Real benefits of HBM?

Discussion in 'Videocards - AMD Radeon' started by Screwdriver, Aug 28, 2015.

Thread Status:
Not open for further replies.
  1. Screwdriver

    Screwdriver Guest

    Messages:
    120
    Likes Received:
    0
    GPU:
    Sapphire R9 Fury Tri-X
    Did AMD ever test Fuji with DDR5 and then saw a performance benefit with HBM? I mean I know the advantages as told by AMD, but just curious if the speed benefit was real or was it propaganda to get them towards HBM2 and to also say they were the first.
     
  2. Barry J

    Barry J Ancient Guru

    Messages:
    2,803
    Likes Received:
    152
    GPU:
    RTX2080 TRIO Super
    I would imagine a little of both
     
  3. ---TK---

    ---TK--- Guest

    Messages:
    22,104
    Likes Received:
    3
    GPU:
    2x 980Ti Gaming 1430/7296
    Gddr5 memory bandwidth on a 980ti or the 390 390x retreads have mode than enough bandwidth. I don't see any benefit other than power consumption until gddr5 becomes a bottleneck. That won't happen before hbm though.
     
  4. Hootmon

    Hootmon Guest

    Messages:
    1,231
    Likes Received:
    6
    GPU:
    XFX THICC III Ultra
    Yeah...

    DDR5 is beginning to approach a 'diminishing returns' zone with its narrower/faster design. Its probably good for a few more iterations before something else will be required.

    A fatter/slower approach will allow more data to be pushed in comparison as going forward.
     

  5. Screwdriver

    Screwdriver Guest

    Messages:
    120
    Likes Received:
    0
    GPU:
    Sapphire R9 Fury Tri-X
    More than enough, but maybe with HBM they might of been faster?
     
  6. Loophole35

    Loophole35 Guest

    Messages:
    9,797
    Likes Received:
    1,161
    GPU:
    EVGA 1080ti SC
    Most of the improvement is in the space saving of having the RAM stacks on the imposer and the power efficiency. We really didn't have a bandwidth issue.
     
  7. ---TK---

    ---TK--- Guest

    Messages:
    22,104
    Likes Received:
    3
    GPU:
    2x 980Ti Gaming 1430/7296
    Overclocking vram on a 384 bit bus 980ti yields little gains, that why I just run it slightly oc. 512 bit 390 390x has more than enough bandwidth as well.
     
  8. Truder

    Truder Ancient Guru

    Messages:
    2,388
    Likes Received:
    1,423
    GPU:
    RX 6700XT Nitro+
    I think it'll be more a case of in terms of trickle down technology is where we'll see the benefits.

    If we look at the mid-range, memory bandwidth is a bottleneck, when ocing the memory does yield better results, but as stated in the highend, memory hasn't been an issue so if we imagine hbm in the midrange segment, the only limitation would be capacity and core configuration/clock.

    This has always been the case with hardware design, the best on the highend and eventually a year later we see the highend technology trickle down to mainstream.
     
  9. PrMinisterGR

    PrMinisterGR Ancient Guru

    Messages:
    8,125
    Likes Received:
    969
    GPU:
    Inno3D RTX 3090
    There never has been "enough" bandwidth for graphics really. And apparently the Fury X seems to scale very well with HBM overclocking. Look at this. He basically got a 20% performance increase out of a 20% increase in memory speed and a 9% performance increase in core speed. It strangely correlates with the findings from this little test/study.

    TL;DR: It matters quite a lot. It would be interesting to see numbers with half the HBM speed, although I bet that HBM also has better timings than GDDR5, since it actually runs in much lower frequencies.
     
  10. ---TK---

    ---TK--- Guest

    Messages:
    22,104
    Likes Received:
    3
    GPU:
    2x 980Ti Gaming 1430/7296
    High end cards have traditionally had plenty of bandwidth and owned many. The only definitive high end cards I have had where overclocking the vram on 680's but that was due to being 256 bit bus. 6ghz oc to 7.2 netted good gains. 580 sli, 780ti sli 980ti sli all had enough bandwidth at stock because they were 384 bit. Overclocking vram on them netted close to zilch. I would rather have 6gb of fast vram 365gb/s than 4 GB of hbm1.
     

