New MacPro predictions? Hopes? Fears?

[Rakesh Malik]

So why is intel abandoning XEON Phi and switching over to AVX-512 etc....

I'm not convinced that Intel IS abandoning Xeon Phi -- it's rocking the supercomputing industry these days.

The bus technology in it has however helped the rest of the processor line also... where do you think Intel got the blazing fast interconnects for the 28-core monsters?

AMD is no slouch in that department, obviously; it's using technology evolved out of HyperTransport, which when it pioneered was truly state of the art high-performance computing technology at consumer prices. It took a while, but obviously AMD hasn't forgotten how to do that, since its 32-core processors are screamers also...

I'm honestly skeptical about Intel being able to compete in the GPU space, but only time will tell.

 

[Misha Engel]

I'm not convinced that Intel IS abandoning Xeon Phi -- it's rocking the supercomputing industry these days.

NVidia is rocking the supercomputing industry these days with their GV100 GPU's

Intel's last announced xeon phi isn't even priced any more https://ark.intel.com/products/series/132784/Intel-Xeon-Phi-72x5-Processor-Family

https://www.top500.org/news/intel-dumps-knights-hill-future-of-xeon-phi-product-line-uncertain/

The bus technology in it has however helped the rest of the processor line also... where do you think Intel got the blazing fast interconnects for the 28-core monsters?

AMD is no slouch in that department, obviously; it's using technology evolved out of HyperTransport, which when it pioneered was truly state of the art high-performance computing technology at consumer prices. It took a while, but obviously AMD hasn't forgotten how to do that, since its 32-core processors are screamers also...

I'm honestly skeptical about Intel being able to compete in the GPU space, but only time will tell.

I don't know what you mean by intels fast interconnects but when I look at NVidia's NVlink 2 with a bi-directional speed of 300 GB/s, that is what I call high speed.
https://en.wikipedia.org/wiki/NVLink

AMD is working on VEGA 20 and the compute speeds look promising 20 TFlops (fp32) grain of salt warning https://wccftech.com/amd-7nm-vega-20-20-tflop-compute-estimation/ and hopefully 10 TFlops (fp64). The current NVidia GV100 get's around 15 TFlops(fp32) and 7,5 TFlops(fp64). Current Xeon Phi's are not even close to these values.

 

[Antony Newman]

"AMD is no slouch in that department, obviously; it's using technology evolved out of HyperTransport, which when it pioneered was truly state of the art high-performance computing technology at consumer prices."

My recollection is that some of the most brilliant scientists developed and then incorporated concepts into the DEC ALPHA 21164 (and later) CPUs.
Unfortunately an ahead-of-their-time science company with in ineffective PR machine does not a profitable company make.

DECs loss was AMD and INTEL's windfall - which is why in the years that followed, INTEL and AMD made huge 'pioneering' gains in their CPU desings... by simply employing the engineers that had developed those ideas while they were at DEC.

http://nerdlogger.com/2007/09/28/dec-alpha-engineers-we-salute-you/

INTEL

Intel are attempting to make offerings where there money to made - even if forcing their way into an established market requires contra-revenue schemes (ie paying companies to use Intel technology), where the bean counters believe that the long term revenue stream will keep their shareholders happy.

There is a lot of money to be made selling dedicated CPU accelerated on-package hardware for hypersclaers that does not get publicity. To more rapidly develop (monetise) this eASIC was brought in-house. Some will use 'Phi' like extensions to accelerate customer workloads.

As there is a huge up and coming market for safety & autonomous compute car system - INTEL will no doubt be working hard to create competitive offerings. The issue with Phi is that it does not have enough dedicated hardware acceleration to deal with NON-CPU style workloads. An INTEL GPU will need 'tensor' like technology (just like the VOLTA architecture) that just happens to also be useful for Graphic Workloads.

Being able to a 4.4 matrix dot product with (lets ~FOUR times) less energy than the existing Phi architecture (for the same process) is going make INTEL relevant in vision processing. If NVIDIA are able eek out 1.5GHz on monstrous GPUs before they combust, and INTEL is able to hit 3GHz with their GPU offering (INTEL are self-destructing themselves at the moment trying to get the transistor to transistor interconnect speed up on their 10nm process) - Intel may actually be able to create a product that meets that very high end market.

