Ryzen vs i7 (Mainstream); Threadripper vs i9 (HEDT); X299 vs X399/TRX40; Xeon vs Epyc

Discussion in 'Hardware Components and Aftermarket Upgrades' started by ajc9988, Jun 7, 2017.

  1. ole!!!

    ole!!! Notebook Prophet

    Reputations:
    2,271
    Messages:
    5,745
    Likes Received:
    3,667
    Trophy Points:
    431
    i said zen2 is 6-7% IPC ahead. guess you can't read, goes to show you let your emotion gets to you. now shut it :)
     
    tilleroftheearth likes this.
  2. ajc9988

    ajc9988 Death by a thousand paper cuts

    Reputations:
    1,586
    Messages:
    5,796
    Likes Received:
    8,189
    Trophy Points:
    681
    And you are using the most conservative estimates of per generation IPC gains for AMD while assuming the best case scenario for Intel, including a possibly unrealistic frequency on the 10nm refinement.

    Reports already have closer to 16-23% IPC of Zen 3 over the current skylake based cores. Two more generations with larger/wider core designs is at minimum 14%, but more likely closer to 20-25%, with an outsized amount up to 32% IPC, added to the minimal 16% IPC of the Zen 3, and you would have closer to 36-41% IPC over that, with the outsize level of up to 48% or so. Also, even this isn't proper, because the year over year IPC changes are actually multiplied, similar to compound interest, meaning that the IPC increase over a current gen will be higher than what I put here as additive.

    You then assume Intel will hit 5GHz on 10nm+++, which I'd like to know the basis for that belief, because nothing suggests they will hit that. I'd say 4.4-4.6GHz would be more reasonable, and even at that, it will be a furnace. That places it at only a 5-10% frequency advantage, assuming AMD doesn't increase frequency with the newer nodes, which would likely be 5nm+ in 2022 against your theoretical 10nm+++ golden cove CPU which you assume to be around 30% more IPC over current core uarch.

    As I said, Intel will be competitive in 2022, likely not before. They will also likely be on chiplets by then, meaning this also does not fully account for the latency hits they will have in the transition.

    So you acting like AMD will barely be competitive or be crushed MAKES NO SENSE! It is what the match should have been this year if Intel had a working 10nm process, etc.

    Another important point is that at 5nm, the density used for HPC will be closer to 100 million transistors, whereas Intel's 10nm variant will be closer to the 60M range. If my estimate is right, AMD will be reducing the core die size to around 51.28mm^2, which is a nearly 33% reduction in size while adding transistors over the current 7nm process. That would give just over a 50% yield on the current defect density and increase the number of dies per wafer to 616 good dies on the current risk production levels. considering 15-18 months before using this node, 21 months at the farthest, AMD will be able to have a descent amount for market. Can the same be said of Intel considering their current "volume" yields are SO LOW that they introduced Ice Lake on mobile with about 1/3d the normal level of SKUs and even then they are on the market in very low quantities and the 14nm++++ chips outperform them. Same with next year's server chips with high core count. They will only get a tiny amount of usable chips, which is why Cooper Lake will be the high performance variant. If that is the case, on what basis do you think that in 2 years, Intel will solve both the yields AND the frequency issue to correct their issues with volume to market to alleviate the shortage (other than having mobile theoretically on a 7nm node, while trying to have both server and desktop on 10nm nodes)?

    Further, why are you assuming Intel will hit another 15-18% IPC increase for golden cove? Please enlighten me on that.

    Source article on TSMC 5nm yields used to estimate current yields for 5nm using this defect density:
    https://www.anandtech.com/show/15219/early-tsmc-5nm-test-chip-yields-80-hvm-coming-in-h1-2020
     
    Last edited: Dec 13, 2019
    hmscott likes this.
  3. rlk

    rlk Notebook Evangelist

    Reputations:
    127
    Messages:
    555
    Likes Received:
    291
    Trophy Points:
    76
    All right, let's be a little careful here.

    AMD's statements about Zen3 gains are "right in line with what you would expect from an entirely new architecture". That statement's vague, and it's qualified by noting that Zen2 was a bigger gain than would normally be expected from an evolutionary upgrade, but unless something else has come out that I didn't spot, 15-20% isn't entirely unreasonable to speculate on. Although this earlier piece was speculating 8% IPC uplift plus maybe 200 MHz frequency.

