New MacPro predictions? Hopes? Fears?

[Mike Halper]

I just built a new hackintosh. Was tired of waiting. 10 cores, sync all cores to 4.5 GHz turbo, 64GB RAM, 1080TI FTW3 Gaming, 43,000+ Geekbench, ~2,500 Cinebench, continuity, handoff, iMessage, etc all works. Completely stable and solid. Thunderbolt 3 works, even with hot-plugging too, as does USB 3.1 gen 2 ports. Even the 18-core iMac Pro is only under 48,000 Geekbench, and the 10-core iMac Pro is only 35,000 Geekbench. About $4000 (not including existing monitors), so a lot less than buying an iMac Pro.

 

[Misha Engel]

Just wait 2 more years and you can do everything (yes everything) on the iPhone XXX.

 

[Antony Newman]

Apple's eGPU over TB3 (5GB/s) has been both a technology stopgap and successful demonstration of modularity.

Perhaps this is indicative of their overall direction - where lightweight devices can doc with GPU powerhouses located out of earshot (or submerged in an Aquarium Cooler ;-) )

If Apple are to modularise everything - they could glue the machine together with bits and pieces of PCIe4, NVLink, CCIX, Infinity, CAPI and OpenCAPI.
My guess (hope) however is that we get to see Gen-Z in action in 2019.

https://www.csm.ornl.gov/workshops/...chnology Overview, Trends, and Alignments.pdf

With 448GB/sec bandwidth on tap - there probably enough to keep a couple of 64 core CPUs and a few GPU's busy.

... might be a project that AMD and Apple get to work together?

AJ

 

[Misha Engel]

Apple's eGPU over TB3 (5GB/s) has been both a technology stopgap and successful demonstration of modularity.

AJ

What kind of magic does thunderbolt 3 use to to get 40 Gbit/s=5GB/s over 4 PCIe 3 lanes (985 MB/s each).
In my world 4 x 985 MB/s = 3.94 GB/s = 31.52 Gbit/s.

What am I missing...

 

[Blair S. Paulsen]

What kind of magic does thunderbolt 3 use to to get 40 Gbit/s=5GB/s over 4 PCIe 3 lanes (985 MB/s each).
In my world 4 x 985 MB/s = 3.94 GB/s = 31.52 Gbit/s. What am I missing...

AFAIK, your'e spot on - if you only have 4 lanes of PCIe3 backing a TB3 port then you would be bottlenecked by that bus, regardless of the theoretical bandwidth. Almost as misleading as buying a single spinning 3.5" drive with a USB 3.1 SS interconnect that claims "up to 5Gb/s" on the box.
In May of 2017 Intel announced plans to support TB3 "on die", which might support the full 5GB/s noted.

In any case, as Andrew notes, newer tech more appropriate for shredder GPUs is at hand. I just want TB3 for storage pools and networking, where even 3GB/s should hold me for a while... ;-)

Cheers - #19

 

[Karim D. Ghantous]

I just built a new hackintosh. Was tired of waiting. 10 cores, sync all cores to 4.5 GHz turbo, 64GB RAM, 1080TI FTW3 Gaming, 43000+ Geekbench, ~2500 Cinebench, continuity, handoff, iMessage, etc all works.

Yes, yes, yes. But what is the Monero hashrate?

 

[Antony Newman]

It looks like Fujitsu is sampling their protype ARM processors (likely to be used in the worlds first ‘green’ exascale supercomputer)

One interesting point is that they are using OPM (On Package Memory).

Rather that have ‘slow’ 80GB/sec memory that is that on the other side of the motherboard, they are connecting memory using a wafer-jigsaw technology (ie directly to the SoC that houses 48 Fujitsu cores).

This proximity enables a > 1TB/sec bandwidth - more than enough to avoid starving any core that needs memory.

https://mobile.twitter.com/ogawa_tter/status/1011236798997991424/photo/3

I wonder if Apple are going to go this route? They are the ones that pushed TSMC to rapidly develop InFO (integrated fan-out technology) to stack 3D memory in the iPhone.

If Apple does go the Gen-Z route - perhaps we will see BIG CPU modules akin to the 2007 8 Core Mac Pros?
It would certainly be a modular to be able to plug in CPU (with OPM) into the Gen-Z ‘backplane’ box.

