Results 1 to 13 of 13

Thread: Asymmetric 3D cache on high end 7000 series

  1. #1
    root Member DanceswithUnix's Avatar
    Join Date
    Jan 2006
    Location
    In the middle of a core dump
    Posts
    12,978
    Thanks
    778
    Thanked
    1,586 times in 1,341 posts
    • DanceswithUnix's system
      • Motherboard:
      • Asus X470-PRO
      • CPU:
      • 5900X
      • Memory:
      • 32GB 3200MHz ECC
      • Storage:
      • 2TB Linux, 2TB Games (Win 10)
      • Graphics card(s):
      • Asus Strix RX Vega 56
      • PSU:
      • 650W Corsair TX
      • Case:
      • Antec 300
      • Operating System:
      • Fedora 39 + Win 10 Pro 64 (yuk)
      • Monitor(s):
      • Benq XL2730Z 1440p + Iiyama 27" 1440p
      • Internet:
      • Zen 900Mb/900Mb (CityFibre FttP)

    Asymmetric 3D cache on high end 7000 series

    Interesting confirmation here:

    https://www.techpowerup.com/303084/a...e-two-chiplets

    If you can afford a 7900X-3D or better, then you only get the stacked cache on one of the CCD die.

    That's potentially cool, as the non stacked die can boost to the same sort of single thread frequency as the non 3D variant. Or it's a nightmare.

    I guess thread scheduling just got more interesting

  2. #2
    Senior Member
    Join Date
    Jul 2009
    Location
    West Sussex
    Posts
    1,721
    Thanks
    197
    Thanked
    243 times in 223 posts
    • kompukare's system
      • Motherboard:
      • Asus P8Z77-V LX
      • CPU:
      • Intel i5-3570K
      • Memory:
      • 4 x 8GB DDR3
      • Storage:
      • Samsung 850 EVo 500GB | Corsair MP510 960GB | 2 x WD 4TB spinners
      • Graphics card(s):
      • Sappihre R7 260X 1GB (sic)
      • PSU:
      • Antec 650 Gold TruePower (Seasonic)
      • Case:
      • Aerocool DS 200 (silenced, 53.6 litres)l)
      • Operating System:
      • Windows 10-64
      • Monitor(s):
      • 2 x ViewSonic 27" 1440p

    Re: Asymmetric 3D cache on high end 7000 series

    Quote Originally Posted by DanceswithUnix View Post
    Interesting confirmation here:

    https://www.techpowerup.com/303084/a...e-two-chiplets

    If you can afford a 7900X-3D or better, then you only get the stacked cache on one of the CCD die.

    That's potentially cool, as the non stacked die can boost to the same sort of single thread frequency as the non 3D variant. Or it's a nightmare.

    I guess thread scheduling just got more interesting
    The kernel thread schedulers have been living in interesting times for a while now. For x86, first it was Intel and their E cores, not this. Plus the various CCD stuff AMD did. Then for mobiles (with which Linux has to cope due to Android), this has been going for a while.

    The other obvious takeaway: with 5800X3D having great DDR4 and timings was nice but not essential. 5800X3D was basically very easy to get great performance from without expensive RAM or having to configure sub-timings.

    For Zen4 3D this is probably no longer the case. The 3D CCD should be fine, but if anything needs to use the "ordinary" CCD then because of that, you'd want fancy expensive RAM with good timings.

  3. Received thanks from:

    CAT-THE-FIFTH (13-01-2023)

  4. #3
    Evil Monkey! MrJim's Avatar
    Join Date
    Aug 2012
    Location
    London
    Posts
    2,313
    Thanks
    300
    Thanked
    472 times in 362 posts
    • MrJim's system
      • Motherboard:
      • MSI Tomahawk X570
      • CPU:
      • AMD Ryzen 5900X
      • Memory:
      • 32gb Kingston 3600 DDR4
      • Storage:
      • Aorus 1Tb NVME SSD, Samsung 1Tb 970 Evo SSD, Crucial 2tb MX500 SSD, Seagate Ironwolf 4Tb SSD
      • Graphics card(s):
      • EVGA 3080Ti
      • PSU:
      • Seasonic Prime Ultra Platinum 1300W
      • Case:
      • Fractal Meshify 2
      • Operating System:
      • Windows 11 Pro
      • Monitor(s):
      • Viewsonic 27" XG2703-GS
      • Internet:
      • BT 900 mb/s FTTP

    Re: Asymmetric 3D cache on high end 7000 series

    One commentator mentioned that most games (so far) haven't really made use of more than eight cores, which is what the vcache is aimed at. It will be interesting to see if any games fail to use the CCD with the vcache, which as you say, will be down to the scheduler.

