New MacPro predictions? Hopes? Fears?

[Wayne Morellini]

You are big on this desktop at! Thing Antony. I agree with you, but I don't know Apple's timing or intent. I'm still unhappy they canned the cube, now that was cuter than an Mac Pro to me, and great for your Arm desktop between the mini and pro. Unfortunately we might just get another mini, and not see it canned for a cube with as much grunt as a mac pro.

Imagination has announced there much better next architecture. Sure Apple can dodge licensing a bit, but it is just a natural progression, and possibly a customisation beyond what competitors do with general purpose GPU processing. I wonder if a few players might already know and be working on products for next generation Architecture? Look at the current Xavier chip could on Nvidia and compare t Apple a10/11. You talk about desktop processor, but to me a desktop is my Sinclair zx81 or my old commodore 64. You might talk about full os featured instead.

Arm has to respond to the interest in the worked out RiscV processor. While RiscV might have university professors working out this most efficient instead auction set and architecture, Arm can spend a billion dollars modeling and figuring something out better. But still, been the these architectures are the wrong way. I also wish that web assembly could have learned a thing or two from RiscV, to make their virtual binary hardware targetable, the would solve a few issues. My proposed JavaScript binary was going to be tailored specifically for efficient hardware, but then web assembly came along. Any little product I would like to do I can use a JavaScript engine on an arm reference design and allow users to use a JavaScript subset in webassembly. I used to have a lot of issue finding suitable arms as they used 10-100 times too much power for my purposes, but my preferred supplier products were not mature enough even though industry leaders in low energy processing. But now arms are low powered enough and mature enough at low energy to host suitable JavaScript engines. So, I am very happy to see RISCV also.

 

[Antony Newman]

Although the ARM architecture mentioned above does support multi CPUs (i.e. Different vendors) - Apple just publically took this option off the table.

+) X86 for the power (CPU) ... not an Apple ARM core.

This does not preclude an Apple GPU (or other accelerator)

AJ

 

[Antony Newman]

Imagination have been inspired by Riddick to offer an architecture that will scale on the limit of will be needed for most mobiles & tablets for a year or two - but not scale in the way the Nvidia Soc does (i.e. All out performance).

Apple's can currently differentiate themselves in performance per watt in the (speculation) 66% of Imagination GPU they already replaced.
It will be interesting to see what they will be offering.

IMO : ARM does not need to respond to RISCV : They already have Cray, Fitjustsu working with them, and one of the top Supercomputer software analytics and tuning companies pouring their time into extreme systems that work well Re Amdahls law. They have already had a 'nod from Microsoft in the Enterprise space.

AJ

 

[Homer Dulu]

new (old) mac pros and new new mac pro (in the works) :

http://daringfireball.net/2017/04/the_mac_pro_lives

YESSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSS!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!11

 

[Jarek Zabczynski]

I'm glad all the kicking and screaming is paying off. Hopefully we get something more like a traditional tower with drive bays and pci slots. TB3 is great, but not for everything. NVIDIA GPU support is a must! Hopefully a dual processor option too!

Dual Processors, Room for 2 doublewide GPUs and power (2xGTX1080ti) One can dream!!!

 

[Nicholas Croft]

I'm glad all the kicking and screaming is paying off. Hopefully we get something more like a traditional tower with drive bays and pci slots. TB3 is great, but not for everything. NVIDIA GPU support is a must! Hopefully a dual processor option too!

Dual Processors, Room for 2 doublewide GPUs and power (2xGTX1080ti) One can dream!!!

"Full length GPU" (single) was mentioned.

Apple probably still like compact - I'm calling a Dan-Case like box with two GPU's (side by side) and the motherboard on top.

Bodes well for Hacks too - now that everyone and their dog can do it :Angel_anim: - *'natively' keeping pace with whatever the chipset after X99 is. (SkylakeX/KabyX/CoffeeX Lake "X299/LGA2066"?).

Something like this:

https://www.dan-cases.com/

And you can decipher probable CPU/Chipset from this:

http://wccftech.com/intel-x299-chipset-skylake-x-kaby-lake-x/

 

[Wayne Morellini]

Frankly guys, like Apple has, we need a PCIe card standard where the ports can be remote through s header. So, this would allow the card to route to whatever case position, split positions, as well as the standard integrated solution. Then case design is more flexible (not to mention being able to use the full slot ends four fans on standard cases).