  11. Loophole35

    Loophole35 Guest

    Messages:
    9,797
    Likes Received:
    1,161
    GPU:
    EVGA 1080ti SC
    You do know the moar VRAM argument only worked when AMD had the most on their high end.


    Honestly the Fiji GPU is a nice card had it come out last year just after the 980/970 it would have been awesome. It didn't. It's late to the party which sucks because it takes the wind out of the sails for HBM.
     
  12. PrMinisterGR

    PrMinisterGR Ancient Guru

    Messages:
    8,125
    Likes Received:
    969
    GPU:
    Inno3D RTX 3090
    The numbers correlate that you do get extra speed for the extra bandwidth though. It will never be as much as for the core, but it is far from insignificant, and I bet that for cards so ridiculously parallel as the Fiji ones it matters even more. The 3DMark results of the first I posted kinda show that actually.
     
  13. Loophole35

    Loophole35 Guest

    Messages:
    9,797
    Likes Received:
    1,161
    GPU:
    EVGA 1080ti SC
    This may be something new we have to learn on HBM as going over 350Gb/s on GDDR5 never really netted much. But this being a new memory configuration the way we understand VRAM bandwidth may be different.
     
  14. Hootmon

    Hootmon Guest

    Messages:
    1,231
    Likes Received:
    6
    GPU:
    XFX THICC III Ultra
    +1

    Bus Width and Clock Speeds matter, but so does Latency. HBM shuffles those variables in a new way.
     
  15. thesmokingman

    thesmokingman Guest

    Messages:
    36
    Likes Received:
    0
    GPU:
    XFX 6950
    Compare what happens with Nv cards in teh VR space vs the Fury. Also, the advantages of HBM are still untapped. It will take time for developers to capitalize on the massive bandwidth. And that's not all because there's still the huge drop in latency with HBM is driving the foundation for VR. Google it. Nv cards cannot handle VR because their latencies are too great, it makes users nauseous. Btw, propaganda?? :3eyes:


    You realize it is the new standard and everyone is on board right? Nvidia, Samsung, Intel besides AMD and Hynix. There is a major declining slope in regards to power consumption with adding massive amounts of vram like with the Titan X. This is especially prevalent in the high performance computing space where the major players have been tackling the issue with variations on 3D stacking for years.
     
    Last edited: Aug 28, 2015

  16. eclap

    eclap Banned

    Messages:
    31,468
    Likes Received:
    4
    GPU:
    Palit GR 1080 2000/11000
    Nope

    384bit bus on the 7950 was again plenty, overclocking the vram yielded next to zero performance.
     
  17. Agonist

    Agonist Ancient Guru

    Messages:
    4,283
    Likes Received:
    1,312
    GPU:
    XFX 7900xtx Black
    Exactly, I left my vram stock. Core clock was a whole different story.
    I had my 7950 crossfire @ 1275 core.
    The 384bit bus was more then enough with 3GB for eyefinity also.
    They were monsters when oc'ed. Better then my 670s I have now.
     
  18. ---TK---

    ---TK--- Guest

    Messages:
    22,104
    Likes Received:
    3
    GPU:
    2x 980Ti Gaming 1430/7296
    Yeah now lowly gdd5 that I got running at 365.4 GB/s is a bottleneck, funny stuff. 512 bit 290, 290x, 390, 390x running at 5ghz for the 200 series and 6ghz for the 300 series. hbm1 would of been better if it were capable of more than 4gb, probably why nvidia is going right to hbm2.
     
  19. Screwdriver

    Screwdriver Guest

    Messages:
    120
    Likes Received:
    0
    GPU:
    Sapphire R9 Fury Tri-X
    4gb has not been show to be a bottleneck.....I am sure when it is, even the mighty 980 ti will be yesterdays news.
     
  20. ---TK---

    ---TK--- Guest

    Messages:
    22,104
    Likes Received:
    3
    GPU:
    2x 980Ti Gaming 1430/7296
    Pretty sure a faster card with more vram will be relevant longer than a slower card and less vram.
     
Thread Status:
Not open for further replies.

Share This Page