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. Papusan

    Papusan JOKEBOOKs Sucks! Dont waste your $$$ on FILTHY

    Reputations:
    31,256
    Messages:
    26,412
    Likes Received:
    48,535
    Trophy Points:
    931
    I haven’t AMD chips so I can’t say how much the OS will do for that chips. But faster ram should help push scores in right direction.
    F423AAF0-F2AD-4261-98C9-7B5609C7C062.png
     
    tilleroftheearth likes this.
  2. ajc9988

    ajc9988 Death by a thousand paper cuts

    Reputations:
    1,651
    Messages:
    5,938
    Likes Received:
    8,414
    Trophy Points:
    681
    Well now I'm informing you of that. GB3, GB4, CB15 are among the ones that, with Ryzen and TR, perform BETTER on Win 10 instead of Win 7. That's why my 4.25GHz 1950X is right behind Mr. Fox's 7960X in GB3 multi while he is at 5.2GHz.

    [​IMG]
    [​IMG]

    And for GB4:
    [​IMG]
    [​IMG]

    Now Intel will certainly still get ST top score in GB, but with an extra 15-35% (yes, some Epyc CPUs are seeing a lift as high as 35% on multi-threaded workloads gen to gen), what do you think is going to happen here, especially since in those workloads my heavy water cooling loop is keeping up with a CHILLED 7960X. Further, looking at the scores on Ryzen mainstream, the CB scores seem like TR will be doing very well against these 14nm Intel chips as well.

    Then, you tell people they could buy a 32-core chip for the cost of Intel's 16 and 18 core chips and dominate those workloads, the HEDT battle is over.

    Also, binning is there and is working. Just wish I could see what those chips could do with my excessive water loop!

    Edit: Here are my linux, win 10, and win 7 top scores submitted to geekbench with links:
    GB3
    Linux - 77106 https://browser.geekbench.com/geekbench3/8675016
    Win 10 - 73784 https://browser.geekbench.com/geekbench3/8793286 ( I have the higher score above because done in debug mode and didn't submit the saved file yet)
    Win 7 - 64687 https://browser.geekbench.com/geekbench3/8794784

    GB4
    Linux - 60289 https://browser.geekbench.com/v4/cpu/9715694
    Win 10 - 50973 (see above)
    Win 7 - 35874 https://browser.geekbench.com/v4/cpu/8726973

    There are quite a few benches that do better on Win 10 than Win 7 on Ryzen, plus there isn't the penalty for security patches that windows makes users install on windows 10 for Intel chips, meaning that they don't get dinged as hard, which may be why your spreadsheet has so much working better on Intel chips for Win 7.
     
    Last edited: Aug 16, 2019
    hmscott likes this.
  3. ajc9988

    ajc9988 Death by a thousand paper cuts

    Reputations:
    1,651
    Messages:
    5,938
    Likes Received:
    8,414
    Trophy Points:
    681
    On another note, people probably remember Intel investing $1B to expand the 14nm node fab capacity.

    Well, due to demand, TSMC is investing $6.5B more this year to expand capacity. They cited 5G tech and increasing capacity at 5nm (volume next year) as the reasons for the increased spending.
    https://www.digitimes.com/news/a20190814PD207.html
     
    TANWare and hmscott like this.
  4. ole!!!

    ole!!! Notebook Prophet

    Reputations:
    2,493
    Messages:
    5,827
    Likes Received:
    3,781
    Trophy Points:
    431
    at that sort of cancerous voltage with low frequency, we will need to wait a few years before jumping into AMD and AMD laptops.

    luckily zen 3 in a yr might see another decent improvement in both frequency & IPC and intel wont have anything "excellent" to be worthy of buying until maybe beginning of 2021 or even 2022. also recently found out ICL ipc is only around 7-8% over CFL and not the claimed 15-18% from intel, at least in CB15.

    this guy did a good job revealing intel's issue assuming most of it is true:
    I've been saying it for over a year, Intel's 10nm wound up as a tangled up experiment. In short, Intel management wanted such a massive shrink that would ensure their domination of the process market and attract customers to their fabs. The load combined with budget cuts and layoffs (and their other stunts) in the manufacturing group killed 10nm.
    In Long: There were three primary mistakes Intel made, Cobalt Metal Layers, Contact Over Active Gate (COAG), and plain ole hubris.

    In 2013, Intel was late. The Self Aligned Dual Patterning (SADP) required by the feature size of 14nm had a bad learning curve, yields were very bad at first, to the point where Broadwell was mostly a paper launch in 2014, two quarters behind schedule. This was not a critical problem and it was fixed gradually, such that Skylake was not delayed. Behind the scenes though, the long ramp time created a problem. As Intel has only a single (large) process development team, not leapfrogging teams, the 14nm delay led to a delay of 10nm. The specifications it would shoot for were not set in stone until 2014.