  5. #4
    root Member DanceswithUnix's Avatar
    Join Date
    Jan 2006
    Location
    In the middle of a core dump
    Posts
    12,978
    Thanks
    778
    Thanked
    1,586 times in 1,341 posts
    • DanceswithUnix's system
      • Motherboard:
      • Asus X470-PRO
      • CPU:
      • 5900X
      • Memory:
      • 32GB 3200MHz ECC
      • Storage:
      • 2TB Linux, 2TB Games (Win 10)
      • Graphics card(s):
      • Asus Strix RX Vega 56
      • PSU:
      • 650W Corsair TX
      • Case:
      • Antec 300
      • Operating System:
      • Fedora 39 + Win 10 Pro 64 (yuk)
      • Monitor(s):
      • Benq XL2730Z 1440p + Iiyama 27" 1440p
      • Internet:
      • Zen 900Mb/900Mb (CityFibre FttP)

    Re: Asymmetric 3D cache on high end 7000 series

    Quote Originally Posted by kompukare View Post
    The kernel thread schedulers have been living in interesting times for a while now. For x86, first it was Intel and their E cores, not this. Plus the various CCD stuff AMD did. Then for mobiles (with which Linux has to cope due to Android), this has been going for a while.

    The other obvious takeaway: with 5800X3D having great DDR4 and timings was nice but not essential. 5800X3D was basically very easy to get great performance from without expensive RAM or having to configure sub-timings.

    For Zen4 3D this is probably no longer the case. The 3D CCD should be fine, but if anything needs to use the "ordinary" CCD then because of that, you'd want fancy expensive RAM with good timings.
    That will again be an interesting point. The cores on the plain CCD should still have access to the 3D L3 cache, just with a bit more latency.

  6. #5
    Moosing about! CAT-THE-FIFTH's Avatar
    Join Date
    Aug 2006
    Location
    Not here
    Posts
    32,039
    Thanks
    3,910
    Thanked
    5,224 times in 4,015 posts
    • CAT-THE-FIFTH's system
      • Motherboard:
      • Less E-PEEN
      • CPU:
      • Massive E-PEEN
      • Memory:
      • RGB E-PEEN
      • Storage:
      • Not in any order
      • Graphics card(s):
      • EVEN BIGGER E-PEEN
      • PSU:
      • OVERSIZED
      • Case:
      • UNDERSIZED
      • Operating System:
      • DOS 6.22
      • Monitor(s):
      • NOT USUALLY ON....WHEN I POST
      • Internet:
      • FUNCTIONAL

    Re: Asymmetric 3D cache on high end 7000 series

    This is going to play havoc with direct die heatpipe coolers! Also sounds like AMD trying to save money by only using one die with the stacked cache. It wouldn't surprise me the next generation will have both with the stacked cache.

  7. #6
    Senior Member
    Join Date
    Mar 2009
    Location
    Manchester
    Posts
    1,669
    Thanks
    52
    Thanked
    123 times in 100 posts
    • Percy1983's system
      • Motherboard:
      • Gigabyte x570 Aorus Pro
      • CPU:
      • AMD 5900x + Cooler Master Nepton 140XL
      • Memory:
      • 64GB (4x16GB ) Corsair Vengence 3200mhz @ 3600mhz CL16
      • Storage:
      • 1tb SP US75 Boot + Fast 4tb SP XS70 + Slow Raid 0 4tb (2tbx2) with 100gb NVME cache
      • Graphics card(s):
      • Vega 56 8GB
      • PSU:
      • 875w Thermaltake Toughpower XT
      • Case:
      • Thermaltake Level 10 GT Snow Edition
      • Operating System:
      • Windows 11 Pro 64bit
      • Monitor(s):
      • 24" Acer UHD x2
      • Internet:
      • Vodafone

    Re: Asymmetric 3D cache on high end 7000 series

    Sounds interesting, can't be any worse for scheduling than intels P and E cores at a guess.

    I am probably skipping this generation and the one after anyway so hopefully it will all be sorted for my next upgrade.

