Driver war: How Nvidia just left AMD’s marketing foils pulverized and pays a lot more | Retro

1 - Von Treibern und vergessenen Zertifizierungen 2 - Benchmarksystem und Teil 1 der Ergebnisse 3 - Der Rest der Benchmarkergebnisse 4 - Zusammenfassung und Fazit

Already 10 years ago there were discussions about marketing films, drivers and certifications. So none of this is new. And I admit it, the title was meant to be a little provocative at the time. But it also reflects very well what can happen to a tester very quickly and unintentionally when he gets caught between the fronts of two world views without foreboding. But always nicely one after the other. At that time, I was preparing the new workstation graphics card charts for 2013. A total of 67 individual charts with 21 graphics cards (seven of which are consumer cards for better comparison) should provide an even better overview of the performance of a total of four different graphics card architectures. And then came what I have now dug out of my archives again in the following. A

At this point, I now seamlessly continue with the original and for the excerpt at that time, 10 cards and a benchmark had to suffice for now, because the problem is a rather special one.

1. Balance – Solidworks 2013 should do the trick

Without anticipating the result: In many pro applications, the more powerful copies of the older Quadro cards with Fermi architecture are currently still leading. GCN has caught up enormously in the workstation field and was even able to conquer the performance crown in some applications, but our freshly compiled benchmark suite did not yet stand for a real balance. Interesting that AMD itself recently referred me to Solidworks 2013 and also fed me with corresponding marketing slides, so why not pay the new star in the FirePro solar system a visit? In the slides, it reads like this:

2. Problems with the benchmark and the thing with the certification

Here, I must now first of all take a bit of a step back. I know and respect that in the pro sector, you generally only use drivers certified by the manufacturer for the respective application. AMD states “Nvidia driver 307.45, 311.15” on the mentioned slides for its own tests for this. Since I actually prefer the newer drivers whenever possible, but the 311.15 was not available, Nvidia referred me to the similar 311.35. The Solidworks homepage: Everything is in the green for FirePro cards:

It has to be added that Dassault even specified a 310.90 for Solidworks on its homepage, e.g. for a Quadro K4000, but only a much older version was listed on the Nvidia homepage itself. Pure chaos. I’m sure it wouldn’t have been noticeable at all if technical problems hadn’t delayed the benchmarks and forced me to test several driver versions against each other once again.

I noticed that the newer Nvidia drivers showed performance gains of over 100% compared to the old ones. However, since very low values were measured in the first runs with Solidworks 2013 and the SPECapc benchmark for all cards compared to a reference measurement, AMD had asked me for a phone call, but interestingly, it was no longer about my actual benchmark problem , but solely about testing only with certified drivers.

In principle, AMD is right at this point, because you will very likely not use non-certified drivers in practice.the certification by the software manufacturer guarantees full functionality with all program features, even if SPECapc “enables” functions in the benchmark run that are otherwise only reserved for certified drivers. Nevertheless, the verbally very dominant insistence on the certified version was quite striking, especially since the two benchmarks from the slides shown above want to suggest an absolute exceptional position, which is even very real if you really use the given old Nvidia drivers.

3. I ask Nvidia and first of all I am astonished

When asked about the fact that no really high-performance, up-to-date driver is certified for Solidworks 2013, I got the answer from Nvidia “that they had somehow lost focus on this application”, but that they would immediately catch up and push for certification. Since this certification has meanwhile been caught up on very short notice, I have again entered the market and recorded quite remarkable results: The now freshly certified driver 311.50 not only provides for considerably better results of the Quadro cards, but also leads to more than a doubling of the results in many areas. With which one would have learned two things again: PowerPoint slides are usually not worth the power that was needed to create them, and what is valid today can turn into the opposite tomorrow, so that one believes to need a run-up even to stand.

On the next two pages, I will anticipate (in abbreviated form) a few results from the big workstation special to show how a botched certification can lead to outbursts of joy on the other side and how quickly such an ideal world can be destroyed again.