    Management gave them a difficult task. To win mobile they had to be power efficient and dense. To win desktop they needed to be fast. To win servers they needed excellent yields. And above all they needed to be better than the competition to attract new customers.
    In order to reach the goals set by management, the manufacturing group had to get creative. To that end a number of techniques never put into a production process before were adopted. COAG, SAQP, Cobalt, Ruthenium Liners, Tungsten contacts, single dummy gate, etc. This push is directly what led to the death of the process. Of those, only really COAG and Cobalt are causing the issues. I'll go into the specific problems next.

    If anyone is to blame, its the management, and their firing of the CEO with a ******** reason shows the board will not accept responsibility for the companies failings. They will not come clean in the foreseeable future. Their foundries are virtually dead after all the firings and cost cutting.

    The idea with Contact Over Active Gate is that instead of extending a gate such that it connects up with a contact to the side (thus using space on the side), the Contact stretches directly from the metal layer to the gate, rather than laying onto the substrate. This means there is NO room for error on manufacturing. The slightest misalignment leads to ****ed contacts. Thermal expansion, micro-vibrations from people walking nearby, changes in air pressure, imagine a cause, and it'll affect yields. I bet you the bloody position of the Moon can affect it. This kills the yields.

    To hit the targets Intel set, a minimum metal pitch of 36nm was selected. When you have Copper wires on a process they need to have a liner around them, this prevents diffusion, electromigration, and other nasty electrical fun. But this liner needs to be a certain thickness, so when the overall size of the wire gets smaller, the liner takes up a larger portion of it. Below 40nm it was thought that Cobalt would have superior electrical properties, despite it having a higher bulk electrical resistance. Its far more resistant to electromigration and needs a miniscule barrier to prevent it, while its resistance decreases at a slower rate as the wire size gets smaller.

    However, Intel overlooked two key problems: ductility/malleability, and thermal conductivity. Even at those tiny levels, Copper wires would be able to handle thermal expansion mechanical loads, bending and stretching ever so slightly as a processor made its rounds. And copper is Very good at transferring heat, letting the lower metal layers sink heat into the upper ones. Meanwhile Cobalt is hilariously brittle and has a sixth the thermal conductivity as Copper. On operation hot spots start to form, heat can't get away, brittle nature creates microfractures, and higher voltage to cross the fracture boundaries. Means the voltage/frequency curve is hilariously bad. This kills the performance and power usage.

    So where does it leave us at?

    10nm was meant to launch end of 2015, after 14nm this was pushed to 2016. It was Q3 2018 when i originally wrote much of what is outlined here, and the only 10nm chip at that time was a minuscule dual core made in a one-off batch of 100k units that took 6 months to assemble. Yields are sub 1%, the GPU doesn't function, and power usage was higher than 22nm.

    TSMC and at the time GloFlo were both ramping their 7nm processes, and while they're comparable to the 10nm in density, they actually work. They both are using SAQP for the transistors, but the choice of 40nm metal pitch allowed for SADP for that particular layer. GloFlo was using Copper with a Cobalt/Tantalum liner while TSMC is using straight Copper/Tantalum. Neither screwed with COAG or Single Dummy, but GloFlo was using Cobalt contacts.

    TSMC 7nm has been volume production, GloFlo meant enter it in a few months after TSMC initially did, but has since put it on hold indefinitely in favor of pursuing nodes beyond that. Consumer products on both of them meant to enter the market in mid-2019. TSMC did, latter didn't due to aforementioned reasons. Regardless, They're both going to outperform Intel's 10nm. Their manufacturing group failed at all levels. To make matters worse they didn't even bother backporting their improved cores to 14nm. Icelake has been inhouse since early 2017. The design is finalized, they just can't make the damn thing and didn't backport it. Management thought they couldn't fail.

    It is in my personal opinion that their 10nm process will never be financially viable. They were literally creating 10nm micro housefires that burned their own chips.
    Cannonlake is virtually dead. Ice Lake-U and Y (sub 15w mobile dual cores) came out as predicted still in mid-2019. I doubt Ice Lake-S (quads) will see anything beyond a paper launch ala Broadwell quads. Ice Lake-SP is a dead proposition, there's no way they're going to be able to make a server core based off that 10nm. Tiger Lake and Sapphire Rapids? Who the hell knows. Why worry about something that isn't going to happen.

    Intel's chance lies with 7nm, but people already predicted it won't arrive until 2021 at the earliest. After Cascade Lake is Cooper Lake; Kaby Lake-U is being replaced by Whiskey Lake, which'll be replaced by Comet Lake to run alongside the Ice Lake-U/Y dual cores. Amber Lake will replace Kaby Lake-Y until Ice Lake-Y. All those new code names are 14nm, and all are just tweaked Skylake.

    Also, The whole 14nm+++++ tripe is meaningless and has been for years. They're standard PDK updates. Everyone in the industry does them. It wasn't until recently that they started being branded, in this case by Intel.
     
  5. ajc9988

    ajc9988 Death by a thousand paper cuts

    Reputations:
    1,651
    Messages:
    5,938
    Likes Received:
    8,414
    Trophy Points:
    681
    So, since the TR 32-core GB4 score was leaked, we now get to see the Intel 18-core Cascade-X GB4 score, which should be priced to compete against each other:
    [​IMG]

    https://www.tomshardware.co.uk/intel-18-core-cascade-lake-x-cpu,news-61453.html
    It is "3.36% [faster] in single-core workloads and around 7.4% [faster] in multi-core workloads" compared to the 9980XE.

