Nvidia said Pascal has 5-10x the performance of Maxwell in certain compute tasks, we now know from the drive PX2 tests that this is true.
No one ever claimed that rlaes directly to games.
PX2 hasn't confirmed that anywhere at all, not even close.
The slides they showed for PX2 are comparing a SINGLE Titan X compared to an entire PX2 unit. That is two gpus, and
FOUR SOCs.
The biggest increase in performance listed between a SINGLE Maxwell GPU and 2 discrete GPUs, two more GPUs in the Tegras and one SOC(looks like a fabric providing one from first glance) and an image processing SOC is in..... image processing.
The rest of it is 1Tflop increase from the gpu to the entire system and tripling the 'deep learning' flops, which with the 4 separate SOCs can't remotely be attributed to the Pascal GPU alone. The best thing Pascal does improved over Maxwell is lower precision, meaning 8Tflop SP is equivalent to 16Tflop half precision IF it scales perfectly, it may only provide a certain ratio as with double precision. Yet they claim 3 times the tflops in deep learning performance... quite obviously a lot of that gain is from something outside of half precision flops and from putting things not suited to the GPU onto the other SOCs.
Making claims about Pascal performance by comparing a Maxwell with 6 separate chips, 2 Pascals, 2 Tegras and 2 other SOCs is incredibly disingenuous. In no way can you take PX2 numbers and suggest how much Pascal has improved performance and they absolutely didn't confirm claims of 5 times the compute performance at all.
Half of the gains Nvidia have claimed in increased compute performance came from a different interface in bandwidth limited systems with 8 gpus and even then Nvidia felt the need to put a disclaimer in that they were '*very rough estimates'.
Actually I would say the best way to phrase what Nvidia claimed was up to 5 times the performance from certain compute tasks(all mixed precision or only half precision, NOT single precision to single precision or double to double precision improvements) and 2x that from preventing bandwidth problems in specific systems. SO Nvidia is only claiming up to 5x compute improvement, then 2 x those improvements from Nvlink, which is just an interface. It's not even about giving the card more bandwidth, it's about preventing bandwidth to that card dropping. IE one card in a system gets 16x pci-e, but 8 cards only share 2x lanes each(theoretically). It's not about giving one card 64x pci-e lanes worth of bandwidth, but in 8 card systems preventing it dropping to 2x lanes each, but say 4x lanes each, hence the 2x increase from Nvlink. I wouldn't call that part of the architecture.
Today if you could make a system with double the available pci-e lanes to the CPU it would improve gpgpu compute performance the same way and really can't be considered a compute architecture/performance increase.