An even more unusual route would be for Apple to switch to Fujitsu’s chip.
With each core ‘almost’ as powerful as a XEON core, but running with the efficiency of a GPU - I suspect that their 48 cores prototype has comfortably overtaken Intel’s top of the line 28C chips.

AJ

 

[Frank Glencairn]

MY prediction is some ass clowns are still going to claim they can build their PC cheaper.

LOL- Classy Stockholm Syndrome victim in denial mode.

 

[Misha Engel]

I expect that Apple will go the ARM way, looking at the current speeds that can be achieved with this architecture.

https://www.servethehome.com/updated-cavium-thunderx2-power-consumption-results/

https://www.servethehome.com/cavium-thunderx2-review-benchmarks-real-arm-server-option/

https://www.top500.org/news/benchmarks-in-hand-uk-academics-see-promising-future-for-arm-chips-in-hpc/

It is probably gone be a very high-speed iphone hence iPhone XXX.

For me the GPU is more of a question since NVidia and AMD have so much GPU IP that it will be hard for a newcomer to develop something with sufficient speed.

 

[Rakesh Malik]

An even more unusual route would be for Apple to switch to Fujitsu’s chip.
With each core ‘almost’ as powerful as a XEON core, but running with the efficiency of a GPU - I suspect that their 48 cores prototype has comfortably overtaken Intel’s top of the line 28C chips.

AJ

The Fujitsu processors look very interesting. I'm going to see if I can find out more about them now.

It's pretty unlikely that any current ARM cores will be genuinely competitive with the flagship x86 cores, especially for floating point. The Sky Lake core and its successors all have dual 256-bit vector FPUs, massive caches with bandwidth out the wazoo, and extreme clock speeds. Per watt, ARM is faster... but when you turn off the power restrictions, x86 runs WAY ahead.

Even Ryzen still only has 128-bit vector FPUs, so AMD is using higher core counts to match Intel's performance. It's working now, but just see what happens when AMD upgrades those to 256-bit vector units (and Intel migrates the currently server oriented AVX 512 to the rest of the line).

Is it possible for ARM to get there? Yes, technically speaking. The open question is whether or not the market will sustain it, and right now, the answer is a resounding "No." Anyone who wants that market to exist will have to create it and be willing to accept a fairly slow ramp initially, playing a rather long game in order to get there.

It IMO be good for everyone; we'd have a genuine competitor for x86, in addition to two x86 titans in competition with each other.

 

[Antony Newman]

CPU

"Is it possible for ARM to get there?"

With NASA going Arm for the long haul, and Ex-Intel (Ampere) going Arm, and Cray & Fujitsu pushing forward for a rounded scale-out solutions - it bodes well for the longevity of the ARMs v8 architecture. If RIKEN/Fujitsu's SVE extensions fit well (ie are as efficient for) many AI workloads - who knows - there could be ARM APU's at some point in the future (2023?).

Cavium is riding the wave on a 16nm surfboard - and EU supercomputer installers due to wait for their next (ThunderX3) iteration have decided that the X2 is good enough (bringing forward their rollout to this year). I hope that Cavium ditch their internally designed (pre-Vulcan) - and concentrate on getting a 7nm part out the door (ie the TSMC 7nm 7HPC process). If the 7HPC process burns 65% less power - they Might be able to create a chip with a Base clock of 3GHz.

In 2019 - ARM themselves may finally release 'Aries' (their Huge Server Core - most likely designed like Fujitsu's to minimise Latency .. and NOT area). At 7nm - we might Finally see what the ARM bods can do (or should that 'Qualcomm' can do when they put a boat load on a single piece of Silicon?)

GPU

If Apples next (A12) GPU on TSMC 7nm Low Power increases their Performance/watt by 40% - I think it demonstrates that architectural implementation of IP has a greater effect on performance that raw IP (especially with 'Lal Shimpi burning midnight oil under the Apple tree). This being why ARMs on A76 (CPU) had a huge jump in performance. If Apple are going to have 8K screens in 2019 - the post-Mojave macOs is going to need 8K GPUs. And if Apple are pushing for Augmented reality - it would not surprise me if Apple enter the market with a mobile oriented 8K GPU?

