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

[suicidalfranco]

3 hours ago, hishnash said:

In perfomance it will be easy, the M1 runs circles around the framework laptop on every spec point (even price). 

Even if its the fastest, would still go framework any day of the week

[Dracarris]

30 minutes ago, Video Beagle said:

This is not a presentation for tech geeks. It's not an engineering presentation

Nothing of what I suggested would qualify it as such. Not even remotely.

[Lightwreather]

So, some information as to the configuration has come out:

https://youtu.be/SNINTJwZzq4?t=408

[hishnash]

1 hour ago, leadeater said:

If you can show me a better benchmark of the M1 storage in small I/O sizes I'll happily compare to better tests on the Windows/Linux side with IOmeter or equivalent. 

 

I would suggest looking into latency numbers as for day to day operations (using the OS) this will make the most difference between drivers. This tool has full cross platform support https://arstechnica.com/gadgets/2020/02/how-fast-are-your-disks-find-out-the-open-source-way-with-fio/ 

[leadeater]

11 minutes ago, hishnash said:

I would suggest looking into latency numbers as for day to day operations (using the OS) this will make the most difference between drivers. This tool has full cross platform support https://arstechnica.com/gadgets/2020/02/how-fast-are-your-disks-find-out-the-open-source-way-with-fio/ 

Yes I know how to use fio and IOmeter, what I do not have is an M1 device or a review of an M1 with anything better than "CrystalDiskMark" (it's MacOS equiv I linked) or Black Magic Speed test.

 

Even with the very basic one I linked if what you said were true the last row, 4K Random QD1, would have been MUCH faster on the M1 than the 980 Pro but it's not, in fact in terms of write way slower.

6618.652 IOPs vs 51596.680 IOPs

 

or

 

0.15ms vs 0.019ms (this isn't actually correct for anything other than inflight I/Os not burst)

 

Storage is literally one of my direct expert subject areas, just last week I spent days testing different RAID configurations (stripe sizes, array groupings, cache ratios) for Netapp StroageGRID that were to run on HPE Apollo 4200 Gen10 LFF servers with 24 NL-SAS 12TB disks. I did this because I disagreed with the vendor provided advice so had to prove I was in fact correct, I was. 

[leadeater]

1 hour ago, Video Beagle said:

This is not a presentation for tech geeks. It's not an engineering presentation to board partners like AMD does. This is a big shiny marketing show aimed at Wallstreet and the mainstream

That's not an excuse to no label a graph, and no it won't make it harder to read. Any other company would be shamed for not doing it and I see no reason Apple should be exempt, they are because eh Apple but well, I can't change it. Doesn't make Apple either any less incompetent or deceitful, the only legitimate reasons to not correctly label and scale graphs.

 

But don't worry I'm not bashing on Apple alone, Nvidia is a repeat offender to, Intel to a lesser degree than Nvidia, AMD pre-Ryzen 1000 too.

 

I just wish Apple would stop pretending they are graphs and just get rid of the graph lines, at least then you can't complain they haven't labeled the graph lol

 

However I'd take what Apple does over this abomintation

Good luck knowing the market share percentages.

[Commodus]

5 hours ago, leadeater said:

I still have a feeling Apple isn't going to release their biggest planned chips just yet, like the big mac Pro replacement ones. So there's still a good chance there will be even higher end stuff coming later as well.

Guaranteed... the pinnacle will be the chips destined for the Mac Pro (and possibly high-end iMac). Although I wonder how Apple is going to tackle high-end GPU needs in those spaces — no on-die wizardry is going to mimic four high-end dedicated GPUs.

[LAwLz]

So, what expectations do people have?

I'm expecting the M1X to outperform the Ryzen 5800X at both single core and multi-core workloads, and use about half the power at load.

I am very excited for the GPU as well. The GPU in the M1 is fantastic for the size and power consumption.

[Commodus]

1 hour ago, LAwLz said:

So, what expectations do people have?

I'm expecting the M1X to outperform the Ryzen 5800X at both single core and multi-core workloads, and use about half the power at load.

I am very excited for the GPU as well. The GPU in the M1 is fantastic for the size and power consumption.

