Hmm, annoying news to hear, but it was always one of the likelier scenarios in mind. Oh well, still doesn't aggravate me as much as certain other hardware stuff... (present day GPU market is still terrible for my specific wants and it's one of the moving parts delaying my next desktop build)
But at the end of the day, I've actually always felt some level of calm ever since the Nvidia hack. Knowing that Drake's being worked on and it being a matter of time helps me a lot. I end up not really caring
that much about the
when and focus more on having fun with thinking about the
what and
how, along with the underlying
why that's behind the details of the
what.
(...no, I don't think that I can fit in who and where in there)
For example, that GPU size of 12 SMs. That's a what detail. But then one wonders, why 12 exactly? Why not a smaller size which would be cheaper? We've then surmised that there must be some practical advantage to go with this size. Probably targeting perf/watt. But is it perf/watt in docked? Is it perf/watt in portable? Can it be optimal for perf/watt in both modes? How much power can you allocate for the GPU, which then presumably gets split up among 12 SMs? If it's only 3-4 watts in portable, what do you get from a quarter to a third of a watt per SM? Are we still above the voltage floor* at this point? Which then introduces the thought that maybe's there power gating going on, so you're not splitting up a few watts across
12 SMs, but instead a smaller number.
Another example/frequent topic is memory bandwidth. We can reasonably infer that the memory bus width is 128 bit. LPDDR5 is what's explicitly listed, IIRC. It's highly likely that 'LPDDR5' here is specifically the standard LPDDR5, but there's an outside shot that it's LPDDR5X. Regardless, options can be assumed. We can calculate 128 bit LPDDR5 (6400 MT/s for a total of 102.4 GB/s). We can calculate current 128 bit LPDDR5X (7500 MT/s for a total of 120 GB/s). We can calculate the peak by spec 128 bit LPDDR5X (8533 MT/s for a total of ~136.5 GB/s). The followup question is then, how does that bandwidth fit with the needs of CPU + GPU? That breaks up into, how much would the CPU need and how much would the GPU need? How much the CPU needs is a tough one to figure out; we can probably discuss that later. What about the GPU? What can be used as a starting point? As far as we're aware, architecturally, we're basically looking at Ampere. Can we then use the RTX 30 series cards for comparison? If yes, can we do something like, calculate (bandwidth) / (SM count * clocks) to get a feel for how Nvidia tried to balance things back in 2020? (or, bandwidth / flops as a shortcut). Would Nvidia have since tweaked how they'd balance things with experience from post-RTX 30 launch?
*quick refresh for the readers:
For chips, power consumption more or less scales quadratically with voltage (ie, voltage squared). As long as you can safely lower the voltage, your power consumption goes down faster than your performance, so your perf/watt ends up improving. The catch is, you can only lower voltage so far (you eventually lose stability). Where that floor is presumably varies with every chip. When you can't continue lowering voltage and can only further dial down current, your power consumption goes down
slower than your performance, so your perf/watt starts going back downward. Thraktor demonstrated this with his RTX 3070 in
this post.