    Speculating beyond that to Zen4 and Zen5 is rather more of a stretch. AMD has certainly done a superb job backing up its claims for Zen2, but my comment above about Zen5 and Golden Cove applies here, as well. Squeezing out gains isn't going to get any easier as chips get faster and architectures are better tuned.

    Besides which, no one here is making purchasing decisions based on what Zen4 or Zen5 will be. That's just too far in the future to be very relevant. At least for mainstream, the chipsets will be different; the memory and I/O platforms will also be different, if not by Zen4, certainly by Zen5.

    My own plan this spring was to wait for Zen2 and get something between 8-16 cores depending upon where the prices wound up. I had to replace my i7-5820K and went with a 2700X, which changed my equation somewhat. There's not a lot of point for me to get a 3700X or 3800X now; it's certainly a very nice chip, but it's not that radically faster than the 2700X. The 3900X is still expensive and in very short supply. So for me the Zen3 question is somewhat relevant (since I presumably won't have to replace my motherboard). But it probably means I'm not going to upgrade my 2700X for a while unless the bottom really falls out of the market for the 3900X (or AMD releases an end user 3900 at a good discount from the 3900X). More likely I'll wait until the Ryzen 4000 chps are out and see what's actually going on, and then decide.
     
    ajc9988 likes this.
  4. ajc9988

    ajc9988 Death by a thousand paper cuts

    Reputations:
    1,586
    Messages:
    5,796
    Likes Received:
    8,189
    Trophy Points:
    681
    You misunderstand what I'm referencing for Zen 3. I'm relying on the reports of 10% integer IPC and 17% average integer+FP that has been leaked, which is from the same outlet that has the 13% IPC increase leak last fall, which was fairly accurate. I then added AMDs Zen 2 IPC advantage of 6-7% over coffee to the 10% integer leak to come to the 16-17% amount over coffee lake. The average is 7-8% IPC, so going with the 10% integer IPC is not unreasonable for the estimate of Zen 3 over Zen 2. It is the reference point that matters. And 10% is roughly in line generationally, matching the statement from AMD.

    (Edit: so the 16-23% IPC range was based on the leaked integer 10% and integer+FP 17% IPC, whereas the FP 32% with some up to 50% was excluded, which 6+17=23%)

    Beyond that, I'm basing it on my own assumptions on effects of where I believe AMD will go with their architecture. Going three years out is fully speculative as no one knows the uarch changes nor fully the yields or frequencies achieved on future process nodes.

    Instead, my point was to show he was casting AMD in a very conservative analysis on upcoming products while raining down an overly optimistic analysis on Intel's products considering Intel has had issues delivering to date.

    Edit:



    Edit 2: your October article references redgamingtech as the source. Here is redgamingtech from Dec. 6, 2019 giving the higher numbers I am using. They were the ones with the 13% IPC estimate for Zen 2 last fall, which was slightly lower than the final 15-17% IPC which materialized later. As such, I will estimate on their numbers again as I've no reason not to.

    On nvidia, rgt got it wrong for 20 series, but so did most of the tech press. I believe Nvidia uses subterfuge, feeding contradictory information down the line to see where it flushes out, along with instead of creating true leaks to generate hype, they use multiple variants on information so people are talking about which rumor is correct.
     
    Last edited: Dec 13, 2019
  5. ajc9988

    ajc9988 Death by a thousand paper cuts

    Reputations:
    1,586
    Messages:
    5,796
    Likes Received:
    8,189
    Trophy Points:
    681
    In addition to the update from last week, RGT has an analysis of IPC leaks from today as well.

     
  6. ole!!!

    ole!!! Notebook Prophet

    Reputations:
    2,271
    Messages:
    5,745
    Likes Received:
    3,667
    Trophy Points:
    431
    sounds like AMD gonna be in back in playing catch up game real soon if this keeps up. assuming the leak is correct, and I have high doubts it will be, zen 3 will be about 20% faster than CFL while running at a max 4.3ghz at best at 16 cores and coming in late 2020. i mean just icelake alone is ~15% higher, when golden cove and 10nm++ or +++ hits in late 2021 or early 2022 AMD will be behind.
     
  7. ajc9988

    ajc9988 Death by a thousand paper cuts

    Reputations:
    1,586
    Messages:
    5,796
    Likes Received:
    8,189
    Trophy Points:
    681
    Are you really that ignorant? Next year on desktop is another 14nm++++ sylake called comet lake. And there is another Intel vulnerability that the fix requires locking the voltage at stock voltage.