Imagination have been inspired by Riddick to offer an architecture that will scale on the limit of will be needed for most mobiles & tablets for a year or two - but not scale in the way the Nvidia Soc does (i.e. All out performance).

Apple's can currently differentiate themselves in performance per watt in the (speculation) 66% of Imagination GPU they already replaced.
It will be interesting to see what they will be offering.

IMO : ARM does not need to respond to RISCV : They already have Cray, Fitjustsu working with them, and one of the top Supercomputer software analytics and tuning companies pouring their time into extreme systems that work well Re Amdahls law. They have already had a 'nod from Microsoft in the Enterprise space.

AJ

I was thinking a post not needing a reply. Can you please explain the previous post Antony?

That is not how business works Antony. If the competitor cash out do you and undercut you, and have the interests and support of a lot of major players in the market, it doesn't matter what present wins they have that are not end game, they can simply win so many future wins they make it look like the o!d AMD Intel fight with you as AMD.

The companies know the above, and support new competition in order to motivate the present supplier to do more and keep things in check, or otherwise get bitten. Arm knows this, so should be up the game to distract from the new competitor so it doesn't become real competition. They know the optional architecture around the chip is years and even more years ahead. (Needing a big IP and circuits cell infusions just to match in years), but they know the ISA is stuffed, and would lead to an x86 like death spiral against a more efficient architecture (such as x86 faced against Arm) if they kept souping it up (like x86 was), which means a new competitive ISA is needed, which fortunately they have already had years of experience with with thumb 1 and 2 etc.

Intel's solution is simple, but they haven't done it yet.

 

[Wayne Morellini]

Anyway, JEFF, has indicated he is under NDA for one thing with Apple. Can he confirm or deny that relates to the modular pro computer they talked about?

 

[Blair Thornton]

This makes me so happy. Now I can stop contemplating the move to Windows. Man I was dreading the thought.

I couldn't imagine that either Jarek.

What I don't understand is this... Apple are saying the current design has thermal limitations because of the graphics cards and the next Mac Pro will be expandable, modular and upgradeable.

Good.

How are they going to do it and be modular?

Thunderbolt 3?

Add the graphics card you want via an expansion case ?
Add any card externally via an expansion case?

What is your guess?

 

[Wayne Morellini]

Read my previous posts on how to do this. You can have slot together modules using a connector between modules. That could be PCIe or whatever, nvlink, TBx50 lines, some optical fibre running at terabits per line long distance. The connectors could be on an active bus, or fork an active or passive backplanes. My proposal back early last decade for a bus like this involved involved bandwidth potentials in petabits with today's technologies. Me first bus solution in the 1990's involved a special unique optical bus proposal which is suitable. These novices are only starting on this stuff.

 

[Jarek Zabczynski]

The classic MacPro was pretty perfect. Go back to that, add Thunderbolt 3 and USBC, paint it black (space grey), done.

 

[Jarek Zabczynski]

This design concept is pretty slick.

http://www.pascaleggert.de/macpro.html

 

[Wayne Morellini]

Lol! great effort.

 

[Antony Newman]

Not sure exactly which bit needs further explanation - but here goes:

+) CPUs are very efficient (performance per watt) for making logical decisions and for unexpected program flow.
+) GPUs are efficient for many currently graphical operations.
+) For new types of 'Professional' applications - there is need of different highly efficient (again in performance per watt) systems that CPU and GPU are inefficient to use to tackle.

AI example
Intel has self-batised itself as an AI company and is Trying use CPU muscle to solve AI problems.
NVIDIA has made a big splash with its Huge GPUs to solve AI problems.
But there are chips out there that that are 15-20 times faster to execute and 1 to 2 orders more Power efficient (perf/per watt).

What ISA make sense going forward :
+) An ISA in which the CPU is efficient & Small - and can efficiently orchestrate the Hardware Acceleration.
+) An ISA that can allow ramping up and down of Watts supplied to cores within this interconnected system
+) An ISA in which BANDWIDTH is the gold dust

If you peruse Goolgle's White papers on their Tensor Flow (TPU) chip, you'll see that even though it is 28nm (ie old old) process technology, it completely decimates both Intel and Nvidias biggest and greatest systems both in performance and Power efficiency (depends on the workload - but Between 20-80 times).

