AMD Graphics Inside Story

Discussion in '[H]ard|OCP Front Page News' started by rgMekanic, Jun 12, 2018.

  1. rgMekanic

    rgMekanic [H]ard|News Staff Member

    Messages:
    5,069
    Joined:
    May 13, 2013
    WCCFTech has posted an article that delves deep into the timeline changes made recently by AMD, and expands on Kyle_Bennett's original From "ATI to AMD back to ATI? A Journey in Futility" article, published in May 2016. While the content of WCCFTech's article is yet to be confirmed, it paints an interesting picture when it comes to the timeline for RTG Graphics, as well as the decision making guiding it.

    Sacrificing some of their gaming blood to make a comeback in the x86 market and a slow but steady stream of income from the mainstream and semi-custom GPU market is the trade-off at play here and while I do not have the actual numbers to back any kind of statement, it does seem like a feasible choice to make. Lisa’s choice might have robbed Radeon fans of their birthright, but it probably saved the company.
     
    Emission, lostin3d, Revdarian and 2 others like this.
  2. R_Type

    R_Type [H]Lite

    Messages:
    68
    Joined:
    Mar 11, 2018
    There hasn't been the money to do a 'proper' ati division in years then is the tl:dr
     
  3. pcgeekesq

    pcgeekesq [H]ard|Gawd

    Messages:
    1,129
    Joined:
    Apr 23, 2012
    Summary: "Rich people and rich businesses have money, so we're making products for them. Can't afford a $1000+ GPU? Then eat our table scraps."

    Of course, AMD is not alone in pursuing this particular business model. Clearly Apple, NVidia, Intel, Tesla, and others are doing so as well.
    They focus on products for the rich for the same reason Willie Sutton robbed banks: "Because that's where the money is."
     
  4. dgz

    dgz [H]ardness Supreme

    Messages:
    4,395
    Joined:
    Feb 15, 2010
    Hey universe, check out the dude with the navi.

    Can't blame the companies for chasing the money on this one, though. Just what the fuck are they supposed to do? It's not possible to make cheap hi end gaming cards that are also shit for anything but games or "soho" use. I think we absolutely don't want such segmentation anyway hehe. When was the last time you bought a GPU or CPU that wasn't purposely crippled in some way for no other reason?
     
  5. Zion Halcyon

    Zion Halcyon 2[H]4U

    Messages:
    2,099
    Joined:
    Dec 28, 2007
    I think this really fleshes it out.

    Sucks because I really like freesync.
     
    F.E.A.R., lostin3d, Flatline and 3 others like this.
  6. Zion Halcyon

    Zion Halcyon 2[H]4U

    Messages:
    2,099
    Joined:
    Dec 28, 2007
    The flip side is that AMD can't exactly be benevolent given how small of an operation they are.

    They were fighting for their lives so I don't blame them one bit for going after the money to save their company.

    If Raja can make some graphics cards worth a damn with Intel that also have freesync in them I will be a happy camper.
     
  7. ZippZ

    ZippZ n00bie

    Messages:
    19
    Joined:
    Nov 9, 2016
    Yup AMD was flirting with bankruptcy. You know a company is desperate when they have to sell their headquarters and rent it back from the people they sold it too.
     
    DrezKill, Darth Kyrie, N4CR and 2 others like this.
  8. R_Type

    R_Type [H]Lite

    Messages:
    68
    Joined:
    Mar 11, 2018
    There was a loooong period where they didn't have a fucking pot to piss in. So gpus went on the chopping block
     
  9. pcgeekesq

    pcgeekesq [H]ard|Gawd

    Messages:
    1,129
    Joined:
    Apr 23, 2012
    Benevolence isn't a good strategy for any for-profit company.
    It's good PR, of course. but there's no inherent correspondence between PR and reality.
     
  10. Croak

    Croak [H]ard|Gawd

    Messages:
    1,125
    Joined:
    Apr 16, 2006
    Jesus. Get help man.
     
    Aardvarck and auntjemima like this.
  11. cdabc123

    cdabc123 [H]ard|Gawd

    Messages:
    1,997
    Joined:
    Jun 21, 2016
    as someone who has quite abite of hardware i thought that would be fairly easy to list abunch of things but now thinking about it the only high end non-crippled hardware i have is a phi 7290, and a pair of es 8180's

    looks like I need to buy some more teslas :p
     
  12. DigitalGriffin

    DigitalGriffin [H]ardness Supreme

    Messages:
    4,528
    Joined:
    Oct 14, 2004
    I said roughly two years back that console development was a win win for amd. Ms and Sony pay for micro architecture r & d, even if near break even on chips. This carries over to the consumer.
     