    That 10-core is going against a 16-core and will likely be between 4.8-5.0 GHz. If AMD at 4.3GHz is around 20-23% higher IPC, that would be AMD winning on outright performance AND having more cores. That is called a win.

    Ice lake is only on mobile chips and caps at 3.9GHz, which is lower than Zen 3 speeds and has a lower IPC. That is called losing.

    AMD will be updating to a new socket, get improved I/O, a new memory standard and interface (AMD has stated publicly lack of memory bandwidth is kneecapping their chips, so they are planning to increase that; this likely means increasing memory channels on server to 12-16 channels, HEDT to 8 channels (see TRX80), and potentially even quad channel for mainstream, if not adding HBM on package), plus a new generation with IPC uplift for Zen 4.

    Now, where is your data on golden cove IPC? Where is your data on 10nm++ being able to even get passed 4GHz, needles trying to suggest 5GHz? Even if you add the 17-18% IPC for sunny cove, which means AMD still would be up 2-5% IPC while running at equal to faster speeds than what we have seen on 10nm+, you then have another jump in IPC, which you assume is 7-8% even though AMD did 53%, 3-4%, 15-17%, and is on track for another 10-17% on Zen 3, depending on if talking pure integer or overall mixed workloads, all while having more ability to do greater changes like going much wider due to a 1.8x density increase going to 5nm, which allows for much higher IPC and who's yields on a pure die shrink would already yield as many dies per water as what they get on 7nm DUV.

    You assume Intel will fix the frequency and then give Intel another 18% IPC, which has no current basis and only then would they be competitive with a likely Zen 4, especially if 10nm+++ runs at 4.0-4.4GHz, all while coming out in 2022 after AMD has already confirmed zen 4 for 2021 and everyone already knows Intel is doing sunny or Willow cove on 14nm++++ for 2021, meaning your speculative Intel chip would need another Zen generation, zen 5, as the direct head to head, not Zen 4.

    Hmmm. Seems you likely need someone to map this out for you.
     
    jaybee83 and hmscott like this.
  8. ole!!!

    ole!!! Notebook Prophet

    Reputations:
    2,271
    Messages:
    5,745
    Likes Received:
    3,667
    Trophy Points:
    431
    im not talking about next yr, but in late 2021 and early 2022. I would definitely want a 5ghz+ CPU with higher IPC than a low frequency CPU with the same IPC.
     
  9. ajc9988

    ajc9988 Death by a thousand paper cuts

    Reputations:
    1,586
    Messages:
    5,796
    Likes Received:
    8,189
    Trophy Points:
    681
    Once again, where are you getting 5GHz from?
     
  10. hmscott

    hmscott Notebook Nobel Laureate

    Reputations:
    6,872
    Messages:
    20,231
    Likes Received:
    25,053
    Trophy Points:
    931
    5 years ago Intel was gonna deliver 10nm desktop / server CPU's in 2 years.

    3 years ago Intel was gonna deliver 10nm desktop / server CPU's in 2 years.

    Last Intel said we were gonna get 10nm desktops and server CPU's at the end of this year in 2019... and here's a few gimped 2 core 10nm ULV CPU's - oh, and Intel added that they have disabled the onboard IGPU in those gimped 10nm ULV CPU's.

    In the middle of this year 2019, just when we were eagerly expecting announcements for the 10nm desktop and server CPU shipments this year, Intel told us that they are now going to ship 10nm desktop and server CPU's at the end of 2021 or early 2022... in another 2 years, and here's a few gimped ULV 10nm 4 core CPU's.

    Intel hasn't been able to get 10nm to work well enough to replace 14nm on desktops and servers, and at the rate Intel are progressing with 10nm it might be another 10 years before they work up to a 16c 10nm desktop / server CPU.

    And, I wouldn't believe Intel on anything they are about to say about 7nm either.

    Intel predictions for new process deliver can't be believed because Intel has proven that they can't deliver the truth about where they are in the process.

    When Intel actually delivers 10nm / 7nm desktop / server CPU's we can buy, that's the time to spend time on Intel.

    Until then, AMD is putting on a great show, pull up a chair and enjoy. :)
     
Loading...

Share This Page