We are in an explosive race right now. There are 1000's of companies requiring 28nm SoC, only 100's that are pursuing 14nm due to the CHIP modelling and development costs at the 14nm (Foundry) process, 10's of companies that can afford the 7nm (Foundry) process costs - and ONLY a handful that can afford the 5nm design (until dedicated SoCs are developed that can accelerate that SoC modelling workload).

The FIRST foundry that gets to 5nm is likely to get ALL those handful of foundry customers.

Even if INTEL fails to deliver on the ex-CEO's 2020 GPU promises - they are in the unique position that they can spend billions to create a dedicated process Just for GPUs. By the time the Intel 10nm++ comes out - two foundries processes will be be using EUV to create (slightly) smaller 7nm+ (TSMC and GF sized) transistors.

Even if INTEL 'fail' at 10nm - they can pour money into their next node down - and create a GPU that (with the magic of contra revenue) will establish themselves back in pole position.

AJ

 

[Rakesh Malik]

NVidia is rocking the supercomputing industry these days with their GV100 GPU's

I'm well aware of that. It's not a club that's exclusive to one or the other after all.

https://www.top500.org/news/intel-dumps-knights-hill-future-of-xeon-phi-product-line-uncertain/

HPC is a lucrative market... I have a feeling that Intel pulled Knight's Hill because of the same manufacturing problems it's been dealing with for Coffee Lake and all, but doesn't want to admit it.

I don't know what you mean by intels fast interconnects but when I look at NVidia's NVlink 2 with a bi-directional speed of 300 GB/s, that is what I call high speed.
https://en.wikipedia.org/wiki/NVLink

The fact that nVidia can do it doesn't mean that no one else can. The 16+ core Xeons are possible because of the same interconnect that Intel used in Xeon Phi after all... they'd perform terribly without it. Intel's ring bus has a near-zero latency for communication between cores on a chip. I don't remember the numbers, but I did read about how it works; it's pretty impressive. IIRC it's called Infinity Fabric.

AMD is working on VEGA 20 and the compute speeds look promising 20 TFlops (fp32) grain of salt warning https://wccftech.com/amd-7nm-vega-20-20-tflop-compute-estimation/ and hopefully 10 TFlops (fp64). The current NVidia GV100 get's around 15 TFlops(fp32) and 7,5 TFlops(fp64). Current Xeon Phi's are not even close to these values.

AMD, nVidia, and Cavium have probably lit a big fire under Intel's ass.

I suspect that a *lot* of this is due to Intel's manufacturing problems. It's well known that the problems have delayed several of Intel's processor design plans, so the management for the Xeon Phi side of the world is probably actually glad that Argonne delayed the project, because it gives Intel another shot at it. If it had NOT been delayed, it would make losing the contract to AMD or nVidia or Fujistu or (etc) more or less inevitable.

But I still doubt that Intel will actually drop the Xeon Phi, because it's a relatively small volume market that also has rather high volumes.

Unless Intel believes that it can scale up it SkyLake successors to the point where they'd outperform Xeon Phi, and it's just sandbagging in the hopes that it can get that "next generation" platform launched before it loses too much ground to the competition (kind of like the partial MacBook Pro update ).

 

[Antony Newman]

The DEC Alpha 21164 guy (JK) is now at Intel to help sort out things :
https://www.anandtech.com/show/13048/an-anandtech-exclusive-the-jim-keller-interview

My reading of the interview:
+) He is helping undo false assumption on 10nm process.
+) Intel are keen to process non traditional workloads.
+) The size of the GPU Intel can create should be quite large,
+) Intel’s 7nm process will be more hiccup-free.
+) Intel are going to have another crack at mobile.
+) Intel could well be looking at much larger bulding blocks for Supercomputers,

AJ

 

[Rakesh Malik]

I wonder what Intel had to offer him to entice him on board in light of how many other semiconductor companies seem to want him working for them

It will be interesting to see if he's able to get Intel through its manufacturing problems with the latest process nodes. Right now, Intel is a process node BEHIND the industry... it will be interesting.