ARM are expecting other companies to build the 'TPU' like capabilities and plug them into a very-high-bandwith Mesh Network that is > 1TB/s

APPLE : They Do need AI like chips for Speech and Image recognition. If Google's White papers are anything to go by, the error rate with AI models is 10 times lower. Always on speech recognition could moved from the Network -> phone, use 1/20th the power of the existing GPU solution, and be 10 times more 'correct'.

Imagination technologies don't have any IP in this area. Apple already have enough knowledge to make their own GPUs.
My guess is that Apple's GPU team is really Apples AI & GPU team. They need their own Silicon to move Network based AI into the Phone.

AJ

<snip> I was thinking a post not needing a reply. Can you please explain the previous post Antony? <snip>

 

[Wayne Morellini]

Although the ARM architecture mentioned above does support multi CPUs (i.e. Different vendors) - Apple just publically took this option off the table.

+) X86 for the power (CPU) ... not an Apple ARM core.

This does not preclude an Apple GPU (or other accelerator)

AJ

I said ”previous" post Antony.

Speech recognition could be done in machine for a long time Antony. We had in this country a pioneer called Dragon Dictate, worked even on slow old Pentiums. With modern accuracy would it not work in machines 10-100+ times more powerful? I had their limited edition android virtual keyboard with speech recognition on my old galaxy years ago. I forget where ever that was local or online. The real reason I fear, that they want it online and to keep it there, is to analysis what you say. What happens past that I can't really say. Do they sell the analysis, pick up who is talking around you and sell that, word by word etc I don't know? We have not to be naive, the days of overriding incentive for competitive performance to the actual customer has gone, replaced by incentive to sell our privacy to illicit third parties who have no right to it. It is just that the law has not caught up to it yet and started throwing people in gaol, retrospectively. If people took enough money around to elected officials they could even make child exploitation legal. It is that bad.

 

[Wayne Morellini]

http://wccftech.com/tsmc-7nm-proces-sram-details/?utm_source=wccftech&utm_medium=related

https://www.nextplatform.com/2017/02/28/amd-researchers-eye-apus-exascale/

 

[Marc Wielage]

The classic MacPro was pretty perfect. Go back to that, add Thunderbolt 3 and USBC, paint it black (space grey), done.

Only if it has six PCI slots (not just 3) and is rackmount. If they do that, every sound mixing company, VFX company, editing company, DIT, and color-correction company would buy it. Six Thunderbolt 3 ports, 6 USB 3.1 ports, upgradable CPUs, upgradable SSDs, upgradable slots for GPUs... it could be done.

 

[Jarek Zabczynski]

Only if it has six PCI slots (not just 3) and is rackmount. If they do that, every sound mixing company, VFX company, editing company, DIT, and color-correction company would buy it. Six Thunderbolt 3 ports, 6 USB 3.1 ports, upgradable CPUs, upgradable SSDs, upgradable slots for GPUs... it could be done.

Well the original had 4 slots. I could live with that if at least two of them were doublewide so you can have two gfx cards without blocking any other slots. Since I wouldn't need a usb3 card, 4 slots would be enough.

 

[Antony Newman]

The classic MacPro was pretty perfect. Go back to that, add Thunderbolt 3 and USBC, paint it black (space grey), done.

(Missed this earlier) It would have made an excellent April 1st video with Ive pouring over every nuance of the new cheese-grater design ;-)

"<snip> a pioneer called Dragon Dictate <snip>

'Speaking of Nuance - this is the 'basic tech that underpins Siri. The algorithm is (IMO) a brute force CPU algorithm that needs hefty hand tuning to increase phoneme separation.

To do more correlation and covariance & without (human) deterministic (algorithm) intervention requires considerably more real time data manipulation than state of the art desktop GPUs can currently muster. This summary view is from the white papers that Google have made available on the subject. These souls admit they are at the starting point in trying to guess the underlying way that this new system should work - but even with TPU chips optimised to perform 65536 MAC (multiply and accumulate) on 6 year old chip fabs - their chips out-performs GPUs by one to two orders of magnitude. Googles white papers on reducing speech error rate are also available.

