Intel's upcoming 10nm and beyond

Discussion in 'Hardware Components and Aftermarket Upgrades' started by ajc9988, Apr 25, 2019.

  1. tilleroftheearth

    tilleroftheearth Wisdom listens quietly...

    Reputations:
    4,695
    Messages:
    12,157
    Likes Received:
    2,204
    Trophy Points:
    631
    ajc9988 and rlk,

    My workloads are not ambiguous, but neither are they 'standard' to anything you or my competitors may like to think or imagine. To understand my workload is simple; I process lots of high-resolution images and transform them into something my clients want. Now, do you know my workloads? No, I didn't think so. But I'll repeat once again; slow high multicore platforms are the bs empty promises around here.

    I never stated that your workloads may or may not benefit from a high multicore platform, but I am stating that the majority of users won't benefit (and haven't for the last three years or so either). Pushing AMD or Intel high multicore platforms as future-proofing at the expense of the present day, real-world performance is kinda sad to me. Those kids don't know any better.

    Finding software that takes advantage of all those cores efficiently and better than a lower core platform is also a little disingenuous too. I would even say that is one more reason why PC sales have stagnated. If 'upgrading' gives you less performance at double the core count, why not wait for the 112 core platform to waste my $$$$$ on? :rolleyes:

    Do I not want actual productivity workloads to be more parallel? Of course, I do. That is the future. Still doesn't bring it here today.

    Let's talk about concrete facts, shall we? I've been saying a version of the above for almost a third of a decade now. What has happened in that time with regards to parallelism in workloads/software that we didn't have before then? This is the only question that needs to be answered. Everything I've been saying rests on that.

    Given that slow, glacial progress over the last two or three years (if we give time for AMD's products to be in the hands of the devs...), I really hope that the next equal time period shows exponential results.

    But coming back to what you buy today? You always buy the most powerful system you can for your current workloads. Predicting the tech future is a good way to go out of business.

    Here's my concrete example: when I joined notebookreview almost 10 years ago looking for info and real-world tests to decide whether SSD's were in my immediate future, I was told to just shut up and buy them. A few years later, I found their use case; OP'ing by almost 50% at the time. The proof that I offered then still wasn't enough to stop the ridicule from the naysayers. This is no different, but now, my 'bs ambiguous workload' is what is attacked, instead. o_O

    Instead of trying to build yourself's up by tearing me down, try answering the questions above and below that I pose to the whole forum too.

    What has transpired over the last few years that has made buying a high core count platform a requirement? And without wishing and speculating on the future, what (if there is any real reason) is buying one today being 'future-proof' to 2024?

    Because from where I stand, in 2024 I won't have a single tech item that I'm currently using today.
     
    Kyle likes this.
  2. custom90gt

    custom90gt Doc Mod Super Moderator

    Reputations:
    7,054
    Messages:
    3,091
    Likes Received:
    3,277
    Trophy Points:
    331
    I'm going to just step in real quick and remind everyone to be civil to each other.

    Having said that, in my opinion, the blanket statement of buy the fastest you can afford doesn't make sense to me. There are so many factors you have to weigh when purchasing a system that you can't make a simple statement like that. You have to look at budget, desired longevity, current/desired usage, cost of downtime, etc...

    When I was doing custom builds, I would take a ton of time trying to figure out what the customer actually needed. Sure I could have just thrown the fastest processor and what have you in there, but that would have been a disservice to the customer. There is no right answer for a build for everyone, that's why there are so many different parts out there.
     
  3. rlk

    rlk Notebook Evangelist

    Reputations:
    92
    Messages:
    451
    Likes Received:
    223
    Trophy Points:
    56
    Interesting. A priori, this workload sounds like it would be highly parallel, and assuming that the transform stage is independent of any other image (which assumption may well be false), that part should be embarrassingly parallel. "Embarrassingly parallel" is not an epithet; it's a term of art from the high performance technical computing community that means that different items to be processed are independent of each other, and so adding more parallelism up to the size of the problem results in arbitrary speedup.

    The key to optimizing any workflow of this nature is to minimize the amount of work that needs to be serialized, and in particular, the amount of time required by a human. That means making the human interaction as fast as possible, even if it means more back end processing. It's worth analyzing one's workflow carefully. Your actual processing steps might be very different, but it's worth doing the kind of analysis I'm outlining below.