I'm still skeptical about the Intel GPU, but at least it might get the processor line moving again, so if AMD keeps on pace, things ought to be very interesting next year.

 

[Antony Newman]

It is incredible that JK is working for a behemoth of company. JK distills why he chooses assignments - indicating that this will be the first time he will oversee such a comprehensive range of technology and diverse range of technicians (5 - 10 times larger than any previous role).

It takes a special type of mind that can be and is meticulous at every engineering level. If he was brought in to find a technical solution for Intel’s 5-7 year strategies, it bodes well for Intel’s future.

With Intel dropping the ball at (Intel) 10nm - it looks like AMD won’t INITIALLY be releasing a 64 Core Monster CPU on TSMC 7nm - because Intel won’t even be able to ‘answer’ a 48core beast that has double the intra-chip and external bandwidth of AMDs first offering (not to mention the IPC tweaks And the likely sizeable jump in base core frequency).

My guess is that AMD will make a deal with GF to offer the 64 core version on (7nm+) EUV (to meeting their minimum manufacturing obligations) and that the product will be the Killer data centre piece that AMD will use to shore up their balance sheets (especially if it is offered at a premium).

In the HPC world - I wonder if SoC area dedicated to hyper-threading will give way to lower latency single threading cores that could operate at higher single threaded performance? At some point throwing more cores is less effective in workload acceleration compared to increasing individual cores GHz + IPC.

Either way, fierce competition and rapid innovation are going to make 2019 an exciting year for equipment purchases; With the jumps in memory bandwidth perhaps 8K footage & 4K timelines will be possible without consuming the lions share of the systems resources?

AJ

 

[Antony Newman]

INTEL PHI.

With no 10mm version of the Phi - Intel just couldn’t push a ‘co-processor’ to compete Head to Head with GPUs with Compute extensions (Half Precision and Tensor processing).

https://www.anandtech.com/show/13111/intel-begins-eol-plan-for-xeon-phi-7200-series-knights-landing

If uptake of Knight Mill is also like warm - Intels attempt to corner the Networking market with Omni-Path & railroad companys to use it could well backfire.

Nvidia is the only other major company pushing a proprietary solution (and having some success) ... I wonder if Intel will be forced to join the emerging open standard .. and finally sign up to the Gen-Z consortium?

(Translated from Japanese) https://translate.google.com/transl...column/kaigai/1053318.html&edit-text=&act=url

In 2017 AMD (Papermaster) stated that they will be migrating their INTER-SOC from (what I had incorrectly thought was the Infinity Fabric) to ... GEN-Z. Rome might be too soon to switch (core Gen-Z 1.0 spec published in 2018_02) ... but who knows how aggressively AMD will leverage their (likely) pole position and try and keep Intel out of their slipstream? (My guess is AMD will use Gen-Z in Milan).

I haven’t found any articles that tie Apple to Gen-Z yet - but if Apple replace the ‘e-Gpu’ Point-to-point with a crazy high speed (Gen-Z) bus - it will be enough to ‘commute their architecture to silicon photonics when it eventually materialises.

AJ

 

[Rakesh Malik]

INTEL PHI.

With no 10mm version of the Phi - Intel just couldn’t push a ‘co-processor’ to compete Head to Head with GPUs with Compute extensions (Half Precision and Tensor processing).

Sounds like Intel's giving up after all... I've been hoping that Intel would double down on Xeon Phi unlike with Itanium because 192 cores on a single chip would be huge when teamed with a contemporary GPU.

It's going to be interesting to see what happens with JK's leadership. This being the first time since the Pentium that Intel has NOT been leading the process technology world, it's a sign of interesting times -- especially when one of the leaders is now the foundry that AMD and Fujitsu formed (Global Foundries).

And finally some serious competition in the processor space... x86 has been so dominant for so long that it's squeezed out just about everything else from the personal computer, gaming, and even server markets, even though ARM sales have been by volume several orders of magnitude larger.

 

[Jarek Zabczynski]

The new RTX offerings from NVIDIA got me pretty pumped. Hopefully these will compatible with Macs going forward.