  8. #7
    Banhammer in peace PeterB kalniel's Avatar
    Join Date
    Aug 2005
    Posts
    31,023
    Thanks
    1,870
    Thanked
    3,381 times in 2,718 posts
    • kalniel's system
      • Motherboard:
      • Gigabyte Z390 Aorus Ultra
      • CPU:
      • Intel i9 9900k
      • Memory:
      • 32GB DDR4 3200 CL16
      • Storage:
      • 1TB Samsung 970Evo+ NVMe
      • Graphics card(s):
      • nVidia GTX 1060 6GB
      • PSU:
      • Seasonic 600W
      • Case:
      • Cooler Master HAF 912
      • Operating System:
      • Win 10 Pro x64
      • Monitor(s):
      • Dell S2721DGF
      • Internet:
      • rubbish

    Re: Asymmetric 3D cache on high end 7000 series

    Quote Originally Posted by Percy1983 View Post
    Sounds interesting, can't be any worse for scheduling than intels P and E cores at a guess.
    I think it'll be a little bit tougher, since it's pretty obvious when you would schedule between a P and an E core, but now you've/the scheduler has to choose between higher frequency or more cache, both of which are valid performance gains.

  9. #8
    Senior Member
    Join Date
    Jul 2003
    Posts
    12,113
    Thanks
    906
    Thanked
    580 times in 405 posts

    Re: Asymmetric 3D cache on high end 7000 series

    Quote Originally Posted by Percy1983 View Post
    Sounds interesting, can't be any worse for scheduling than intels P and E cores at a guess.
    Bit OT, but how do the P/E cores seem to effect things in Windows, I've not touched any of the newer Intel stuff for a while, but have just had an AIO with a 12th Gen i7 in it, need to test it to see how much it screws stuff up.

  10. #9
    Evil Monkey! MrJim's Avatar
    Join Date
    Aug 2012
    Location
    London
    Posts
    2,313
    Thanks
    300
    Thanked
    472 times in 362 posts
    • MrJim's system
      • Motherboard:
      • MSI Tomahawk X570
      • CPU:
      • AMD Ryzen 5900X
      • Memory:
      • 32gb Kingston 3600 DDR4
      • Storage:
      • Aorus 1Tb NVME SSD, Samsung 1Tb 970 Evo SSD, Crucial 2tb MX500 SSD, Seagate Ironwolf 4Tb SSD
      • Graphics card(s):
      • EVGA 3080Ti
      • PSU:
      • Seasonic Prime Ultra Platinum 1300W
      • Case:
      • Fractal Meshify 2
      • Operating System:
      • Windows 11 Pro
      • Monitor(s):
      • Viewsonic 27" XG2703-GS
      • Internet:
      • BT 900 mb/s FTTP

    Re: Asymmetric 3D cache on high end 7000 series

    Quote Originally Posted by [GSV]Trig View Post
    Bit OT, but how do the P/E cores seem to effect things in Windows, I've not touched any of the newer Intel stuff for a while, but have just had an AIO with a 12th Gen i7 in it, need to test it to see how much it screws stuff up.
    I think you need to be using Windows 11 in order for the scheduler to work correctly with the P & E cores, although I don't have an Intel cpu so can't verify that. You can always turn off the the E cores in the bios, as one way of preventing problems (if a bit extreme).

  11. #10
    root Member DanceswithUnix's Avatar
    Join Date
    Jan 2006
    Location
    In the middle of a core dump
    Posts
    12,978
    Thanks
    778
    Thanked
    1,586 times in 1,341 posts
    • DanceswithUnix's system
      • Motherboard:
      • Asus X470-PRO
      • CPU:
      • 5900X
      • Memory:
      • 32GB 3200MHz ECC
      • Storage:
      • 2TB Linux, 2TB Games (Win 10)
      • Graphics card(s):
      • Asus Strix RX Vega 56
      • PSU:
      • 650W Corsair TX
      • Case:
      • Antec 300
      • Operating System:
      • Fedora 39 + Win 10 Pro 64 (yuk)
      • Monitor(s):
      • Benq XL2730Z 1440p + Iiyama 27" 1440p
      • Internet:
      • Zen 900Mb/900Mb (CityFibre FttP)

    Re: Asymmetric 3D cache on high end 7000 series

    Quote Originally Posted by CAT-THE-FIFTH View Post
    This is going to play havoc with direct die heatpipe coolers! Also sounds like AMD trying to save money by only using one die with the stacked cache. It wouldn't surprise me the next generation will have both with the stacked cache.
    What problem do you see with coolers? In a normal 7950X you can have one of the compute die generating more heat than the other, and the IO die doing its own thing. Just seems more of the same to me.


    Quote Originally Posted by kalniel View Post
    I think it'll be a little bit tougher, since it's pretty obvious when you would schedule between a P and an E core, but now you've/the scheduler has to choose between higher frequency or more cache, both of which are valid performance gains.
    I'm sort of expecting a whitelist tbh.

