LPDDR6 is actually the biggest single Q. There’s really no other question of that caliber.
indeed.
NPU upgrade yeah though I don’t know about 100 TOPs? We’ll see. N3E/N3P also a given I think.
Microsoft will double the NPU TOPS requirment for next generation Copilot PCs accoriding to rumours...
But LPDDR6? Big Q. Does it arrive on time, and do they want to rely exclusively on that? If not, then one issue is then they’d have to ship a memory controller compatible with LPDDR5 and LPDDR6. With LPDDR4 and LPDDR5 this was a PITA apparently.
And it will be a bigger PITA with LPDDR5/6.
5 years ago, the Snapdragon 865 was announced as the first Snapdragon SoC to support LPDDR5. It had a hybrid LPDDR controller that supported both LPDDR4X and LPDDR5.
I expect this year's Snapdragon 8 Gen 4 to top out at LPDDR5X-9600. Perhaps it will also support Samsung's LPDDR5X-10667. But I don't think we can expect a LPDDR5X/LPDDR6 hybrid controller. There is really next to no benefit in supporting LPDDR6.
1. If they drop LPDDR5X support entirely and replace it with LPDDR6, OEMs are going to revolt. I expect LPDDR6 will be extremely rare and expensive until 2025H2.
2. They could support both LPDDR5X/LPDDR6 in a hybrid controller. But implementing this is going to be difficult, due to the substantially different architecture of LPDDR6. And because LPDDR6 supply will be limited, few devices really make use of it.
Hence why I believe 8G4 coming in 2024Q4 will not support LPDDR6. I expect 8G5 in 2025Q4 will completely drop LPDDR5X support and fully adopt LPDDR6.
If they do LPDDR5-only, then it kind of hurts for future proofing and it’d hurt more than doing that in phones because a PC can really benefit from LPDDR6 assuming they’d also widen the bus which I suspect they would to 192B.
Yeah, they need the bandwidth to feed the huge NPUs.
Also, I expect to see mobile E Cores from phones thrown in as long as scheduling can make good use of them.
Are you talking about Custom Oryon derived E-cores or ARM Cortex A5xx? The former is likely, while the latter is never going to happen.
Wouldn’t work the same way as macOS with APIs but they could just migrate from E to P for increased user interaction and stuff with kernel scheduling. It’s too tempting not to if only for area gains and mild battery gains.
People always talk about P/E core scheduling and stuff when it comes to Intel, and AMD to some extent (becuase of their Zen C cores). But Qualcomm was the first to bring a hybrid CPU architecture to Windows PCs with their Snapdragon 8cx. Doesn't Windows-on-ARM have a good in-built P/E core scheduler?
We could probably see either 8 + 4 or a 12+4 setup for the big one, and I expect to then see a smaller die like Purwa doing 4 + 4 or 4+6 — could also use it in Android tablets or handhelds probably. (Or they do 8+4, who knows)
I don't think it makes sense to go beyond 8P+4E for consumer/prosumer devices for now. Speaking of which, it's time to reveal this:
Snapdragon X2 Speculation/Concept V7