I caused a minor stir at Citrix Synergy last week when I commented on the price difference between NVIDIA and AMD’s virtual GPU platforms (to the point where someone walked up to me in the hall and said “You know NVIDIA is here at the show, right?”). I asked a half dozen people who I know have done vGPU work if they had any information on pricing for the complete solution–hardware and licenses–and they expressed the same frustration that I had: information was hard to come by, and when you did get something it was both ambiguous and expensive.
In fact, one person I talked to said that his customers were so turned off by pricing on NVIDIA’s M60 Grid 2.0 offering that they simply went back to buying K2 cards, which are cheaper than the M60 and do not require licenses. We all had a shared interest in the AMD platform’s cost and performance, but without hard numbers from NVIDIA it was hard to compare the former. The latter will require time more than anything since the hardware/software support with AMD is not as broad as NVIDIA’s.
Eventually I found what I was looking for (Thanks @kbaggerman): concrete pricing information in the form of a blog post from Jits Langedijk. NVIDIA’s licenses break down into three categories (as opposed to four or five in the original Grid 2.0 model): Applications, PC, and Workstation. Each can be bought as a perpetual or annual subscription license:
- Applications (RDSH workloads): $10 per user/year, $20 per user perpetual ($5/year maintenance)
- PC: $50 per user/year, $100 per user perpetual ($25/year maintenance)
- Workstation: $250 per user/year, $450 per user perpetual ($100/year maintenance)
When you talk to NVIDIA, you will no doubt hear the annual subscription price, and it seems like everyone leads with the Applications pricing tier which does not apply to VDI desktops. Just keep that in mind. When you hear “It’s just 85 cents per month,” that’s the lowest possible price you can pay for a Grid 2.0 license, and it doesn’t count the price of hardware.
So let’s look at the PC level, which is the lowest level that supports VDI. If we have 64 users in our environment, we’re looking at $6,400 in perpetual licenses. I like perpetual here because the cost after three years is the same as the cost of an annual subscription after three years. That price for virtual GPUs for users isn’t unbearable, but that’s only for software. You still need the cards, and pricing for that is harder to come across since NVIDIA leverages channel partners that pretty much charge what they can. Assuming you can get an M60 card in the $4,000 range, you’re looking at $14,400 (you need two cards to get to 64 users) to support high-end graphics acceleration in VDI desktops.
That’s a huge price tag, especially compared to the ~$8,000 total cost of the AMD offering for the same 64 users. We don’t yet know what performance/user experience is like with the AMD solution, though, so that comparison might not be apples to apples. NVIDIA certainly believes there’s a gap, because the week before Synergy they announced the M10 card, which has been optimized for user density instead of performance.
The M10 card features four GPUs, each with 640 CUDA cores (compared to the M60 which has two GPUs, each with 2048 CUDA cores) and 32 GB of memory. Each card is capable of supporting 64 users (double that of the M60), albeit at a good-not-great user experience. That’s ok, because this card is intended for task/knowledge workers, not workstation-level use cases. The cost of the M10 card really makes things interesting: $2,500. (Note, this is a ballpark figure that I received from people at Synergy, but also confirmed elsewhere).
If we adjust the cost model above for the M10 instead of the M60 (assuming the AMD 32 users per card number aligns better with the M10’s target audience), we see that the NVIDIA solution comes in at around $11,400 ($6,400 in licenses plus $5,000 in hardware) versus AMD’s $8,000 total cost, which is a bit easier to swallow if you’re married to NVIDIA.
The real proof is in the pudding, though. If AMD’s performance can rival that of NVIDIA, then companies are going to have a hard time justifying the extra spend and complex licensing scenario. NVIDIA has time to wait and see, as well as a lot of room to work on price since the bulk of the solution’s cost comes from licensing. I have hardware coming from both companies, and I can’t wait to put them to the test.