Weapon = FPGA???

[Wayne Morellini]

Gavin, jpeg2000 is an extreme processing hog, it is not the most suitable example. The cineform technology walks over it. Also we are looking forwards to 16-64 core server level Arm technology at the moment. Something will have to break sometime, but I think cine form just might be capable of that on present technology with the latest GPU coprocessing. Just a shame cine form is not on arms.

 

[Antony Newman]

I think that the majority of real CPU software workloads favour the XEON microarchitecture over ARM v8 mainly because the datasets are small, and the code is not designed with parallelism in mind.

But as Moore law runs out of steam - there will be a move to multi core solutions.

The 'big' XEON core maxes out GHz core speeds, Multiple threading, out of order execution performance ... great for single thread performance, but not necessary ideal compute-dense-wise compared with having multiple 'simple' x86 cores.

This is why Intel Xeon Phi doesn't use Multiple 'big' cores. And have been concentrating so heavily on increasing the bandwidth of local memory that is core coherent. And to adapt their Parallel processing engine for a wider variety of tasks (which is why eeking out the FPU performance with less chip area is so important).

From this standpoint ... I think that ARM Microarchitecture is future ready. When software moves to harness the Xeon Phi .... it will also also run effectively on 48 Core ARMs.

Intel : Xeon Lights Landing (14nm) / 60-72 cores / Omnipath / 3TFlops
ARM : Designs down to 16nm / max of 48 cores / hooks to connect to h/w accelerators.

IBM + Nvdia may end up producing the most power efficient silicon (for DARPA/DOE?)
Samsung may buy AMD GPU & APU tech to compete with Qualcomm (when their 14nm yield is up)
Intel will charge everyone ($$$$) for 22 core Xeons ... or ($$$$) 72 Core Xeon Phi Knights landing.
Or TSMC will charge ($$) for a 48 core 3GHz for whomever wants To Cook one.

AJ.

 

[Gavin Greenwalt]

From this standpoint ... I think that ARM Microarchitecture is future ready. When software moves to harness the Xeon Phi .... it will also also run effectively on 48 Core ARMs.

ARM has a *lot* of work to do though that they haven't even really started on such as R&D on virtualization support just off the top of my head. ARM has been so focused on mobile that they haven't invested in the peripheral features that a modern server (or desktop) cpu offers that we take for granted. Nor have I seen much movement on solving their inter-core bandwidth issues. I feel like ARM is today with HPC where intel was with mobile 5-6 years ago. I'm not sure though that ARM will see a compelling reason to even pursue that market or spend money the way Intel has to catch up on mobile.

 

[Antony Newman]

The parallel processing, datarate optimisations and virtualisation extensions in ARMv8.1-A are IMO a suitable microarchitecture for servers. It would not surprise me if Apple adopt this even before ARM release their next big core that is 8.1 compatible.

ARM CoreLink CCN-512 already supports a bandwidth of 1.8 terrabits a second.

My point about 'future ready' is that:
+ ARM (by virtue of not having the stellar single thread performance of Intel, and the evolution of microarchitecture in which a large number of components communicate at very high speed using the on chip AMBA bus) will be well suited for 48 core workloads ... when software is written to advantage of this.
+ Intel (who can squeeze around 22big cores on 14nm, or 72 small cores) is paving the way for software to be written to be more parallel as they are not going to squeeze much more out of their single thread performance.
+ As Intels small cores are comparable in performance to (speculation) ARMs big cores, future code design that targets the Xeon Phi workloads will also help ARM break into that very market.

AJ

 

[Wayne Morellini]

Antony, I think it pays not to argue that a dominant CPU has optimiseds software but to argue what the SRM can do with optimized software down the line. This has been the problem for Intel for years that has stopped them pulling ahead of arm in the mobile race, as their heavily optimized platform is only ahead until arm companies optimise on a faster node.

I'm pretty dick of hearing about processing design, they are just doing what I have been saying for years and years and years ago. As for Phi, I have been aiming to do a x86 compatible processor smaller and more efficient than Arm.