    Old games will only have a few threads, don't need the 3D cache and will benefit from increased frequency. They will already be fast enough, but using the high freq cores will make benchmarks look better.
    Modern stuff will more likely use more cores, need more cache, and so will lean on the 3D cache cores.
    Productivity programs will just max out all the cores, so are moot.

    Besides, if the scheduler gets it wrong then the fast clocked core will still see the cache on the 3D die like an L4 so should still go faster than a straight 7950X.

  12. #11
    Senior Member
    Join Date
    Mar 2005
    Posts
    4,932
    Thanks
    171
    Thanked
    383 times in 310 posts
    • badass's system
      • Motherboard:
      • ASUS P8Z77-m pro
      • CPU:
      • Core i5 3570K
      • Memory:
      • 32GB
      • Storage:
      • 1TB Samsung 850 EVO, 2TB WD Green
      • Graphics card(s):
      • Radeon RX 580
      • PSU:
      • Corsair HX520W
      • Case:
      • Silverstone SG02-F
      • Operating System:
      • Windows 10 X64
      • Monitor(s):
      • Del U2311, LG226WTQ
      • Internet:
      • 80/20 FTTC

    Re: Asymmetric 3D cache on high end 7000 series

    Quote Originally Posted by DanceswithUnix View Post
    Besides, if the scheduler gets it wrong then the fast clocked core will still see the cache on the 3D die like an L4 so should still go faster than a straight 7950X.
    Is the L3 shared between CCX's? I thought the L3 was "Per CCX" - i.e. the other compute die could not use it.
    "In a perfect world... spammers would get caught, go to jail, and share a cell with many men who have enlarged their penises, taken Viagra and are looking for a new relationship."

  13. #12
    Moosing about! CAT-THE-FIFTH's Avatar
    Join Date
    Aug 2006
    Location
    Not here
    Posts
    32,039
    Thanks
    3,910
    Thanked
    5,224 times in 4,015 posts
    • CAT-THE-FIFTH's system
      • Motherboard:
      • Less E-PEEN
      • CPU:
      • Massive E-PEEN
      • Memory:
      • RGB E-PEEN
      • Storage:
      • Not in any order
      • Graphics card(s):
      • EVEN BIGGER E-PEEN
      • PSU:
      • OVERSIZED
      • Case:
      • UNDERSIZED
      • Operating System:
      • DOS 6.22
      • Monitor(s):
      • NOT USUALLY ON....WHEN I POST
      • Internet:
      • FUNCTIONAL

    Re: Asymmetric 3D cache on high end 7000 series

    Quote Originally Posted by DanceswithUnix View Post
    What problem do you see with coolers? In a normal 7950X you can have one of the compute die generating more heat than the other, and the IO die doing its own thing. Just seems more of the same to me.
    Because it's an asymmetrical orientation. This is going to cause problems with direct die air coolers for sure.

  14. #13
    root Member DanceswithUnix's Avatar
    Join Date
    Jan 2006
    Location
    In the middle of a core dump
    Posts
    12,978
    Thanks
    778
    Thanked
    1,586 times in 1,341 posts
    • DanceswithUnix's system
      • Motherboard:
      • Asus X470-PRO
      • CPU:
      • 5900X
      • Memory:
      • 32GB 3200MHz ECC
      • Storage:
      • 2TB Linux, 2TB Games (Win 10)
      • Graphics card(s):
      • Asus Strix RX Vega 56
      • PSU:
      • 650W Corsair TX
      • Case:
      • Antec 300
      • Operating System:
      • Fedora 39 + Win 10 Pro 64 (yuk)
      • Monitor(s):
      • Benq XL2730Z 1440p + Iiyama 27" 1440p
      • Internet:
      • Zen 900Mb/900Mb (CityFibre FttP)

    Re: Asymmetric 3D cache on high end 7000 series

    Quote Originally Posted by badass View Post
    Is the L3 shared between CCX's? I thought the L3 was "Per CCX" - i.e. the other compute die could not use it.
    That had me wondering if I made a bad assumption somewhere, and I think you might be right (though I've had a hard time finding definitive references).

    The L3 seems to be wired such that cores can see all the L3 through the Infinity Fabric, and there seem to be plenty of "cross core latency" sort of benchmarks. That makes sense for efficient sharing of data across threads. *But*, it's a victim cache so only the cores local to the L3 write to it.

    I did find an article that said the big cache cores would be scheduled first, with the normal 8 cores kicking in when there are enough threads for things like Blender. That might be rubbish though, or for that matter that's software so it could change

Thread Information

Users Browsing this Thread

There are currently 1 users browsing this thread. (0 members and 1 guests)

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •