Jump to content

October 18th Apple Event - Unleashed - Apple Silicon, MacBook Pro upgrades, HomePod mini, AirPods 3rd Generation

BondiBlue
Go to solution Solved by BondiBlue,

Summary

The Apple Unleashed event is over! Here are the new products that were announced:

  • AirPods
    • New AirPods 3rd Generation: MagSafe wireless charging, Adaptive EQ, and longer battery life
  • HomePod mini
    • In addition to Space Gray and White, HomePod mini now comes in Blue, Yellow, and Orange
  • Apple Music
    • New Voice Plan starts at $4.99/month, allows for Apple Music through Siri, including new custom playlist
  • And yes, new Macs and Apple Silicon
    • The M1 chip is now part of a lineup of three SoC designs, including the M1, M1 Pro, and M1 Max
    • The MacBook Pro has been redesigned, bringing back more ports, MagSafe charging, better battery life, and more
      • The 14" MacBook Pro starts at $1999, and the 16" starts at $2499. The 13" M1 MBP is now the base model
      • Support for up to 64GB of unified memory and 8TB of flash storage
      • M1 Pro and Max both have 10 CPU cores, and M1 Max can have up to 32 GPU cores
      • Fast charging has been added to the MacBook Pro, allowing for up to 50% charge in only 30 minutes

 

My thoughts

I'm really excited for the new MacBook Pros. I plan on upgrading to a new 16" MacBook Pro within the next couple months, and I can't wait. 

 

Sources

Apple Events

The Verge

3 hours ago, hishnash said:

Less that 5% of macPro customers are going need to buy additional add in card GPUs and given these are the same dies as the 14" and 16 macs

That's still based on a rumor about a GPU SoC configuration that may not come to be, such a large GPU in the SoC is going to require a lot of memory bandwidth that I just don't see Apple implementing due to cost and package size.

 

They could probably do it with 4 LPDDR5X chips in a quad channel configuration which would give around 180-200 GB/s memory bandwidth however that may still be too low. Apple does in general need less bandwidth for their applications and the targeted use cases however broadly speaking it's not going to be a GPU as powerful as a RTX 3090 because picking sub tests in that way is just a flawed and misleading as using AVX-512 to say Intel CPUs have that kind of broad performance, when they do not. Don't be Intel and use their "real world benchmark" schtick. Being good at a few things doesn't make you god at everything.

 

The M1 is a great SoC, it would still very happily sit in a larger MacBook Pro just fine and do everything ever needed of that device. So there are two paths Apple could go with an SoC revision for that sort of use case and target devices; More performance cores and TDP increase to match those extra cores so same clocks, increase in TDP to increase clocks and otherwise the exact same SoC.

 

The actual need for a much larger GPU basically isn't there, what you get now with the M1 is more than adequate and pushing for more than this in a single die SoC with limited memory bandwidth is going to be very strongly in diminishing returns. I can certainly see a double size GPU along with more performance cores as being quite likely but these 4x things are nothing but aspirational based solely on what people want to see and not at all grounded in technical parameters that would be required to actually make it possible.

 

M1 is a ~15W TDP SoC with device total power around that 25W-30W at the most high end. I just do not see Apple targeting more than 35W-40W SoC TDP in any MacBook Pro because it's a battery ran device and they care a lot about battery run times, which at least in my opinion is more important than 15% extra performance.

 

I do not see Apple using the same SoC in a Mac Pro as what is in any MacBook Pro either, not unless they do a mid sized Mac Pro.

Link to comment
Share on other sites

Link to post
Share on other sites

3 hours ago, hishnash said:

Yep but they sell much larger volumes, here we are talking about just the the macPro optional additional GPU market this is likely less than 5,000 units per year.  

I think you are massively underestimating how many Mac Pros Apple would sell and how many GPUs would be purchased if Apple did make a dedicated GPU.

Link to comment
Share on other sites

Link to post
Share on other sites

I think we’re mixing rumors about 2 different rumored Mac Pros

- the so called “half sized” Mac Pro

- the full size Mac Pro

 

(pic below is just a made up rendering for fun)

 

 

I could see the former using the rumored Jade-4C SoC (32 performance cores + 8 efficiency cores + up to 128-core GPU, that’s 4x times more GPU than the highest end 16” M1X laptop). I could se it having a big heatsink and a single big fan. And no dGPU whatsoever. And I could even see it being soft-launched tomorrow.