 

[Bob Zelin]

Jarek -
why would you even think that the new NVIDIA RTX will be Mac compatible. As of 2018, Apple is doing everything they can possibly do to make NVidia drivers NOT work with their operating systems. Unless you hack the macOS with instructions from sites like egpu.io (which often don't work), you cannot get any NVidia cards to work with a Mac - not even with eGPU boxes that are approved by Apple. And Adobe is not currently supporting any AMD eGPU rendering enhancements (only Blackmagic is right now). I think that Apple will do everything in their power to make sure that NVidia cards NEVER work with any Mac computer that comes out in the future. I think that Apple will put out special effort to make sure that if NVidia DOES release a driver that works, Apple will release a "dot" release of their OS to make sure that things stop working. That's certainly the way things are right now, and if by some miracle a NEW Mac Pro does come out in 2019 or 2020, I bet you a nickel that NVidia cards will NEVER be allowed inside this new Mac Pro. Let me take this one step further. I believe that NO HARDWARE from any manufacturer will EVER be allowed inside a Mac product ever again, unless they become an "Apple partner".

Bob Zelin

 

[Brian Timmons]

Jarek -
why would you even think that the new NVIDIA RTX will be Mac compatible. As of 2018, Apple is doing everything they can possibly do to make NVidia drivers NOT work with their operating systems. Unless you hack the macOS with instructions from sites like egpu.io (which often don't work), you cannot get any NVidia cards to work with a Mac - not even with eGPU boxes that are approved by Apple. And Adobe is not currently supporting any AMD eGPU rendering enhancements (only Blackmagic is right now). I think that Apple will do everything in their power to make sure that NVidia cards NEVER work with any Mac computer that comes out in the future. I think that Apple will put out special effort to make sure that if NVidia DOES release a driver that works, Apple will release a "dot" release of their OS to make sure that things stop working. That's certainly the way things are right now, and if by some miracle a NEW Mac Pro does come out in 2019 or 2020, I bet you a nickel that NVidia cards will NEVER be allowed inside this new Mac Pro. Let me take this one step further. I believe that NO HARDWARE from any manufacturer will EVER be allowed inside a Mac product ever again, unless they become an "Apple partner".

Bob Zelin

If this is this case then everyone who cares needs let Apple clearly know that completely blocking out NVIDIA is unacceptable for Pro needs.

Below is the feedback form for the MacPro (yes it for the old model but express concerns for the new model in post).
https://www.apple.com/feedback/mac-pro.html

If you give a damn, let them know.

Brian Timmons
BRITIM/MEDIA

 

[Misha Engel]

If this is this case then everyone who cares needs let Apple clearly know that completely blocking out NVIDIA is unacceptable for Pro needs.

Below is the feedback form for the MacPro (yes it for the old model but express concerns for the new model in post).
https://www.apple.com/feedback/mac-pro.html

If you give a damn, let them know.

Brian Timmons
BRITIM/MEDIA

Pro is more than just .R3D.

Pro to apple is more something like ProRes (smart move from ARRI/Apple) and now ProResRAW with Sony FS7, FS5 and Panasonic EVA1(which is a hugh market compared to RED camera's).
Apple gives their cash cows what they want, stylish expensive all-in-ones/laptops, light codecs that runs on their hard/software, GPU/storage expansion with Thunderbolt.

Why would they help a company they don't like(NVidia) or help a competing codec(.R3D)?, when they have a partner that dominates the professionel camera market(ARRI) and partners that help them sell more Mac's and promoting their new codec(Sony, Panasonic and Atomos) for wedding shooters/lower budget productions.

 

[Antony Newman]

Bob,

Thats sounds slightly pessimistic.

+) I am not convinced that eGPU in its current form has the sort of extensible bandwidth to scale at the rate that the Apple Pro Workflows team would have mandated for the MMP.

+) Apple MAY have been banking on Intel 10nm & Chiplet tech to be out in 2019 (which is now 2020). But as Apple have NOT made an announcement of a slip in the schedule - MMP doesn't look like its going to be on Intel 10nm.

+) What might be interesting is if Apple mandate that GPU's include Metal 2.0 drivers for certification. If Apples 'attacks the high end 8K market - would NVidia be willing to meet Apple requirements? AMD definitely would; Nvidia - not so sure - as they are fighting to keeping themselves out of emerging high end open standards, and pushing for their own proprietary interface.

