It was back in May when NVIDIA first revealed its intentions to bravely enter the cloud with its new VGX processor line-up and now, the firm has officially unveiled its first two models.
The NVIDIA VGX K1 is constructed from four low-end cores to achieve the maximum possible performance-per-watt ratio, whilst supporting 16GB of DDR3 RAM, the focus, of course, is to maximise the number of users that can be simultaneously virtualised on the card and to reduce the power footprint of each one. The K1 is targeted at large firms with a high number of users to cater to and, is capable of handling up to 100 clients.
The VGX K2, on the other-hand, features two high-end Kepler cores, not dissimilar to an underclocked GeForce GTX 690, however with 8GB of GDDR5 RAM, enabling a greater client count, which NVIDIA claims is a maximum of 64. The K2's focus is on delivering superior per-client performance, at the expense of power efficiency.
| NVIDIA VGX K1 | NVIDIA VGX K2 | |
| Number of GPUs | 4 x entry Kepler GPUs | 2 x High-end Kepler GPUs |
| Total NVIDIA CUDA cores | 768 | 3072 |
| Total memory size | 16GB DDR3 | 8GB GDDR5 |
| Max power | 130W | 225W |
| Board length | 10.5in | 10.5in |
| Board height | 4.4in | 4.4in |
| Board width | Dual Slot | Dual Slot |
| Display IO | None | None |
| Aux power | 6-pin connector | 8-pin connector |
| PCIe | x16 | x16 |
| PCIe generation | Gen3 (Gen2 compatible) | Gen3 (Gen2 compatible) |
| Cooling solution | Passive | Passive |
It's interesting to note that NVIDIA has been able to passively cool both boards and is using this fact to promote high 24x7 reliability, though, just how clock speeds have suffered in exchange remains to be seen.
With a little quick math and of course, assuming outrageously optimistic operating conditions and bottlenecks; at maximum user-loads, virtualised clients could expect the equivalent performance from their cards:
(768 | 16) / 100 = 7.68 cores | 160MB RAM
(3072 | 8) / 64 = 48 cores | 125MB RAM
Of course, the real magic behind virtualisation is that it's unlikely all users will be active or at full-load at any one time and so those needing performance should be able to claim a larger slice of the pie, at least for a while, though, games design firms or a CG houses should perhaps consider running fewer clients off a single VGX card!
Virtualisation is all well and good, however, access still typically requires a thin-client and, with a plentiful supply of efficient APUs on the market, it's a shame to see this power go to waste.