    I have a schematically similar workflow in my avocation (sports photography for my alma mater) that I've put a lot of work into optimizing. I typically take about 2000 frames and keep 300-400 of them, which I upload (you can see this at https://rlk.smugmug.com/Sports). The steps amount to:

    1) Shoot the game.

    2) Offload the photos onto my system.

    3) Import the photos into my image management system (KPhotoAlbum).

    4) Review the photos and select the ones I want to keep.

    5) Crop and rotate the selected photos.

    6) Apply a watermark.

    7) Upload the photos.

    Step (1) of course is on the game time. Step (2) is sequential; it's limited by the I/O throughput (but if I had a fast enough card, it might be worth investigating parallelizing that, to achieve a deeper I/O queue depth.

    Step (3) is partially parallelized, helped by some coding I did to partially parallelize checksum computation and thumbnail generation (so there's some data parallelism and some control flow parallelism there) in addition to using an I/O scout thread to pre-read the images into memory. With a fast SSD, it would be worth increasing the number of scouts to improve queue depth, but I don't have an NVMe drive to tune that. More threads might allow greater parallelism of thumbnail and checksum computation if I had an NVMe. Between this and some other improvements, I'm basically I/O limited on a SATA SSD and am completely I/O bound to a hard drive.

    Step (4) is, of course, sequential, although KPhotoAlbum preloads images so I don't have to wait to skip to the next image. This is also human-intensive; KPhotoAlbum lets me tag images with a single key and use the space bar to move to the next image (being able to tag-and-next-image in one key stroke might have benefit). This steps is one of the two time-consuming steps, in this case because I have to review a lot of images.

    Step (5), the processing step, is partly on my time (decide on crop and rotation) and partly computation. There are two basic apps I can use for this, Darktable and RawTherapee (on Linux). I use RawTherapee because the crop workflow is faster; I can do it with one click-and-drag rather than having to position the mouse in the corner and do it in more steps. It's about 5 seconds faster per image because of that; with 300 images, that's not negligible! This is the other time consuming step, and I'd like to see what I can do to further optimize it.

    But actually applying the crop, rotate, and watermark (step 6) is something else. Neither Darktable nor RawTherapee efficiently parallelize image export. They can perform certain operations using multiple threads, but not multiple images simultaneously. So I wrote a script that extracts the crop and rotation from the sidecar files generated by RawTherapee and use ImageMagick to apply the crop. This part is parallelized; my script processes multiple images simultaneously. That saves about 10 minutes processing a typical game.

    Step (7), of course, is network bound.
     
  4. tilleroftheearth

    tilleroftheearth Wisdom listens quietly...

    Reputations:
    4,695
    Messages:
    12,157
    Likes Received:
    2,204
    Trophy Points:
    631
    You have a pretty good workflow for a single 'shooter', single user workflow.

    I make workloads as parallel as possible by using multiple 'shooters', dozens of workstations and multiple staff. ;)

    Multiple workstations are much more productive than a single monster workstation in my experience - especially when a workstation goes down (and they will and they do). 'Chunks' of each shoot are processed on multiple NAS and even more workstations and the entire job process stops when/if the 'perfect' required/contracted images are processed and recognized early on.

    Machines simply can't replace humans when selecting 'keeper' images, except for things like focus, etc. If they are used like that, sooner or later the images just all kinda look the same (this was tried and abandoned already). I doubt that this will change in my lifetime, or at least for my clients' needs.

    When I was also shooting not that long ago, I would capture up to 1K images per 10 minutes, continuously for hours. And I seldom shot alone. When the shoots were (known) to be shorter, time-wise, 2K+ images per 5 minutes was easily reached, per photographer.

    The less an image is retouched, the more life it has (yeah; even RAW images). After effects are more packaging and getting some special images print-worthy at the sizes requested, but seldom significantly slow the above process anymore. The cameras are used to create the 'feel' contracted. The software is only a safety net. ;)

     
  5. rlk

    rlk Notebook Evangelist

    Reputations:
    92
    Messages:
    451
    Likes Received:
    223
    Trophy Points:
    56
    Yep, that's certainly one way to do it :) It sounds like a highly tuned workflow for your needs, and that you're doing even less post than I am. I agree that parallelizing the photographers and having the fastest processor for a very simple workflow -- which likely does mean high clock rate and low core count -- makes perfect sense for what it sounds like you're doing.
     
