When discussing the comparisons between Nvidia and Apple in
@Jimmyjames's other
thread, it occurred to me how similar Apple's GPU design, at least in terms of TFLOPs/Watt, is to Nvidia's MaxQ line for laptops. In fact, had the M3 Ultra existed it would've almost perfectly lined up with the
4090 MaxQ:
| 4090 MaxQ | hypothetical M3 Ultra |
| execution units | 9728 | 10240 |
| clock speed | 1.455 | 1.398 |
| TFLOPs | 28.31 | 28.63 |
| Watts | 80 | ~80 |
Now obviously there are key differences: the 4090 MaxQ has only 16GB of GDDR6 RAM and a bandwidth of 576 GB/s while the M3 Ultra in the above config would have had a minimum of 96 GB of RAM and a bandwidth of
800 700 GB/s, the M3 Ultra has a TBDR GPU, the M3 Ultra would likely have suffered a performance hit from the interconnect, and of course Nvidia designs the 4090 MaxQ for laptops while Apple designs the Ultra for mid-sized desktops ... and its tower. Commensurate with this difference in design philosophy is a difference in how each user base expects their machines to perform. Apple has prized quiet and cool operations and sells the benefits of that to its users whereas PC laptop makers
will literally hide that they are using MaxQ designs over the more power hungry "mobile" line. Basically a
4070 MaxQ performs equivalently to a
4060 mobile at less than 1/3 the power, but laptop makers are worried that users would be turned off because those users want the performance of the 4070 mobile. So they hide that it's a MaxQ despite the fact that the MaxQ is the more sane design for a laptop. Even more extreme, as
@mr_roboto found to his amusement in a MR thread where a user posted this as a good thing, some PC laptop makers will even brag about how many insane watts they let their GPUs burn ... again in their laptops.
All that aside, I just thought it was interesting how Nvidia has a mobile line of graphics cards, the MaxQs, that are actually quite similarly architected to Apple's (I don't mean in terms of TBDR, just the design philosophy of good performance at low watts through width and low clocks). The 4070 MaxQ is similar to a cut down M3 Max (even lower clocks!) - although the
4050 MaxQ has a little bigger and more power hungry design than the M3 Pro. It seems 35W is as low as Nvidia wanted to make its dGPUs go. I know Apple doesn't prioritize desktops, it's not their bread and butter, but I do wonder if we'll see a desktop oriented SOC from Apple and what that might look like - even just a speed boost like the 4090 mobile compared to the MaxQ nets a 16% increase in performance, but at the cost of 50% more power - worth it for a desktop certainly if more than a bit dubious at 120W for a laptop. Then again, Apple may just build in even more cores, though that is the more expensive option (higher clocks require better bins but adding more cores costs die area, almost certainly the more expensive of the two).
I'm a fan of threading the needle in a monolithic Ultra that relative to the 2x Max Ultra cuts down on cores, boosts clocks (at least on the GPU), and then an Extreme is made by glueing those two together - I know others, e.g.
@NotEntirelyConfused, have likewise created a similar "what if" cut down monolithic Ultras. Were I in charge of Apple, and it is probably a good thing I am not, that is the approach I would take to competing in the desktop space if that became a priority (with the AI boom it could be).
or what is michy-kun? I forget. They are like in unholy duality for me. Great training for my anger management issues though.