    Compare this to the 32-core below, priced the same, approximately:

    [​IMG]
    "The Castle Peak chip performs up to 4.72% and 14.63% faster than the Threadripper 2990WX in single-core and multi-core workloads, respectively." This is likely while also not seeing the prior gen's performance regression due to using the I/O chip.

    Looks like faster single thread at stock, plus a massive 74% faster at multi-threaded workloads. If I was Shania Twain, I'd have to say of Intel "that don't impress me much."

    Overclocking is yet to be seen on the Cascade-X chips, but the Skylake-X originally, for the 7 series chips, used 14nm+ (Kaby) process. I do not know if the refresh (9 series) used 14nm+ or ++ variant. If not ++, there might be a little more in the OC tank, but otherwise, not really.

    Meanwhile, what was pointed out by a person in another forum:
    "Looks like the first sample was running all 64 cores between 4150 - 4200 Mhz. The second sample logged clocks between 4150 - 4350 Mhz.
    https://browser.geekbench.com/v4/cpu/14281644.gb4
    https://browser.geekbench.com/v4/cpu/14281648.gb4 "
     
    jaybee83 and hmscott like this.
  6. TANWare

    TANWare Just This Side of Senile, I think. Super Moderator

    Reputations:
    2,547
    Messages:
    9,583
    Likes Received:
    4,994
    Trophy Points:
    431
    If all used Linux as the OS I would be more impressed by the charts. I did use the browser and compared the all Linux scores and the 32 core AMD easily beat out even Intel's 28 core in geekbench.
     
  7. Papusan

    Papusan JOKEBOOKs Sucks! Dont waste your $$$ on FILTHY

    Reputations:
    31,256
    Messages:
    26,412
    Likes Received:
    48,535
    Trophy Points:
    931
    Intel Tiger Lake-U 4 Core, 8 Thread Processors With Gen 12, Xe GPUs Spotted – Faster Than Core i7-8700K In Single-Core Tests at Just 3.6 GHz Clock Wccftech.com
    Now, two Tiger Lake CPU entries have appeared on the user benchmark database along with early performance numbers. Both CPUs are part of the Tiger Lake-U family which includes 15-28W SKUs. Both processors are configured as a 4 core and 8 thread design, featuring clock speeds of 1.2 GHz base and 3.6 GHz boost. Since the launch is a year away, these clock speeds don’t look final & may end up being much higher in the final retail variants.
    upload_2019-8-19_16-43-33.png
     
    tilleroftheearth and ajc9988 like this.
  8. jaybee83

    jaybee83 Biotech-Doc

    Reputations:
    4,105
    Messages:
    11,564
    Likes Received:
    9,127
    Trophy Points:
    931
    thats +30% IPC over CFL-R and still 23% over ryzen 3000

    Sent from my Xiaomi Mi Max 2 (Oxygen) using Tapatalk
     
    ajc9988 likes this.
  9. ajc9988

    ajc9988 Death by a thousand paper cuts

    Reputations:
    1,651
    Messages:
    5,938
    Likes Received:
    8,414
    Trophy Points:
    681
    How do you get 30%? On quad core, I get around 26.6%. [Edit: nevermind, whereas AMD gets a higher percent jump with multicore, Intel has a higher percent on single core.]

    Also, on pure performance on quad core score, it is still lagging Zen 2, while this uarch is made to go against Zen 3, which is already finished and ready for tapeout. The 3700U and 3750H are Zen+ while the mobile chips it will go against are Zen 2 cores, which is why I selected the 4-core score and looked at the 3900X. If the mobile parts next year can reach anywhere near the 4.3GHz range like most Zen 2 chips do, then that is a solid measure for comparison.

    Also, Ice Lake -SP comes in Q2 2020. Tiger Lake - U comes around Q2-Q3 currently, although it may be pushed out later. Desktop gets 14nm Comet Lake-S in Q4 to Q1 2020 and Rocket Lake-S comes around a year later for desktop, also on 14nm. That means, since these numbers are around what current Zen 2 gets, even with a speed boost, it will have a hard time against Zen 3.

    Kind of sad to see the peak.

    Edit: @Papusan - didn't you say it was pathetic AMD couldn't hit 5GHz, and that the speeds were bad. Yet now you herald even slower speeds as magnificent. Why the change in position?
     
    Last edited: Aug 19, 2019
  10. Papusan

    Papusan JOKEBOOKs Sucks! Dont waste your $$$ on FILTHY

    Reputations:
    31,256
    Messages:
    26,412
    Likes Received:
    48,535
    Trophy Points:
    931
    I still mean low clocks is pathetic (even more pathetic if all cores can't be put equal max single core boost). Where did I herald my own meaning that even slower speeds as magnificent in mentioned post #2927? I just post info found from the web. I have two choices, bruh. Either not post... Or post:rolleyes:
     
Loading...

Share This Page