 

The latter is a late 2022 pipe dream based on the fact that collectively 

- we can’t accept the fact that the 2019 Mac Pro design was just a one-shot thing

- we can’t fathom Apple failing some professionals again and doing the 2013 MacPro mistake again after they had to revert that 

 

We’ll know by late 2022 what they will come up with (discrete pcie GPUs, etc.) to make a real successor to the 2019 Intel Mac Pro. But the other one, the “Mini-Me” Mac Pro, doesn’t necessarily needs dGPUs.

7A69344D-62E9-4ED2-8C6A-147A3AA00FD1.jpeg

Link to comment
Share on other sites

Link to post
Share on other sites

It’ll be interesting to see what they do about the memory bandwidth situation…I wanna see Apple go full bananas both on the bandwidth to the unified memory and the bandwidth to the flash storage..

 

I read LPDDR5X products are expected for h2-2022.

 

Meanwhile if they used LPDDR5 in the M1X it would already be an upgrade over the LPDDR4X used in the M1.

 

Could they also use HBM for unified memory or it would be incredibly pricey? Could this be why they will (supposedly) downgrade the max RAM of the 16” from 64GB to 32GB?

Link to comment
Share on other sites

Link to post
Share on other sites

32 minutes ago, saltycaramel said:

I could see the former using the rumored Jade-4C SoC (32 performance cores + 8 efficiency cores + up to 128-core GPU, that’s 4x times more GPU than the highest end 16” M1X laptop). I could se it having a big heatsink and a single big fan. And no dGPU whatsoever. And I could even see it being soft-launched tomorrow.

The only thing about a small Mac Pro that doesn't really make much sense is a higher end Mac Mini using whatever this new SoC is. I just also get the feeling Apple isn't going to push the boat out as far as many are hoping with this one. Either way dGPU doesn't make sense in either as you say, high end Mac Mini or a small Mac Pro.

 

Also I suggest you strongly keep in mind that the current Firestorm cores use 4W, so do the new Avalanche cores for that matter, so if you want to believe Apple is putting in 32 cores then that is 128W just in those cores alone, no memory power, no GPU power, no efficiency core power so there is a lot of reason why Apple would not put that many cores in an SoC for a device like that. Do you honestly think Apple intends to put a 150W-160W SoC in to a double(ish) sized Mac Mini? I do not.

 

Something like that could possibly work at TSMC 3nm but it would still be around 100W which is far more reasonable.

Link to comment
Share on other sites

Link to post
Share on other sites

The baby MacPro was described by (sometimes unreliable lately) leaker Jon Prosser as “3 to 4 Mac Minis stacked” sized.

 

Bloomberg’s Mark Gurman calls it “half sized”.

 

Both (second hand) definitions make it sound bigger than “doubleish a Mac Mini”.

 

A worst case estimate of the TDP of this giga-die SoC would be 4x the TDP of whatever we’ll see tomorrow in a thin 16” laptop. 

 

If the whole case is filled by a big heatsink with an high air flow design and a big fan (mimicking the 2019 Mac Pro), I wouldn’t put it past Apple to find a way to cool a 160W-200W monster die in a cube or half size case.

Link to comment
Share on other sites

Link to post
Share on other sites

15 minutes ago, saltycaramel said:

If the whole case is filled by a big heatsink with an high air flow design and a big fan (mimicking the 2019 Mac Pro), I wouldn’t put it past Apple to find a way to cool a 160W-200W monster die.

Apple could cool it no problem, I just don't think they'll do it is all. It's just quite an un-Apple thing heh. Basically even with how large it is it would still be kind of noisy. A faster/bigger M1 chip along with a couple of new MacBook Pro's, Mac Mini's plus a few other things seems like a really nice launch event, then next year November ish Apple could release their 3nm SoCs.

 

Apple actually has a decent range or products and they all need to go to Apple SoC's at some point and I envision that current products will stay on their current SoCs for a little while so Apple will be using successively better SoCs as they go up the product stacks. So with that I don't think any Mac Pro products will actually feature any TSMC 5nm at all and will be 3nm, iMac's not sure so wouldn't want to say until after the 18th. Don't see the need to make too any guesses when official information coming so soon. 

