real CAD users. There’s an awful lot of manic rotating parts in Cadalyst, https://www.youtube.com/watch?v=RLFEFYymk1M
- I blogged here about my specific issues with rotating, panning views of shaded parts, here.
- Particularly when virtualizing I’ve seen Cadalyst abused to make performance assessments that in my opinion are pretty meaningless.
- The GRID technologies enable vGPU (GPU sharing and resource consolidation), no real engineering company has dozens of engineers running Cadalyst simultaneously in a frenzy of part rotation. CAD usage and workloads are typically very bursty, users spend a lot of time pausing, going to the toilet, looking at parts etc. Why benchmark a workload that bears so little resemblance to any real usage???
- With the cards designed for the GRID vGPU virtualized sharing, I’ve frequently seen users turn off vsync and FRL (frame rate limiter – a feature designed to help fair sharing and also limit frame rates to what VDI can cope with). This is usually to assess the raw power of the card in some way as that’s how these cards were assessed in physical workstations. The FRL is feature that is critical to VDI enablement ensuring applications generate sensible frame rates that can be remoted. You do not want to bombard a VDI network with ridiculously high frame rates, limit your server scalability nor pay for the bandwidth to do so. Again so artificial I fail to see the use in such a benchmark if you are planning VDI.
Consumer cards come to market faster than professional graphics GPUs
Ed’s article singled this out in a manner that to some extent suggested a secondary conspiracy that in some way CAD /professional users were being palmed off with out-dated hardware relative to consumer cards. Enterprise customers like things certified, tested and then test themselves. After all that when they are finally sure they like to standardize hardware.
- Variants are simply bad for QA, every new product adds another dimension to QA; the more varieties of cards a vendor produces the less each is tested.
- Software and Hardware vendors want long-term releases and availability
As I asked in my last blog:
Would you run your CAD software unsupported? Or Microsoft Windows?
The model of paying for the software development loaded into the GPU card is in my own opinion flawed. Enterprise software is about testing and support as well as interoperability with other products and hardware. If the price of that is loaded into a GPU, that in turn can be marked up by server OEMs leaving the user paying more to an OEM which is not allocated to development or support of the GPU software which is the main product a professional user requires. A GPU without drivers and software development is just raw silicon, a rather expensive paperweight or brick!
The sophistication of the software and support needed for GPUs today makes them comparable to an OS or hypervisor within the graphics stack. I simply can’t imagine any serious enterprise being willing to run their Microsoft OS, VMware stack or CAD software unsupported. The idea of just buying a GPU as a piece of hardware and having no guaranteed way forward if there is an issue with the driver seems a complete anomaly.