+) If Apple has the sort of (Silicon) engineering prowess that Tesla has - perhaps we will see Apple offer an auxiliary GPU of their own.

AJ

 

[Brian Timmons]

Pro is more than just .R3D.

Pro to apple is more something like ProRes (smart move from ARRI/Apple) and now ProResRAW with Sony FS7, FS5 and Panasonic EVA1(which is a hugh market compared to RED camera's).
Apple gives their cash cows what they want, stylish expensive all-in-ones/laptops, light codecs that runs on their hard/software, GPU/storage expansion with Thunderbolt.

Why would they help a company they don't like(NVidia) or help a competing codec(.R3D)?, when they have a partner that dominates the professionel camera market(ARRI) and partners that help them sell more Mac's and promoting their new codec(Sony, Panasonic and Atomos) for wedding shooters/lower budget productions.

I hear you Misha.
Consider though that we are the Pros they sell to.
We determine the market based on how we use the product to meet the demands of our craft.
While Apple can influence industry trends they do not dictate them.

I’m a big fan of Apple but admit they occasionally drop the ball if not damn near lose it.

Apple, however, does listen.
https://www.theverge.com/2017/4/4/15175994/apple-mac-pro-failure-admission

Discussion like this thread is great. Discussion is even better when it’s delivered directly to the subject of the discussion.
https://www.apple.com/feedback/mac-pro.html

If that fails vote with your wallet.

Apple didn’t become a trillion dollar company by being morons.

Brian Timmons
BRITIM/MEDIA

 

[Misha Engel]

Apple didn’t become a trillion dollar company by being morons.

Brian Timmons
BRITIM/MEDIA

The people at Apple are certainly no morons.
They look for consumers (I can't even call them customers, because customers expect customer service) that are willing to pay a high price for medio-core hard/software and don't complain, even defend Apple when it doesn't work. History tells us Apple did their marketing well (knowing what wannebee's and other lifestyle types want).

Pro's buying underpowered hard/software, knowing that it's underpowered on the otherhand...

 

[Brian Timmons]

The people at Apple are certainly no morons.
They look for consumers (I can't even call them customers, because customers expect customer service) that are willing to pay a high price for medio-core hard/software and don't complain, even defend Apple when it doesn't work. History tells us Apple did their marketing well (knowing what wannebee's and other lifestyle types want).

Pro's buying underpowered hard/software, knowing that it's underpowered on the otherhand...

I can agree to an extent with some of what you write.

I do believe part of the problems with Apple’s relationship with the
Pro user is they started treating them like consumers and dismissed the
unique needs of that user base.

I have to ask though.
Are you mainly a PC user with passing interest in the Mac?

Brian Timmons
BRITIM/MEDIA

 

[Jarek Zabczynski]

Jarred met with Apple several months ago, and while we don't know the specifics of those meetings, Jarred seems pretty positive towards Apple on what went on there. Yes, shortly after this Apple began selling the Raven kits, but I'm pretty sure it was about more than just selling cameras. So that leaves me hopeful that we'll see a machine that's more "pro" friendly than the trashcan we have now.

 

[Brian Timmons]

Jarred met with Apple several months ago, and while we don't know the specifics of those meetings, Jarred seems pretty positive towards Apple an what went on there.

That says a lot and is very encouraging.
Hope the results live up to the potential.

Brian Timmons
BRITIM/MEDIA

 

[Misha Engel]

Are you mainly a PC user with passing interest in the Mac?

Brian Timmons
BRITIM/MEDIA

At the moment we only use Mac for ProRes encoding (2..4 times a year) when a customer demands it (not that often in Europe) is willing to pay for it and/or if we otherwise loose the customer.
We have a mac in the reception(yes as a design statement) and the big boss has one(and yes he runs windows on it because he couldn't get used to MacOS) and we have some old macbook air's that are oftened used by the radio department.

Our threadripper ws runs everything, one storage server runs BSD the other windows server, the 4 older workstations run windows and the other pc's/laptops also run windows(defacto standard in Europe).

All operating systems have their pros and cons.