Link to comment
Share on other sites

Link to post
Share on other sites

58 minutes ago, saltycaramel said:

Could they also use HBM for unified memory or it would be incredibly pricey? Could this be why they will (supposedly) downgrade the max RAM of the 16” from 64GB to 32GB?

 

What if Apple pulls a Broadwell and adds some “eDRAM” (not really) solely dedicated to the integrated GPU?

 

The top 16” could have

32GB LPDDR4X (cheaper and power efficient)

8GB HBM2 (pricier and power hungry) dedicated solely to the GPU

 

The current top Intel 16” MBP has

64GB LPDDR4X

8GB HBM2 for the 5600M

 

That’s a total of 72GB.

 

Are we ready to believe Apple would downgrade it to just 32GB LPDDR4X of unified memory?

Link to comment
Share on other sites

Link to post
Share on other sites

16 minutes ago, saltycaramel said:

What if Apple pulls a Broadwell and adds some “eDRAM” (not really) solely dedicated to the integrated GPU?

It would actually be more useful for the CPU cores, HBM isn't great latency wise and CPUs like that more than they like bandwidth.

 

Biggest LPDDR4X chips are 12GB and LPDDR5 are 16GB (however not listed on Samsung's site they have press release saying they have them).

 

Current MacBook Pro 16" uses DDR4 not LPDDR4X btw, that's how they can do 64GB. They are off package on motherboard chips.

eF4gevhZhctZqlMD.medium

XZQEBkLiWWjMF6hP.medium

 

Of course we are all basically assuming Apple is sticking with on package memory for this SoC, while the more likely just something to think about. They don't actually have to and not doing so opens many more options because putting many memory chips on a single SoC package is rather costly and also bad to package yields.

 

16 minutes ago, saltycaramel said:

Are we ready to believe Apple would downgrade it to just 32GB LPDDR4X of unified memory?

They downgraded the MacBook Pro 13" so 32GB is still double that of any current M1 device. Intel MacBook Pro 13" supports 32GB.

Link to comment
Share on other sites

Link to post
Share on other sites

Bout the notch, it is really a question of SW.

 

I do use an iPhoneXR, the notch is 0 problem as pretty much everything is designed to either work around it or ignore the whole top area.

Company issued phone is a Samsung with a hole-punch and pretty much everything has UI elements made unusable by it if I am "holding it wrong".

 

-> Apple will just implement it in a way that does not disturb apps unaware of and gives some extra real estate for those that know about it.

Link to comment
Share on other sites

Link to post
Share on other sites

11 hours ago, hishnash said:

Yep but they sell much larger volumes, here we are talking about just the the macPro optional additional GPU market this is likely less than 5,000 units per year.  

Sure...

Apple will develop and produce a GPU made from defective SoCs for a market of less than 5,000 units per year.

Link to comment
Share on other sites

Link to post
Share on other sites

Last year there were some rumors about an Apple dGPU codenamed “Lifuka” probably meant for the 32” iMac Pro..

Link to comment
Share on other sites

Link to post
Share on other sites

Rumor: it may be possible to spec out the 14” exactly like the 16”, making it the first time ever that the smaller MBP has access to full GPU power

 

 

Just pick your screen size.

Kinda like the M1 iPad Pros. (well not actually this year because the 12.9” has miniLED and the 11” doesn’t, but next year they will have feature parity again)

Link to comment
Share on other sites

Link to post
Share on other sites

7 hours ago, HenrySalayne said:

Sure...

Apple will develop and produce a GPU made from defective SoCs for a market of less than 5,000 units per year.

No im saying they will re-use the SoC packages they are already making just with dies that have defective cpu cores and fully functional GPU cores. 
 

The alternative is to create a dedicated GPU die, that i think is very unlikely there is a reason apple used an off the self FPGA for the afterburner card they did not expect high enough volume to pay the tooling costs to make a dedicated die and packaging etc for it.

 

6 hours ago, saltycaramel said:

Last year there were some rumors about an Apple dGPU codenamed “Lifuka” probably meant for the 32” iMac Pro..

I don think apple will have any dedicated GPUs outside of optional additional units for the macPro. They have been very clear to use devs that we should expect apple apple silicon devices to have the shared memory address space functions that are only possible with a GPU that is on package and shares the memory controller with the CPU, NPU, etc of the SoC package. 

 

12 hours ago, leadeater said:

They downgraded the MacBook Pro 13" so 32GB is still double that of any current M1 device. Intel MacBook Pro 13" supports 32GB.