The only other contender for a GPU is ... (don't laugh!) .. Intel. Timeline : 2018 Q3 Intel announce quest for CEO & that they seem to have lost a rather large customer that was hoping to use their 10nm process. Q4 new CEO is announced. 2019 Q1 10nm goes into High-volume-manufacturing (with Contra-venue while they fixed the (guess) dielectric / brittleness issues with Cobalt and announce a shift back to Copper (or Ruthenium) ). 2019 Q2 Intel announce a 'Better than Vega' GPU ... which Finally arrives in 2019 Q4 using Intel 10nm...

AJ

 

[Antony Newman]

"It's pretty unlikely that any current ARM cores will be genuinely competitive with the flagship x86 cores"

It would not surprise me if Fujitsu's Post ARM chip has a higher IPC than Intel - however, Intel's process is pushed to the extreme to exact a higher GHz than the foundries (which might be a contributory factor to their 10nm disaster - where minimal tolerances for Cobalt at stellar frequencies might be causing the Intel chips to ignite).

Even if TSMC offer 4.4GHz Cache-line IP on their 7HPC process ... I doubt Fujitsu are going to be pushing the cores over 3GHz.
I would therefore speculate that for (every) Fujitsu's core to be 'almost' Xeon performance - what they are really indicating is that:
1) Their Cores have Higher IPC than the XEON - but even when clocked 30% slower the Fujitsu ARM CPUs process at approximately the same rate as the Xeons.
2) ALL the cores can run flat out (ie no Power bottleneck or Bandwidth bottleneck when they are all running)
3) Fujitsu likely did Not 'do' and AVC512 and slow the clock processor down. They designed the SVE to outperform both (existing) SIMD and IntelAVC512.

As Fujitsu also extended the ARM ISA (instruction set architecture) to include HALF precision operations (recently added to GPUs to make them more effective in AI loads) - it would also not surprise me if Fujitsu have made these operations as efficient than the equivalent function in a GPU (otherwise it would be a waste of silicon). It follows that half precision could well surpass all x86 implementations in terms of performance / watt).

(much of the above is conjecture!)

AJ

 

[Misha Engel]

The Fujitsu processors look very interesting. I'm going to see if I can find out more about them now.

It's pretty unlikely that any current ARM cores will be genuinely competitive with the flagship x86 cores, especially for floating point. The Sky Lake core and its successors all have dual 256-bit vector FPUs, massive caches with bandwidth out the wazoo, and extreme clock speeds. Per watt, ARM is faster... but when you turn off the power restrictions, x86 runs WAY ahead.

Even Ryzen still only has 128-bit vector FPUs, so AMD is using higher core counts to match Intel's performance. It's working now, but just see what happens when AMD upgrades those to 256-bit vector units (and Intel migrates the currently server oriented AVX 512 to the rest of the line).

NVidia TitanV/GV100 have 15 TFlops (fp32) and 7.5 TFlops (fp64) CPU's aren't even close to those kind of value's.
GV100 has a mem bandwith of 900 MB/s, CPU mem bandwith is a lot lower.
(Almost) everything AVX 128, 256, 512, 1024, etc... can do, GPU's can do faster.

CPU

"Is it possible for ARM to get there?"

GPU

If Apples next (A12) GPU on TSMC 7nm Low Power increases their Performance/watt by 40% - I think it demonstrates that architectural implementation of IP has a greater effect on performance that raw IP (especially with 'Lal Shimpi burning midnight oil under the Apple tree). This being why ARMs on A76 (CPU) had a huge jump in performance. If Apple are going to have 8K screens in 2019 - the post-Mojave macOs is going to need 8K GPUs. And if Apple are pushing for Augmented reality - it would not surprise me if Apple enter the market with a mobile oriented 8K GPU?