Wcctech article : Not an insightful piece IMO.

NextPlatform piece : Excellent & Informative. It will be fascinating to see if China will move from 256 mini-monster core chips for their prototype Exascale, or if they'll wait until they get access to a modern fab. In terms of a 'real' super (i.e. One that scales) - that will most likely pop up in Japan. As with Googles view on TPU efficacy, Bandwidth looks to be a key limit ... something that the Fujitsu engineers are no doubt 'chewing on (ie Tofu!).

AJ

Edit : Just seen the following that helps illustrate the efficacy of the TPU for 'that type' of load.

 

[Wayne Morellini]

WPU Bro, WPU. I was meaning back them 20 or so years soon basic hardware. I am surprised they haven't advanced the art enough to gets mid range phone to do this in the cpu's. Notice I didn't say advance the technology, because it is the art to get the technology to do more with less. Of course that is the problem all around, rather than figure out how to do it better minimally, they just brute force primitively. Like when
I learned 3D modeling on some of the most advanced hardware and software around, I got in there wondering where all the good stuff was. It they only had the most basic concepts as state of the art (the cloth module had only just come out), took ten many years to catch up. Just keep overestimating people's abilities.

Anyway, somewhere you were saying if only a small transistor count CPU for arrays. Try 4k transistors (plus memory) at atomic level. Can somebody tell me what volume that could fit in in 3D. I'll to get that design as close to 1000 transistors as practical and eliminate a few things. This is the sort of figures a reverse polish notation stack processing architecture can aim for. My own design target hosts a number of improvements to implement normal features on the small count. With such systems you get into the realm of implementing ten on very fast hardware. It is small enough to attempt a regular photonic implementation, it is so small you could implement either implementation in an optical fiber. Imagine, a whole network of fibre with end points to displays control and IO devices and no other processing needed. If course I just made that up, not what I'm really aiming for but doable, but not the best way to use it. I'm approaching different angles of optical, magnetic, fpga, state machine and model description instead of programs, a number of regular IC and architectural improvements, etc, and something better. If anybody (good) wants to start up a new processing technology company (is chip company), seriously, a hybrid high end platform could be made. The platform would use a mixture of what is practical in different paertrs of the circuit at he time. For starters,one part might be high speed silicon, enhanced by photonics off chip likely, the rest could be low energy silicon still services by the photonic part if needed, even magnetic processing array, and some of my improvements. Now, these technologies allow a number of options in combination, but leaving that behind, we can outperform the mundane commodity orientated silicon they are forcing on the workstation market, holding onto the consumer scale of production processes world rather than performance. Because they still have to shorter run customise the silicon in the commodity factory, they still incur a markup in costs. The aim to sell them the workstation for double, but with at least 10x the performance. The workstation itself largely costs the same, the rest of the price is made up of margin and the extra cost of the processing unit, allowing the processing unit to cost much more than double. If you sold it like normal chips the price of he third party workstation would likely skyrocket and reduce sales. Once new technology is developed internally, the optical and magnetic sections should expand with stepping through 100x to a thousand times speed increases, before getting onto the really big numbers these technologies offer, ten, the really big one using a new technique I've come up with, which is doable now, but is brand new research costly. The figures I'm trying to conservatively quote are based in the real potential of these types of technologies already calculated in the scientific and engineering realms, it is just a matter of how you get there. Now, you notice what I am doing, is taking advantage of the relative isolation of the workstation market (and whatever other fits, maybe com's sector) to generate funds for advanced parts and run numbers to start to push those new parts into lower markets. Tell me you smnart guys, wat sort if processing solution are you going to use when they have 50nm or less pixels (real quantum level digital cinema sensor technology figures) for recording holographic images (real camera technology under development) at nearly 800k squared in surround vr environment?? So, see the future cometh, and not just 8k definition at 96k holographic. So, we could potentially offer the processing card that gets the most resolution on the ramp up to 96k, which I expect will start at 24k and 48k (4k definition by what I have read.). Real old fashion development, not just minor changes. Sure everything is super expensive to R&D now, and setup manufacture, but there still is simpler things to develop I keep my eye on, and strategies to get enough profit margin to do something new.

Sorry about the typos, this new device's keyboard is difficult.

Thanks L for that TPU information, I want to look tat up.