No the 13" that they replaced was the 2 port one that was not configurable to 32GB only the 4 port one was configurable to 32GB. 

12 hours ago, leadeater said:

Of course we are all basically assuming Apple is sticking with on package memory for this SoC, while the more likely just something to think about. They don't actually have to and not doing so opens many more options because putting many memory chips on a single SoC package is rather costly and also bad to package yields.

The issue is bandwidth, if they go with off package memory that will be hard to provide enough bandwidth without much higher power draw (GDDR like the consoles for example).  

12 hours ago, leadeater said:

It would actually be more useful for the CPU cores, HBM isn't great latency wise and CPUs like that more than they like bandwidth.

 

I expect this is something that can be configured to some amount when you do a large enough custom order, Intel will be deploying HBM in Xeons soon as a L4 level cache there is no point in a cache if its latency is higher than the next layer (aka DDR5).   The M1 has pretty high memory latency at around 100 ns this is about the same as current HBM2 specs apple int he M1 apple mitigate this by having a large L3 system cache (they call system cache as it shared by all co-prosseosr not only get cpu). 

12 hours ago, saltycaramel said:

32GB LPDDR4X (cheaper and power efficient)

8GB HBM2 (pricier and power hungry) dedicated solely to the GPU

Apple will not have any memory dedicated to the GPU as this completely breaks the api model they have been pushing us devs to move to. The key ability of Apples GPUs, NPU, Cpu etc being able to share address points between each other (without copy operations) just int he same way 2 cpu cores can share the same memory is key to apples success, its why macOS is so snappy on the M1, being able to jump between Cpu, matrix unit, NPU, video encoder decoder, and GPU without needing any memory copies.. being able to have all these read/(write if you are very careful) at the same time! is extremely powerful and really does help you use the GPU, NPU, etc in situations were it a dedicated memory system would not be worth it due to the copy overhead. 

 

Link to comment
Share on other sites

Link to post
Share on other sites

53 minutes ago, hishnash said:

No the 13" that they replaced was the 2 port one that was not configurable to 32GB only the 4 port one was configurable to 32GB. 

And what makes you think anything will change? Once they stop offering the Intel models it's very likely all you'll have is the M1 option as it is right now until the next refresh. You can either have the M1 with 16GB maximum or Intel with 32GB maximum right now. Really doesn't matter about there being a 2 port whatever with less configurable ram, that was just Apple creating a cheaper model and wasn't a capability difference with the CPU or motherboard within it.

 

If you wanted or needed 32GB then you could get it, you still technically can but only if you want an Intel MacBook Pro and I would argue few want that.

 

53 minutes ago, hishnash said:

The issue is bandwidth, if they go with off package memory that will be hard to provide enough bandwidth without much higher power draw (GDDR like the consoles for example).  

Hardly, the difference is very small. Also the biggest power difference is in the fact they are using Samsung's 4 layer 4 channel LPDDR4X memory modules, 2 of them, giving 128bit wide memory bus. That uses a lot less power than 16 DDR4 memory chips, that's where almost all the power difference is. That and LPDDR4X uses 0.6v operating voltage rather than 1.2v.

 

The higher bandwidth in the M1 is just from that it is using LPDDR4X, the Samsung packages give 34GB/s each where as DDR4-3200 is only 25.6GB/s. LPDDR4X is just simply a lot faster than DDR4 is.

 

The latency and bandwidth difference from having on package versus on the main board close by is rather minor.

 

GDDR is just a bad example to look at because it's not optimized for low power, has a higher operating voltage and has many modules as part of each board design. That's like saying LPDDR4X uses less power than DDR4, of course is does. One is literally optimized for Low Power (LP) and the other is not. The trade off there is smaller capacities.

 

53 minutes ago, hishnash said:

Intel will be deploying HBM in Xeons soon as a L4 level cache there is no point in a cache if its latency is higher than the next layer (aka DDR5).

For Intel yes there is. HBM on Xeons is to provide the required memory bandwidth for AVX-512.

 

53 minutes ago, hishnash said:

The M1 has pretty high memory latency at around 100 ns this is about the same as current HBM2 specs apple int he M1 apple mitigate this by having a large L3 system cache (they call system cache as it shared by all co-prosseosr not only get cpu). 