The only other contender for a GPU is ... (don't laugh!) .. Intel. Timeline : 2018 Q3 Intel announce quest for CEO & that they seem to have lost a rather large customer that was hoping to use their 10nm process. Q4 new CEO is announced. 2019 Q1 10nm goes into High-volume-manufacturing (with Contra-venue while they fixed the (guess) dielectric / brittleness issues with Cobalt and announce a shift back to Copper (or Ruthenium) ). 2019 Q2 Intel announce a 'Better than Vega' GPU ... which Finally arrives in 2019 Q4 using Intel 10nm...

AJ

Meanwhile AMD and NVidia sit on their hands and do nothing... I don't think so.
NVlink, chiplets, infinity fabric, 7 nm,... just to name a few.

 

[Antony Newman]

Misha,

It seems like an impossible feat - or a very impressive one if they succeed.

(For Rakesh)
https://www.nextplatform.com/2016/06/23/inside-japans-future-exaflops-arm-supercomputer/

POWER EFFICIENCY

If the Post-K Delivers 1,000 Petaflops in 40 MW of power - very roughly that is 25 TFlops/Watt (if my maths is correct).

To put that into another perspective - Fujitsu reckon that their Post-K architecture is 30 TIMES more power efficient that their previous supercomputer (from 2011). Quite a jump!.

It could be that XEON Phi's and other GPUs are going to be working in conjunction with the Fujistu ARM chips to reach this stellar efficiency target. This could imply that the Next batch of GPU's have a similar jump in Performance per watt. If Fujistu's EXAFLOP performance depends on the in-house developed ARM chip .. I speculate it would have to Perform at a similar efficiency as the GPUs. (I note that Fujitsu's own slides suggest that it has in places GPU like efficiency)

Sidenote : HBM3, Infinity Fabric 2, PCIe5, NVlink 3.0 - none of these Future technologies match the Bandwidth (or low power consumption) of what Fujitsu are using for its (relatively small quantity of) On Package Memory. Fujitsu's research found that this per core bandwidth a major limitation in being able to effectively keeping all the cores busy.

AJ

 

[Misha Engel]

How about this one AJ https://spectrum.ieee.org/tech-talk/semiconductors/design/amd-tackles-coming-chiplet-revolution-with-new-chip-network-scheme On Package Everything.... http://www.guru3d.com/news-story/amd-is-considering-chiplets-to-connect-future-cpus-and-gpus.html

 

[Antony Newman]

Misha,

CHIPLETS

Thankyou for the links - I remember AMD had white papers on supercomputer architecture in which APU's (or clusters of CPU and GPU) were coupled with local high bandwidth memory as a way reducing latency and reducing power consumed when routing data on motherboards (by keeping it closer proximity to the CPU).

Intels developed EMIB which is (I believe) still more advanced at incorporating 'chiplets' than what was diecribed by the AMD engineers and fellows. Intel's PR machine extolled the virtues of how amazing EMIB was - because it was "going to be sooo much cheaper to manufacture" for them.
Roll forward - I have not seen any evidence of high volume customer parts that use this technology.
My guess is Intels (supposed) large customer (that might be going under due to Intels 10nm woes) was hoping to use EMIB & Intels 10nm process (something we should find out about before the fiscal end of year).

Real World Performance vs FLOPS

What has differentiated RIKEN/Fujitsu in their supercomputer design has been that they championed a system to maximise real world performance over BIG PR numbers.

Fujitsu : K Supercomputer : 2011 : 10 PFlops
China NSC : Sunway TailuLight : 2016 : 93 PFlops

http://www.hpcg-benchmark.org/custom/index.html?lid=155&slid=295

The Sunway has an order of magnitude more FLOPS, but the HPCG test shows it runs about 25% SLOWER in real world applications.
(For reference - the 200 PFlop Summit chocked out with Volta GPUs is 5x faster than the 10 PFlop K super on the same real world test)

With the Post-K Supercomputer - Fujitsu are concentrating on Real world performance again - and if successful are aiming for a super that is TWO Orders of Magnitude faster than the K (or about 20 times faster than the Just build Summit that uses top end XEONs and Top end VOLTAs).