I'll be conservative and say that the M1X should perform as well as or better than comparable Core/Ryzen chips in terms of raw CPU performance, but that GPU performance might not live up to the nicer dGPU options on the market. The 14-inch model may be the sweet spot for performance and value.

[BuckGup]

The real hidden king is the mac mini. It's the only M1X based desktop computer on the market.

[Ravendarat]

Ive been waiting forever to replace my 2009 Macbook pro, which by the way I still daily drive to this day, and if the new 14 inch is similar in price to the current 13 than Ill be ordering a new Macbook Pro the first day its available. That of course is provided the rumors are somewhat true in regards to the new options that should be on them

[saltycaramel]

 

lol that pic

[Video Beagle]

2 hours ago, Ravendarat said:

Ive been waiting forever to replace my 2009 Macbook pro, which by the way I still daily drive to this day,

 

2007 right here

 

[Dracarris]

1 hour ago, Video Beagle said:

2007 right here

You guys are hardcore. I succumbed in 2018 to replace my 2010 15" aluminium chunker with a 13" butterfly that feels like it is 1/4th of the size and weight, has 4x the pixels and 3x the trackpad real estate. And don't mention the speed when you come from a Core 2 Duo..

 

Just to be clear, the 2010 15" still works flawlessly.

[StDragon]

On 10/13/2021 at 7:26 AM, LAwLz said:

So, what expectations do people have?

I'm expecting the M1X to outperform the Ryzen 5800X at both single core and multi-core workloads, and use about half the power at load.

I am very excited for the GPU as well. The GPU in the M1 is fantastic for the size and power consumption.

I really want to see a comparison between the M1X and i7-12700T (8P / 4E cores @ 35W)

[hishnash]

21 hours ago, Commodus said:

Guaranteed... the pinnacle will be the chips destined for the Mac Pro (and possibly high-end iMac). Although I wonder how Apple is going to tackle high-end GPU needs in those spaces — no on-die wizardry is going to mimic four high-end dedicated GPUs.

I expect they will have a GPU on the SoC Package, likely using the same dies a step 16" MBP but 4 of them. That would lead to a 128 Core GPU for the SoC. (this would put it above the 3090 in performance by a good margin).

Then they will also mount these SoCs to PCIe/MPX cards and offer them as add in GPUs, this will let them use up the dies that have defective cpu cores and mean they do not need to bother making a dedicated GPU die for the extremely low volume optional add in market of the mac pro.  These will be rather powerful as add in compute cards since they will not only have 128 GPU cores each but also have a load of NPU and other video encoders/decodres but most importantly could also have quite a lot of local memory etc.

 

[saltycaramel]

 

“Big”

the jump-to-lightspeed logo

 

What is Apple trying to say?

It’s too subtle, I don’t get it. 

[hishnash]

3 hours ago, saltycaramel said:

What is Apple trying to say?

 

macPro   

[leadeater]

17 hours ago, hishnash said:

I expect they will have a GPU on the SoC Package, likely using the same dies a step 16" MBP but 4 of them. That would lead to a 128 Core GPU for the SoC. (this would put it above the 3090 in performance by a good margin).

Zero way you are going to get 3090/A40 performance out of a GPU with dual channel LPDDR4X bandwidth. Blood from stone on that one I'm afraid, or interdimensional magic.

 

68GB/s x 4 = 272 GB/s vs 1008 GB/s (696 GB/s for the A40)

 

And that's without the additional complications of multiple chips.

[Laborant]

Regarding CPU performance, I personally expect them to use 4 low-performance cores and 8 high-performance cores. Why four low-performance cores? For the small background tasks, your OS (also Linux) constantly runs in the background. A high-end machine has approximately the same demand for low-performance cores as a standard MacBook Air has. You don't want to keep your high-performance core busy just because your OS issues another WiFi-Printer-search or some "Finder"-Indexing or a refresh on the Weather-Widget. That allows a high-performance core to run with fewer context switches.