You need to compared First Word Latency, there is a massive difference.  I can't remember what is is exactly for DDR but it's around the 10ns whereas with HBM it's 100ns. DDR has difference latencies for different things where HBM is rather static at that 100ns.

Link to comment
Share on other sites

Link to post
Share on other sites

36 minutes ago, hishnash said:

No im saying they will re-use the SoC packages they are already making just with dies that have defective cpu cores and fully functional GPU cores.

Ok. So they need to develop a way of communicating to the GPU directly without the CPU in between (not to mention memory access), they need to design a completely new PCB, a new cooling solution for 4 individual dice and a way to add inter-die communication and - if this wasn't enough - they have to verify and test everything and they need drivers. For less than 25,000 units sold over a period of 5 years?

Generally speaking, the idea of a SoC is that all components are interconnected directly and tightly and have shared access to some ressources like the memory controller. Untangling the GPU to work without the CPU is almost impossible. If you design the entire chip to be easily separable, you can just make two individual dice (like Ryzen APUs) and increase your yield and versatility without any crazy Frankenstein creations. The whole idea really far-fetched.

Link to comment
Share on other sites

Link to post
Share on other sites

1 hour ago, leadeater said:

And what makes you think anything will change? Once they stop offering the Intel models it's very likely all you'll have is the M1 option as it is right now until the next refresh.

The 13" 4 port MBP will be replaced with the 14" macBookPro (that will not have an M1). 

The 2 port MBP should never have been called `pro` as it was always using the ultra low power intel cpus that gave it very lacklustre perfomance it should have just been macBook.
 

1 hour ago, leadeater said:

For Intel yes there is. HBM on Xeons is to provide the required memory bandwidth for AVX-512.

 

And apple need it for GPU, NPU, video encoding/decoding and of course multiple cpu cores under heavy load, remember with apples SoCs is not just the Cpu that is pulling data but all the other co-prosesors many of which can really such a LOT through (video encoder/decoder, matrix co-prososrs, ISP, and of course GPU) and many tasks that might put load on one of these will also be putting load on others, if your doing video decoding your likely also streaming that data directly into the GPU and NPU and maybe ISP to do a range of effects with the result being read by the encoder. The total bandwidth needs of a Pro level apple SOC could be just as much as a AVX heavy Xeon workload. 
 

1 hour ago, leadeater said:

DDR has difference latencies for different things where HBM is rather static at that 100ns.

Yep but M1s latency to main memory (not to system cache) is currently at 100ns that is why I think they would be ok moving to HBM since this would not be an increase in latency they clearly can already handle this high latency (likely by having such a long out of order instruction buffer and large on die cache).  

 

1 hour ago, HenrySalayne said:

So they need to develop a way of communicating to the GPU directly without the CPU in between

Yes apple already has this in fact compute shaders on the GPU can schedule draw calls and do everything you can do from the GPU this has been a key feature of apple gpus for a long time.  Also Core to core communication in an apple GPU happens through system memory/cache layers so having GPU cores spread over multiple dies is not different to having cpu cores spread over multiple dies. The GPU cores (like all the other co-prososrs) on each die talk to the memory controller (the same as the cpu cores) this is not different from having cpu cores spread across multiple dies.  Fundamentally the GPU Cores are not considered any differnt from any other compute unit on apples SoC, be that CPU cores, NPU, matrix-co-prososr, video encoders, decoders, ISP etc all talk to the same memory controller this is how they are able to safely share points to memory without needing to have copies.

having the Gpu cores spread out across multiple dies helps a lot with power draw and thermals since you are unlikely to max out all the compute units of any given SoC. (very very few tasks max the CPU, GPU, NPU, MXU, Video encoders etc all that the same time).

 

1 hour ago, HenrySalayne said:

if this wasn't enough - they have to verify and test everything and they need drivers. For less than 25,000 units sold over a period of 5 years?

Apple sell a lot of high end iMacs likely in the 500,000+ a year these will likely use the same complex but with some dud-placeholder dies (like some thread ripper and ryzen systems).  And i expect the macPro will will many more than 25,000 units per year more like 200,000. And remember this follows from them using the same die as they will be using int eh 14" and 16" MBPs so that is a die that they will be silting many millions.




 

Link to comment
Share on other sites

Link to post
Share on other sites

44 minutes ago, hishnash said:

Yep but M1s latency to main memory (not to system cache) is currently at 100ns

No it's around 8ns for First Word like I said. Those latency measurement tests aren't actually "correct".