  6. Talon

    Talon Notebook Virtuoso

    Reputations:
    1,140
    Messages:
    3,139
    Likes Received:
    3,589
    Trophy Points:
    331
    Intel Process Technology Update: 10nm Server Products in 1H 2020, Accelerated 7nm in 2021

    10NM Ice Lake shipping in June for laptops?

    https://www.anandtech.com/show/1431...r-products-in-1h-2020-accelerated-7nm-in-2021

    https://www.reddit.com/r/intel/comments/bmaslc/intel_confirms_10nm_to_be_released_this_year_10nm/

    I'll be holding onto my 9900K until 2021 me thinks. 10nm will be great an all, but 7nm should be a healthy jump for me. Looks like AMD will a short lead over Intel, but won't be nearly long enough to dramatically shift market share.
     
    Last edited: May 8, 2019
  7. Robbo99999

    Robbo99999 Notebook Prophet

    Reputations:
    3,878
    Messages:
    6,465
    Likes Received:
    5,511
    Trophy Points:
    681
    Could be an extra year, until 2022, that you might have to wait for consumer 7nm according to that article. I'm imagining that I'll be going with a 10nm 8core+ product before then, don't think my 6700K will hold out in usefulness to 2022 (but we'll see).
     
    Papusan, joluke, Talon and 1 other person like this.
  8. ajc9988

    ajc9988 Death by a thousand paper cuts

    Reputations:
    1,484
    Messages:
    5,577
    Likes Received:
    7,901
    Trophy Points:
    681
    They only mention 10nm server chips, which fits in with the Xeon E entry level, low core count Xeons previously disclosed.

    We already know comet lake is 14nm. Unless it's successor is 10nm, which there is little to no guidance on, and which would be a year after comet lake (read late 2020), then it would be late 2021, which may be the 7nm chips, which could slide to 2022.

    So, the only question is if Intel will skip directly to 7nm for mainstream desktop. Nothing they said discredits my prior analysis.

    Sent from my SM-G900P using Tapatalk
     
    hmscott likes this.
  9. Talon

    Talon Notebook Virtuoso

    Reputations:
    1,140
    Messages:
    3,139
    Likes Received:
    3,589
    Trophy Points:
    331
    Intel's latest road map shows 10nm this year/next month, 10nm+ in 2020, with 10nm++ and 7nm in 2021. Exciting times ahead for all. Somewhere in that mix we will see desktop chips obviously.
     
  10. hmscott

    hmscott Notebook Nobel Laureate

    Reputations:
    6,309
    Messages:
    19,551
    Likes Received:
    24,367
    Trophy Points:
    931
    More Intel Marketing BS and future unfulfilled promises for 10nm, and now 7nm production best wishes is all I see.

    As far as I have seen so far, the only Ice Lake 10nm CPU's are ULV Quad Core CPU's, nothing like a 6c/12t or 8c/16t model. Who wants a 4c CPU in this era of higher core count consumer CPU's?

    These are supposedly higher yield versions of the same 10nm process used for last years low production 10nm ULV CPU's that had disabled iGPU's, maybe this year the iGPU's will work? I doubt that 10nm process will match current 14nm IPC or performance at the same clocks, maybe it will match it?

    Does anyone see a desktop 10nm part on the charts? I didn't see any such listing. It wasn't obvious to me, and I don't think Intel is sure enough to even suggest a wish date for delivery for 10nm desktop or H level laptop CPU's.

    Based on the 10nm/7nm overlap with no 10nm desktop parts showing, and no 7nm desktop part showing, I still don't know what to think as far as Intel finally delivering any kind of useful 10nm / 7nm desktop / H laptop CPU's.

    To me it looks like a nicely filled out chart with 10nm / 7nm BS sprinkled in between the real 14nm production runs, in the same way as the last 3-4 years of missed deliveries for 10nm production promises.

    The only difference is that now Intel has added 7nm to their wish list.
     
    Last edited: May 9, 2019
Loading...

Share This Page