 

GPU-wise, I see one power target and two possibilities for how to achieve that. The Graphics need to be faster than in their best-performing MacBook Pro - better than the Radeon 5600M. To achieve that, Apple either continues with AMD and simply attaches the GPU via PCIe or spec-bump their own GPU solution.

 

I guess their upper tier will still be a multi-chip solution, most probably with PCIe 4.0 x16 and fast memory for the GPU. The question is AMD or Apple Silicon. 
Apple really stirs up the competition now. And yes, that's good for us, the consumers.

 

 

[saltycaramel]

15 minutes ago, Laborant said:

Regarding CPU performance, I personally expect them to use 4 low-performance cores and 8 high-performance cores.

 

 

While your reasoning is right, rumors point to M1X being 8p+2e tho.

 

A “bad” thing I could see about your hypothetical 8p+4e M1X would be that then the M1X_dual (codename Jade 2C) and the M1X_quad (codename Jade 4C) would have way too many efficiency cores…16 efficiency cores in the Jade 4C (the ~40-core beast for higher end desktops) sounds a bit of a waste of wafer real estate..

[hishnash]

23 hours ago, leadeater said:

Zero way you are going to get 3090/A40 performance out of a GPU with dual channel LPDDR4X bandwidth. Blood from stone on that one I'm afraid, or interdimensional magic.

 

You assuming that a single die (as it will be in the 16") will only have 1/4 of a channel of LPDDR4x. I expect the 16" MBP will be 4 channels of LPDDR5x or 2 stacks of down clocked HBM2e. So a 4 die arrangement will have either 16 Channels of LPDDR5x or 8 HBM2e (maybe even not down clocked as they are not in a laptop).

 

23 hours ago, leadeater said:

68GB/s x 4 = 272 GB/s vs 1008 GB/s (696 GB/s for the A40)

 

The chip in the 16" will not use the same memory subsystem as the M1 (that has 68GB/s)!  it will be at least 4x faster but i expect more than that since a single M1 performance core can use up 60GB/s and in a pro machine providing enough bandwidth for all cores to run maxed out will really push the needle. 

 

[leadeater]

51 minutes ago, hishnash said:

You assuming that a single die (as it will be in the 16") will only have 1/4 of a channel of LPDDR4x

I assumed 4 dies using the current quad channel used which is very unlikely to change, remember LPDDR4X is 2 channels per module (2x 16 bit). Apple isn't going to quad channel LPDDR5X on these devices, they don't need it and it's too costly, in size as well. 4 dies = 16 memory channels in total with any single die only having access to 4 channels as is right now. Interconnect bandwidth just isn't there to have broad cross die communication like that for GPU, just have a look at first generation EPYC and Threadripper for why and that's for lower bandwidth workloads with sub-NUMA clustering to prevent just this type of cross die communication.

 

They may or may not use LPDDR5X, if they do it'll be only 34% increase in bandwidth as LPDDR5X is a single 16 bit channel as opposed to two 16 bit channels in LPDDR4X. Hence my prediction due to minor bandwidth increase, significant cost increase, potential supply issues Apple will be using LPDDR4X so they do not have to create a new memory controller and will save that for another time with a wider SoC architecture update.

 

The Apple A15 released mid September this year continued to use LPDDR4X and did not move to LPDDR5X.

 

51 minutes ago, hishnash said:

The chip in the 16" will not use the same memory subsystem as the M1 (that has 68GB/s)!  it will be at least 4x faster but i expect more than that since a single M1 performance core can use up 60GB/s and in a pro machine providing enough bandwidth for all cores to run maxed out will really push the needle. 

Unlikely Apple with go with quad channel and LPDDR5X, LPDDR5X itself still is not a given right now. Right now is the very earliest feasible time that LPDDR5X could be used but that does not mean Apple has actually committed to trying to use it.

 

Also HBM on a general purpose CPU, nope that would be horrible.

 

Also your point was Apple could reuse these chips in the MacBook Pros in some dedicated GPU, by way of using 4 dies which means they WILL be using the same memory configuration as in the laptops.

 

If Apple is going to create a dedicated GPU they'll make a dedicated GPU, they won't frankenstein CPU dies in to some weird GPU thing that'll cost more than it needs to and have massive drawbacks in not just making an actual GPU die. They were happy enough to make an accelerator card for the Mac Pro so why would they not make a GPU.

