Last week AMD launched a new FirePro GPU aimed at VDI, say hi to the S7150 line. A new mid-range professional GPU isn’t really news but SemiAccurate thinks the hardware virtualization it offers is.
Lets start out with the hardware, and as the S7150 name suggests, these cards are for server use. Since VDI is by its very nature is server-centric, this is not a great shock, nor is the layout of the three cards. Two of them are labeled S7150, a single GPU, single slot, full height card, one with active cooling, one with passive.
The third card is the S7150 X2, a dual GPU, dual slot, but still full height device that only comes in passive cooling form. If you have spent any serious time in a datacenter, this won’t come as a shock, servers chassis are designed for airflow and the S-line is meant to take advantage of that hot air. Anything needing the density offered by the S7150 X2 doesn’t need additional fans, it would only make things worse in that setting. The two card specs look like this.
The spec in convenient table form
Tonga is the ASIC that both of these FirePros are based on, along with a ton of other W- and S- badged cards, plus an innumerable flock of Radeons. Like any other passive server card, the S7150 line has out-of-band temperature monitoring and no video outputs on the back-end to slow down airflow. There is also a fairly slick mechanism for telling if the card gets way out of it’s intended temperature range called APoGS or Acrid Plume of Grey Smoke. Just kidding, it will throttle back long before things get out of hand, really.
Things get interesting when you start looking at the drivers, they are the same as the W-series cards that lack hardware virtualization. If you don’t think this is a big deal, you don’t work in corporate IT, to them driver sprawl is a major pain point. And a major cost/ROI factor. Don’t overlook the value of this otherwise forgettable bullet point feature, it matters a lot to some very large customers. Similarly important is the fact that hardware virtualization means the drivers are native, not dependent on a software overlay/shim/VMM. This saves huge time and thus cost.
Each GPU can support up to 16 users doing light duty work, less if you are doing CAD or compute heavy tasks. Since the drivers are native, anything you can do on the metal you can do virtualized on the S7150 line including OpenCL in shared mode and low-level compute libraries. Once again this is a big step forward and it is all due to SRIO-V, other GPUs can’t come close to these features. This means the VMM sees the GPU as 16 physical units and that is where most of the driver commonality flows from.
On the OS side, things are a bit more mixed with certifications for VMWare vSphere/ESXi 6.x and Horizon 6.2, but that is about it. Shunning Hyper-V might cost AMD tens of card sales but supporting it would never pay for the certification costs much less make them money. On the client OS side at the moment only Windows 7 and 8.x are supported with 10 coming in the near future. Others may follow but that is the meat of the VDI market, real OSes have native tools to accomplish the same goals.
One huge differentiating feature is that AMD does not charge software licenses to do VDI like some others do, a massive cost headache that bloats the TCO of competing solutions. The price of the hardware, $2399/$3999 for 1/2 GPU cards, can pale in comparison to the software licenses involved. Native drivers and hardware virtualization mean there is no need for specialized driver development, and that cost is not passed on to the customers. If you are pricing VDI solutions, pay close attention to the hidden license fees before you get bitten, hard. For cloud providers, this is a killer feature.
So what is the big deal about these cards? Well the hardware virtualization allows for hard time slicing, at the moment it is round robin but there are hooks to skew things in later updates if needed. If you look at the hardware of any recent AMD GPU’s architecture, you will see multiple ACEs, Hawaii has eight for example.
While AMD would not specifically talk about how the virtualization is implemented, SemiAccurate thinks doing so at the ACE level isn’t it. Instead take a look at one of the least understood features of HSA called hQ. If you set up a queue or a queue group per user, hardware timeslicing would be almost trivial to accomplish as long as SRIO-V is there. Our bet is that hQ and HSA are how the S7150 line accomplishes its VDI mission.
So in the end the hardware is not new, the drivers are not new, the idea is not new, but the features exposed sure are. On top of that the business model behind the S7150 line is a shot across the bow for those charging exorbitant license and software fees for VDI solutions, this trio of cards will cut that revenue stream off at the knees. Between driver simplicity and TCO, this line will probably make cloud providers and corporate IT quite happy.S|A