4 ms·
Apple announced the M3 on October 30 last year, roughly 6 months ago. Have they ever announced successive chip designs in such a short time frame? Unless Mark
by tromp 3y ago
Apple announced the M3 on October 30 last year, roughly 6 months ago. Have they ever announced successive chip designs in such a short time frame?
Unless Mark Gurman has a record of making reliable predictions about Apple, I wouldn't give this much credence.
- aurareturn 3y agoIt's not completely insane. 1. N3B TSMC node used by M3 is reportedly low yielding. Apple likely wants to move away from it asap and move to N3E, which is the higher yielding node. 2. Rumors are that M3 Ultra is not planned due to low yields. M3 Max die shot does not have UltraFusion silicon. This supports the first theory. 3. Apple base the M chips on the A series cores. A18 Pro has already finished in design because Apple always starts making A series chips in Spring before a Fall debut. Therefore, A18 Pro is already in production using N3E node. Given enough early planning, it isn't surprising that M4 can be released early. 4. Reportedly, Apple plans to focus on GenAI at WWDC. They're also behind in GenAI. In order to catch up, they need to be aggressive. M4 is likely to have a big improvement in inference speed via the NPU and/or GPU.
- CSSer 3y agoThink that whole major, unpatchable security flaw thing could also play a part? I could see them wanting to put distance between themselves on that fast too.
- aurareturn 3y agoI don't think so. I think M4 was planned years in advance to debut around this time. You can't change this stuff in a few months.
- saagarjha 3y agoI don’t recall any that affected silicon?
- kevlened 3y agoKey extraction from the HSM. It's unclear if this was ultimately patched in the kernel, but it appears patchable. [1] https://gofetch.fail/ https://gofetch.fail/ [2] https://social.treehouse.systems/@marcan/112238385679496096 https://social.treehouse.systems/@marcan/112238385679496096
- saagarjha 3y agoDefinitely not the HSM, this only affects the AP
- 1123581321 3y agoM3s have a flag to disable the caching vulnerable to GoFetch on p-cores. M1s and M2s are the ones that have no mitigation besides running workloads on slower e-cores.
- lwkl 3y agoThat‘s not true DMP can be disabled on M2 chips and probably also on the M1 chips. This was found by Hector Martin (@marcan) and will probably supported in Asahi Linux.
- 1123581321 3y agoThat's great news. Direct link for anyone else not up to date: https://social.treehouse.systems/@marcan/112238385679496096 https://social.treehouse.systems/@marcan/112238385679496096
- listic 3y ago> They're also behind in GenAI Where can I read up on that and what's state of the art in consumer GenAI chips? Also thanks, all great points!
- TrainedMonkey 3y agoI think 4 is underrated. From what I've seen Apple is making a play for on device inference. Gains on better screen, thinner phone, better camera selling points are pretty incremental now. But with on-device inference they now have access to: doubled the number, no cloud we care about privacy, on device translation capability, etc.
- WoodenChair 3y ago> Unless Mark Gurman has a record of making reliable predictions about Apple, I wouldn't give this much credence. He does have a great record.