 

Edit:

Add some clarifications around LPDDR4X vs LPDDR5X, as they are quite different.

[hishnash]

5 minutes ago, leadeater said:

I assumed 4 dies using the current dual channel used which is very unlikely to change. Apple isn't going to quad channel on these devices,

Given a single Cpu core can fully strut the dual channel on the M1 they can't continue to just go with that, and then consider that they are going from an 8Core GPU to a 32Core gpu even for the single die they need to support more channels (if they are staying with LPDDR). 

 

8 minutes ago, leadeater said:

Interconnect bandwidth just isn't there to have broad cross die communication like that for GPU, just have a look at first generation EPYC and Threadripper for why and that's for lower bandwidth workloads with sub-NUMA clustering to prevent just this type of cross die communication.

So one thing to remember with apples GPUs the 8 core GPU in the M1 really should not be considered a single GPU, it behaves much more like 8 GPUs running in crossfire. Apple strongly! discourage us devs from doing any core to core communication (since that runs through system memory) there is no direct core to core communication change (at least not in the public apis).  

 

11 minutes ago, leadeater said:

If Apple is going to create a dedicated GPU they'll make a dedicated GPU, they won't frankenstein CPU dies in to some weird GPU thing that'll cost more than it needs to and have massive drawbacks in not just making an actual GPU die. They were happy enough to make an accelerator card for the Mac Pro so why would they not make a GPU.

You misunderstand the rumors are that the macPro SoC will be 4 of these dies on one package, that will include a GPU. What I suggesting is for users that want more GPU power what apple will do is use these SoC packages that have defective Cpu cores. So this would not be them making this 4 die solution for the GPU but rather them re-using the SoC that they already made for the system.  Since the SoC will already have a very powerful GPU option very few people will need to buy additional Gpu power (as things like video encoding/decoding etc is all handled by other co-prososors within the package leaving a lot fo GPU bandwidth for the typical macPro user). Apple could go an make a dedicated die for the macPro optional additional Gpu card but this would have a missive cost. There is a reason apple used an off the shelf FPGA for the Afterburner card rather than use a custom chip as the tooling costs would be astronomical given the low sales volume. 

[leadeater]

15 minutes ago, hishnash said:

Given a single Cpu core can fully strut the dual channel on the M1 they can't continue to just go with that, and then consider that they are going from an 8Core GPU to a 32Core gpu even for the single die they need to support more channels (if they are staying with LPDDR). 

That's the case for basically every modern CPU and core arch, it's not unique to Apple and it doesn't actually prevent general performance increases when adding cores. Also I'll leave "rumors" about GPU increase alone because until official specs it's just as likely it's 8 GPU cores as it is 32 cores, both are just guesses with the only real difference is one is based off something actually used now.

 

15 minutes ago, hishnash said:

Apple strongly! discourage us devs from doing any core to core communication

Not much different to Nvidia and AMD with SMs and CUs. That's the point of the wave scheduling etc.

 

15 minutes ago, hishnash said:

You misunderstand the rumors are that the macPro SoC will be 4 of these dies on one package, that will include a GPU. What I suggesting is for users that want more GPU power what apple will do is use these SoC packages that have defective Cpu cores.

I understood just fine, it's the second part I'm pointing to as a no. Apple will not do this, completely illogical for many many reasons. And you're mistaken it would cost more for Apple to create a GPU die, in fact it would cost more to try and use CPU dies because namely there will not be enough "defective" ones so they will inevitably have to use perfectly functional dies and disable all the CPUs etc in them rather than just creating a much smaller GPU only die that at the scale of Apple will cost less.

 

Why is it do you think Nvidia and AMD make multiple and successively smaller GPU dies for lower models, that they sell more of. Because it costs less even after the huge design cost.

 

Edit:

P.S I suggest going back and reading my other post again, I had to fix up the difference between LPDDR4X and LPDDR5X because I used dual channel when it's actually quad, but also quad at half the width of desktop channels so eh, it's confusing.