 

https://blog.logicalincrements.com/2018/11/choosing-ram-cas-true-latency/

https://www.overclock.net/threads/question-when-is-speed-latency.1603630/

 

DDR doesn't operate the same as HBM. DDR starts supplying data to the CPU much sooner than HBM does.

 

44 minutes ago, hishnash said:

And apple need it for GPU, NPU, video encoding/decoding and of course multiple cpu cores under heavy load, remember with apples SoCs is not just the Cpu that is pulling data but all the other co-prosesors many of which can really such a LOT through (video encoder/decoder, matrix co-prososrs, ISP, and of course GPU) and many tasks that might put load on one of these will also be putting load on others, if your doing video decoding your likely also streaming that data directly into the GPU and NPU and maybe ISP to do a range of effects with the result being read by the encoder. The total bandwidth needs of a Pro level apple SOC could be just as much as a AVX heavy Xeon workload. 

And using HBM would cripple CPU performance in the name of slightly better GPU performance. Does that sound like a good idea?

 

And while sure a much bigger SoC for a Mac Pro would need just as much that's a long way off and 6-8 channel DDR5 would be just fine for that. Apple could also go with on package HBM like Intel and Xeons as well as DDR5 and have that all properly handled in the memory controller, microcode and compilers.

Link to comment
Share on other sites

Link to post
Share on other sites

4 hours ago, leadeater said:

No it's around 8ns for First Word like I said. Those latency measurement tests aren't actually "correct".

 

No not according to https://www.anandtech.com/show/16252/mac-mini-apple-m1-tested random reads are in the 93ns+  the system cache is around the  size and there are not instructions that let to explicitly skip the system cache so valid memory latency tests need to be at larger than 32MB (the fast the first word is amassable after 8ns just means that work was in L2 or the SLC.  Any other form (non random) of read is very difficult to benchmark since the cpu will be massively prefetching with its branch prediction (that you can't turn off).  The consensus in the community is the the M1 memroy latency is rather high (this might well be imposed by the shared memory controller that all the co-prososrs use).  

 

 

4 hours ago, leadeater said:

And using HBM would cripple CPU performance in the name of slightly better GPU performance. Does that sound like a good idea?

 

No as i pointed out it would likely not affect memory latency compared to the M1, all benchmarks of the M1 memory latency were you are reading data larger than what can fit in the cache layers returns in the 93ns + range.  Why apple are not running the memory system at a lower latency is a valid question for sure but given they have very good perfomance with this latency i don't see how moving to HBM would have a big impact.

If apple do a 32Core GPU (4x the M1) on the die (as is strongly rumoured) they will need to 4x the bandwidth, they could do that with 4 LPDDR5x dies or 8! LPDDR4x, and your correct that 8 LPDDR4x would be very costly to have on package and having off package would also not line up with the semantics that leaked for the upcoming 16" (were the memory was on package) The package size from these schematics would also rule out 8 LPDDR4x on package, would even be very tight to get 4 LPDDR5x on package given typically packages will include some caps for power soothing etc.  2 HBM2e (16GB stacks) down clocked or 1/2 the bandwidth inteface (to increase yield and reduce power) would provide more than enough bandwidth for such a solution and possibly be cheaper than spending LOTs of 5nm die area on such a wide DDR controller (since the grunt of the controller for HBM is on the bottom for the stack the proportion of die area in the compute die used is much less, for the same bandwidth).   Remember the order volumes apple will be putting through for this will be massive, this one SoC package will likely be the most sold single chip package in laptops int he coming year. (yes it won't by the most laptops but most other laptops solution from other brands will be a mixture of current generation, last generation and the generation before chips from intel and some AMD).  
 

5 hours ago, leadeater said:

Apple could also go with on package HBM like Intel and Xeons as well as DDR5 and have that all properly handled in the memory controller, microcode and compilers.

For the macPro i do expect them to have a hybrid solution (ether DDR5 off package) or DDR over PCIe (so that they do not need to build a large DDR controller on die and just re-use the PCIe lanes) (yes that the cost of latancy) thinks swap but without the issues of ssd write ware. It would be quite an apple solution, increase the  number of PCIe slots and sell (custom apple only provided) PCIe/MPX memory expansion cards that effectively add a form of swap (that is much faster than SSD backed swap) apple could even use Optane like solutions if they wanted to provide extremely high capacity.  





