Note thought, that AMD also labels the lowest performing card MI6 as an inference accelerator and the most expensive one MI25 as a training accelerator. They are well aware of different target markets.
Maybe, but more dice at lower clocks should be more power efficient. Leaning more towards an SSG design.
SSG does that with a Fiji, no reason Vega couldn't. I'd say it's rather likely. Even the K888 with 4x MI25 lists "Optional 8x NVMe support". SSGs did use two drives in RAID0.
Well you market what hardware you can
Does that make sense for a deep learning / inferencing card, though? I get that a dedicated SSD may be great for graphics and video workstations, but is it good for neural networks to have lots of data spread throughout "local storages" in each GPU?
They might not be local. Using a high speed fabric in place of a local drive could work for connecting a large volume of cards sharing a static data set. Even without the fabric each card in theory has more than enough space to hold most data sets without having to reach out to the CPU. Keep the GPUs largely self contained while they do their thing. You wouldn't necessarily need x16 PCIE lanes to the CPU as the cards wouldn't be going there often. No reason a single CPU with a bunch of switches couldn't handle 16 cards effectively for most workloads.
What if we're looking at a dual Vega? 8192 cores at 750MHz? If that's the top of the line card in theory they maxed out the die space leaving a single Vega 10 around the 600mm2 mark on par with Fiji. A pair of ~300mm2 wired together makes more sense. Each with 8GB HBM2 for the consumer Vega. Those would be fairly conservative clocks, but about as power efficient as you could get.
Given the size of the card shown as well as AMDs sweet spot of ~850-900 MHz for perf/watt in Fiji, I could follow your line of reaseon. However, with the move to 14nm, the sweet spot should have signifcantly moved up as well as 300 Watts with the more energy efficient HBM2 (compared to GDDR5X) already included and for "only" being slightly ahead of Titan X I would think this were no good signs for AMD. So I prefer to believe that this is a single GPU card with either some very complex power circuitry or some other nice surprises (SSD-attached, NCU "Network")
I'm saying two Vega10s on the same interposer providing a package roughly the same size as Fiji. So it makes sense that if AMD has 25% higher FP16 performance than P100 at ~600mm2 they would at least have have similar areas. Vega11 was in theory the "big" one. While possible, I doubt they made a chip significantly larger than 600mm2. Vega10 at ~300mm2 benchmarking a bit faster than an overclocked 1080 (314mm2) makes sense. Then a pair of them at lower clocks for the big card with double the ram. Otherwise they're horribly inefficient if 50% more bandwidth and double the die size is only marginally faster than GP104 in games, but outpaces P100 (which already has 2xFP16) in deep learning with 30% less bandwidth.
I see no reason why it wouldn't. Keep in mind that Vega 10 will probably be in >250W cards (save for an eventual Nano 2 solution), so unless it keeps the Fury X's AiO cooling solution then the cards themselves should be quite big.