CORE SCRATCHPAD MEMORY <--> ON PACKAGE MEMORY

I am guessing here - but I think in analysing how to exa'scale up Fujitsu's old K super - they determined that:
+) Utilisation of a lump of ON-CORE memory with >10 times the bandwidth of the fastest (external-to-cpu-core) On-Package (HBM) memory was mandatory to minimise power consumption.
+) That AI like functionality running in a dedicated core will ultimately trump GPUs in perf efficiency. Fujitsu will most likely create their own AI accelerators (they call the DLU chips) with (same at the post-K chips) their own 6D network-on-a-chip fabric. The DLUs will probably be retrofitted into the Post K (when the DLUs eventually see the light of day).
[Sidenote .. I am guessing this is what Nvdia will be announcing .. a fork in the road for Gaming vs ML_n_Compute in Q4).

One of the finer aspects of all the research that RIKEN and Fujitsu are undertaking is that they are going to open-source it back to the HPC communities.

As Fujitsu have been working with the Intel Phi SoC engineers (to optimise real world 32bit maths into 16 bit representation) - my understanding is that Intel themselves will be updating their own (Phi) accelerators that improve the perf/watt by 4 times (for a 2019 release window).

AJ

 

[Misha Engel]

Always nice to hear something about super computing, for the mortals among us the chiplets are more interesting for fast, energy efficient and portable compute power.
Weather it's Apple (ARM + GPU), NVidia (ARM + GPU), AMD (X64 + GPU) or intel(x64 + GPU + FPGA) as long if it is afforable, open, (portable) and fast enough I am a potential customer.

 

[Rakesh Malik]

Intels developed EMIB which is (I believe) still more advanced at incorporating 'chiplets' than what was diecribed by the AMD engineers and fellows. Intel's PR machine extolled the virtues of how amazing EMIB was - because it was "going to be sooo much cheaper to manufacture" for them.
Roll forward - I have not seen any evidence of high volume customer parts that use this technology.

Kaby Lake-G. It's out in the wild, and is available in laptops from HP and Dell.

What has differentiated RIKEN/Fujitsu in their supercomputer design has been that they championed a system to maximise real world performance over BIG PR numbers.

Fujitsu knows it shit. It's been making better SPARC processors than Sun for ages, even before Oracle snapped the dying remnants up. Fujitsu has also been making supercomputers for a long time... so it's no surprise that it can make machines that can compete with the stuff from NEC and Sandia and the like.

So if Fujistu is developing a high performance ARM core, then it's something to pay attention to.

As Fujitsu have been working with the Intel Phi SoC engineers (to optimise real world 32bit maths into 16 bit representation) - my understanding is that Intel themselves will be updating their own (Phi) accelerators that improve the perf/watt by 4 times (for a 2019 release window).

In the x86 world, I believe that the Xeon Phi is the most promising look at the future. It's based on streamlined cores and uses an architecture that resembles that of Tesla and Pascal a lot more than anything else. And Intel has a lot of experience with interconnects and bus technologies, which is the secret sauce for Xeon Phi.

 

[Misha Engel]

In the x86 world, I believe that the Xeon Phi is the most promising look at the future. It's based on streamlined cores and uses an architecture that resembles that of Tesla and Pascal a lot more than anything else. And Intel has a lot of experience with interconnects and bus technologies, which is the secret sauce for Xeon Phi.

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

 

[Antony Newman]

Intel’s Xeon Phi engineers will still be able to share what they have learned in optimising the data representation and increasing math processing (by a sizeable jump).

The Intel 2019 release window now looks like 2020 window.

In what universe would you ever have thought that you would want an AMD CPU and and INTEL GPU?!

AMD_Milan (TSMC EUV 7nm+) + INTEL_GPU (Fingers crosses 10nm++)

AMD for the CPU because a DUAL SOC - no (Intel) crippling; probably 96 core and a boat load of PCIe5
INTEL GPU - because Intel’s process is less likely to be less ‘dark’ at high frequencies ** ***

** : GPUs have more active transistors proportionally than CPUs (more parallel compute).
Their transistors are designed for lower leakage (which means less wasted heat).
(Dark silicon is where you have to trottle down or switch off transistors due to heat issues)
Which should mean a sisazble jump in SoC base frequencies compared to the foundries.
It remains to be seen if Intel hit this one out of the park - but am betting the
new CEO will be looking to win investor confidence back.

*** : Intel might price themselves above Nvidia...

AJ