Just out of curiousity why is HBM out of the question*? I think we discussed this like ~20 pages back but is the cost really that prohibitive? Like, do we know if it's something like 2-3x the cost of LPDDR4 or more like 20-30x the cost?
*I don't expect it because everyone constantly says not to but I'm curious about the reasoning.
It's probably too expensive, and it's probably consumes too much power, but it's hard to make definitive statements on either.
It's certainly unlikely, but we can't 100% rule it out, as we know that
Nvidia has a part in production with one logic die and one four-die HBM stack using TSMC's InFO packaging, and the Switch SoC is one of only two plausible candidates for that chip (the other being Xavier, which shouldn't technically be "in production" yet, for that matter). Using InFO would potentially bring the cost down quite a bit, although again it's impossible to say by how much.
Besides cost, power. Theoretical low for current HBM is 14.6W from what I found. While LPDDR4 would be 1.02W. HBM would kill batteries faster. Would be nice to have, but there would probably only be 2GBs at best inside the Switch in order to save as much battery as possible.
That's for four HBM stacks, whereas Nintendo would only be using one. They could also clock HBM2 quite a bit lower than the standard 2GHz (which gives 256GB/s per stack), potentially to 1GHz or below, which would reduce the power draw quite a bit (they could even clock it down in portable mode and up in docked mode alongside the GPU if they felt like it). Sub 2W power draw would in theory be possible, which is certainly still more than LPDDR4, but not impossibly so.
I thought HBM got into far lower power envelopes than that. The 7th slide at this
link says 3.3W at 128GB/s bandwidth for HBM, which I think could be clocked lower to suit the Switch's needs. Unless I'm reading that slide wrong which I very well could be.
Edit: Ah that might be for the new LPHBM which isn't available yet?
That's first-gen HBM. As far as I'm aware HBM2 consumes slightly more power operating at full speed (2GHz compared to HBM1's 1GHz), but should consume quite a bit less when clocked down to give ~128GB/s bandwidth.
At manufacturer's prices and at large volume it shouldn't make an enormous difference if it's 2-3x more expensive for one component. I would say it would take the BoM from, say, $175 to $185-190 or something. Completely out of my ass though.
According to IHS, the
4GB of LPDDR4 used in the Galaxy S7 cost about $25 for a launch in March this year. A year later, and assuming Nintendo wouldn't use PoP, it would probably cost Nintendo somewhere from $15-$20.
HBM2 is much harder to price, and there's both the component cost and the packaging cost to consider. From the packaging costs point of view, the use of InFO should bring costs down substantially over the use of a silicon interposer (InFO is used in, for example, the Apple A10, although that implementation is a little different). Then, obviously the use of only one HBM stack (versus four in all existing HBM implementations) would bring the component cost down by 75%. The issue is that we don't really know how expensive existing HBM implementations are, how much of that is component cost versus packaging cost, or precisely how big the savings are on the packaging side.
Just for fun, though, let's make a first-order approximation. To do so, let's assume that the savings of InFO over a silicon substrate are between 50% and 75% (this is roughly plausible, but more importantly it means we can ignore the cost breakdown between components and packaging). Let's also assume that AMD launch their new Vega GPUs at around the same time as Switch, to use them for comparison purposes. Judging by current rumours, it seems likely that AMD will have a "smaller" version of Vega with 8GB of HBM2 (two stacks) and a higher-end version with 16GB of HBM2 (four stacks).
The 8GB version seems to be competitive with the GTX 1070/1080, so let's say it launches at $500.
To give a benchmark of memory cost on AMD's ~$500 GPUs, we know that in late 2013, when PS4 launched, 8GB of GDDR5 cost about $88. AMD at the time were using 4GB of GDDR5 in their top-end GPUs, although they used a wider bus than PS4, so let's say it was costing them about $50 for memory for their ~$500 cards.
I don't think it's safe to assume that Vega will have the same memory cost as the R290/290X, as GDDR5 was a mature tech at the time, and HBM2 isn't now. Let's put an upper limit at double the cost though. So, at a maximum, we'd be looking at $100 for a two-stack HBM2 solution on a silicon substrate, and with a 50% saving with InFO, that would translate to a cost of $50 for Nintendo to use one HBM2 stack with InFO packaging.
For the lower limit, let's look at the "big" Vega with 16GB HBM2. Assuming it's going to be competing with Titan XP/1080Ti, it should be comfortably over $500 (how far over doesn't really matter too much). We'll assume, on a lower limit, that the cost of a four-stack HBM2 memory pool is $100 in this case. If we take a 75% cost reduction from moving from a silicon substrate to InFO, then that would put the lower-limit cost for a one-stack InFO HBM pool to $25 for Switch.
So, with some very rough approximations, we can put the cost of 4GB of HBM2 using InFO packaging at somewhere between $25 and $50. Compared to LPDDR4, this would be an increase of from $5 to $35. I should emphasise that these are very rough estimates, and I would personally feel that the higher end of that spectrum would be more likely, but it's hard to say with any degree of certainty.