 

Link to comment
Share on other sites

Link to post
Share on other sites

1 hour ago, hishnash said:

No as i pointed out it would likely not affect memory latency compared to the M1, all benchmarks of the M1 memory latency were you are reading data larger than what can fit in the cache layers returns in the 93ns + range

Did you actually read what I wrote? Yes full access time is around that high, so it is on Intel and AMD (although typically around 60ns to 80ns). However FIRST WORD (this is a literal thing look it up) latency is around that 10ns mark, this is how soon the CPU starts receiving data and can start processing it. The time between when a CPU can start processing data is a ten fold difference between DDR and HBM. This is why it's only used on latency insensitive applications.

 

You can quote that 100ns figure all day long if you like, it's not correct. That's not how DDR works and the latency tests you are looking at, like those of AIDA64 can only test full/complete latency time not First Word.

 

Apples memory latency isn't any higher than anyone else's, not applicably, but if they were to move to HBM then yea it would be a lot higher where it counts.

 

lat-5950.png

 

Zen 3 using the same test is 78ns, this will of course be different on a system by system basis depending on ram timings and frequency.

 

1 hour ago, hishnash said:

If apple do a 32Core GPU (4x the M1) on the die (as is strongly rumoured) they will need to 4x the bandwidth, they could do that with 4 LPDDR5x dies or 8! LPDDR4x, and your correct that 8 LPDDR4x would be very costly to have on package and having off package would also not line up with the semantics that leaked for the upcoming 16" (were the memory was on package) 

Yes and that is your assumption that Apple will be using the same die in a Mac Pro, I however am not making that assumption. Two things we know; MacBook Pro is not a Mac Pro, Mac Pro will not using the same SoC as the MacBook Pro. Both of these statements are true. Now rumors and assumptions are being made that Apple with go with some kind of multi chip/die SoC, may or may not be true. Now if we go with yes this will be the case using the same die as the MacBook Pro is still yet another assumption.

 

The Mac Pro is an entirely different platform with different needs and design goals. Whatever they do on the MacBook Pro may not be entirely directly usable in the Mac Pro. Even if they use the same die this does not in any way rule out expandable DIMM memory or off package memory and for all we know the memory controller already support both LPDDDR4X and DDR4 (this is the case with Intel and AMD).

 

HBM has been around for a long time, there is only a single CPU that has used it (and only it) which is the Fujitsu A64FX and that is a purposed design CPU for vector math and it optimized for throughput and the workloads it runs is specifically suited for all of this. It's not a CPU you would want to use in a "desktop" type system.

 

Edit

image.png.e611b31ce6268fef3b4556d004d21575.png

https://www.utmel.com/blog/categories/memory chip/will-hbm-replace-ddr-and-become-computer-memory

Link to comment
Share on other sites

Link to post
Share on other sites

 

That’s an interesting use of the world “mere” since he’s talking about a desktop 3080.

 

That said, I think Apple’s main focus will be looking really good against their current top AMD offering (5600M 8GB HBM2) on MBPs. And, for the first time, doing it on the smaller 14” model as well. 

Link to comment
Share on other sites

Link to post
Share on other sites

6 minutes ago, saltycaramel said:

 

That’s an interesting use of the world “mere” since he’s talking about a desktop 3080.

 

That said, I think Apple’s main focus will be looking really good against their current top AMD offering (5600M 8GB HBM2) on MBPs. And, for the first time, doing it on the smaller 14” model as well. 

If these things can pull off 70% of a desktop 3080, they would instantly be the best bang for your buck notebooks on the market for both CPU/GPU performance.


The tears this would generate would be entertaining at least… Shame there wouldn’t be a game catalog to go along with it.

MacBook Pro 16 i9-9980HK - Radeon Pro 5500m 8GB - 32GB DDR4 - 2TB NVME

iPhone 12 Mini / Sony WH-1000XM4 / Bose Companion 20

Link to comment
Share on other sites

Link to post
Share on other sites

In 15 minutes Linus will have something extremely important to do and won’t be able to watch. (heard in every post-Apple-event WAN: “actually I didn’t watch the event..”)

Link to comment
Share on other sites

Link to post
Share on other sites

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!

Register a new account

Sign in

Already have an account? Sign in here.

Sign In Now


×