A look at NVIDIA's SLI Multi-OS and new QuadrosMonday 30th March 2009, 11:55:00 PM, written by Arun
NVIDIA introduced a number of new Quadro products today, along with SLI Multi-OS which, amazingly enough, allows for full 3D acceleration in guest OSes for virtualization. A beta version of their CUDA raytracing solution has also shipped to their partners in early March. We won't be covering everything in massive detail, but we did feel quite a few points as well as presentation slides were worth mentioning.
Before we begin, let's have a quick reality check: the professional market is really big, especially for NVIDIA. They had revenue of $693M (in 2008 alone) with very high gross margins, resulting in an operating profit of $322M. this is nearly three times as large as their GPU profit, and more than compensates their highly unprofitable MCP and handheld groups.
Part of the reason for this, of course, is NVIDIA's overwhelming position in the market - notice how their unit sales increased drastically in 2008 despite the major slowdown in the fourth quarter and AMD's unit sales seemingly decreasing below 2005 levels:
Where does that share come from? It's not just one big market, it's many small markets with different requirements and many of them want developer support or specific features. If you create a professional line-up aimed exclusively at the CAD market, you're only addressing a small part of this ~$700M industry; and even in CAD, the brand's reputation and the existing customer relationships mean a lot which explains why AMD hasn't had much success even there despite having had competitive products recently. NVIDIA has a great halo effect from being adopted in mission critical applications, and is currently required and specified by over 90% of Fortune 1000 companies.
NVIDIA gave many examples of niches they target, and they include product & industrial design, seismic analysis and visualisation, medical imaging, broadcast graphics delivery, film & post production, etc. - but that's not what we want to focus on today. Instead, they're introducing a new technology which will allow them to monopolize a different kind of niche, and one which they claim is fairly large: customers who need to run multiple operating systems at the same time. Here's the problem as it stands today:
NVIDIA's solution to this is to use a 'hook driver' that allows the guest OS in a virtualized system to have full access to fully benefit from one of the system's GPUs. This obviously improves productivity as well as saving a lot of money and space because you don't need two boxes, two CPUs, two noisy fans, etc. - meanwhile, NVIDIA's revenue remains identical because you can only expose one GPU per OS, and therefore still need two boards.
In fact, revenue per customer could even go up because the feature is not supported on the lower-end Quadros. Therefore it should come as little surprise that there is no plan to expose this feature on GeForce and that today it's only available on a $899 board based on a massively cut-down GT200b chip. Oh right, we didn't mention that yet, did we? So that brings us to NVIDIA's new Quadro line-up with both existing and new products:
We won't bore you with the too many specifics, but our understanding is that the first three boards are based on GT200b; however the FX3800 only has 192 SPs on a 256-bit memory bus making it substantially slower for bandwidth-limited workloads than the others. The FX1800 has 64 SPs and 192-bit, making it either a G94b or a G92b. The FX580 and FX380 both seem based on the G96b.
In addition to all this, you've also got new Quadro NVS models although some of those were announced a bit earlier. That includes the NVS 450 with four DisplayPort connections on a single bracket, something we first described in late 2007. These cards have had a lot of success for financial trading, but obviously right now most of the sales are in other areas like call centers and digital signage, as well as for general business applications in Vista.
They also talked about a few other things, including NVScale which is multi-GPU scaling solution that works very differently from classic SLI allows you to benefit from all the video memory on each board (which is very useful with massive datasets). However, it requires software modifications so that the driver has some knowledge of what's really happening in the scene. It's pre-integrated in three scene graphs including NVIDIA's own NVSG, but there's also a simple API for third-parties. You can read more about it here.
There's also 3D Vision Pro and SDI, but we won't go into that here. So finally, they also talked about their previously disclosed raytracing solution which will start shipping in the fall, but entered beta in early March. How is this related to the workstation business? Well obviously some customers like a fast but lower-quality raytraced preview before going for the full thing (which still has to be done on the CPU), but perhaps more importantly it can also be used for Real Work(TM). The rays are physically accurate yet can be computed at interactive framerates, and so the results can be used for measurement purposes etc. - for example, PSA in France (aka Peugeot Citroën) uses it for what we understood to essentially be safety engineering.
Related nvidia News
NVIDIA Fermi GPU and Architecture Analysis
NVIDIA's Parallel Nsight finally released
NVIDIA GeForce GTX 460 - GF104 breaks cover
PhysX87, ancient tragedy in 5 acts by RWT
So long, Chris, and thanks for all the fish
NVIDIA GF100 graphics architecture details
NVIDIA Fermi: new GPU architecture, starting with GF100
NVIDIA release OpenCL GPU drivers for Linux and Windows
NVIDIA GeForce GTX 275 at $250 to fight HD 4890