    Darth Kyrie likes this.
  13. Stoly

    Stoly [H]ardness Supreme

    Messages:
    5,725
    Joined:
    Jul 26, 2005
    I think I now understand what AMD meant with new GPU every year.

    Year One... Professional cards
    Year Two... Mainstream gaming cards
    Year Three... High end gaming cards

    Rinse and repeat.

    I don't really hate the idea, except AMD is still far behind nvidia and still has to catch up: I don't see it happening anytime soon with the new roadmap.
     
  14. KazeoHin

    KazeoHin [H]ardness Supreme

    Messages:
    7,332
    Joined:
    Sep 7, 2011
    Although, Navi for PS5, that's pretty cool news.
     
    R_Type likes this.
  15. cdabc123

    cdabc123 [H]ard|Gawd

    Messages:
    1,997
    Joined:
    Jun 21, 2016
    amd has been neglecting the professinal cards in favor of gaming cards as of recently. (try to find a amd mi25) i see it more as a tick tock switching between big die and small die cards.
     
  16. pcgeekesq

    pcgeekesq [H]ard|Gawd

    Messages:
    1,129
    Joined:
    Apr 23, 2012
    Meanwhile, NVidia may be preparing to roll out the next-gen GeForces: https://www.techpowerup.com/245088/nvidia-briefs-aic-partners-about-next-gen-geforce-series

    "The BoM also specifies a timeline for the tentative amount of time it takes for each of the main stages of the product development, leading up to mass-production. It stipulates 11-12 weeks (2-3 months) leading up to mass-production and shipping, which could put product-launch some time in August (assuming the BoM was released some time in May-June)."​
     
  17. Zion Halcyon

    Zion Halcyon 2[H]4U

    Messages:
    2,099
    Joined:
    Dec 28, 2007
    True, but my understanding is the scuttlebutt out of Nvidia is that they are going to be slowing down production of gpus after that release. Frankly there isn't enough competition for them to want to keep up the Frantic Pace at which they released new gpus from, particularly from AMD.

    Intel will be the next potential rival for NVIDIA, but I think that would end up being at least 3 to 5 years out. And 0 guarantees Intel can pull it off.
     
    Cobra likes this.
  18. Uvaman2

    Uvaman2 2[H]4U

    Messages:
    2,327
    Joined:
    Jan 4, 2016
    Serious question.. on the article it references AMD changing to non GCN architecture... But what does that mean? Is that a good thing? A big deal? Anybody any ideas?
     
  19. Meeho

    Meeho 2[H]4U

    Messages:
    3,801
    Joined:
    Aug 16, 2010
    This. Wasn't Navi reported to be another GCN iteration with a process shrink?
     
  20. viscountalpha

    viscountalpha 2[H]4U

    Messages:
    2,309
    Joined:
    Oct 16, 2011
    if they approach it like they did with larrabee, then NO. They will not pull it off.
     
    dragonstongue and pcgeekesq like this.
  21. dragonstongue

    dragonstongue 2[H]4U

    Messages:
    2,860
    Joined:
    Nov 18, 2008
    coming from wccftech who bloody well knows lol.

    Vega is 5th gen GCN, Navi (7nm) is also supposed to be GCN (going by their numbers) would be 6th gen, likely the "next gen" would be a top to bottom new uArch, new memory, new cache design etc etc. (that is the one after Navi (7nm+ for awhile just called "nextgen")

    Only AMD truly knows at this point, they have been effectively building from a "middle ground" point from the 4xxx generation (RV700) sometimes building really over the top from this "baseline" midpoint, that is they targeted a "sweet spot" and built up or cut down from there, for example, 5870-5890, 7770-7970 and so forth, it has generally played out "reasonably well"

    but also has kind of gimped the lower end stuff more than it likely should have been and made the highest end offerings become very inefficient in regards to chewing a bunch of extra power or dishing out significant heat in the process. Kind of like AMD with bulldozer until they released some of the APU that reduced wattage/TDP budgets (they had no choice) they actually ended up with quite potent performance chips for the power being consumed.

    I know fanboys will be fanboys, but at least AMD has not chopped everything away to go "mach 1" in certain "races" while losing the ability to run other races like Nv has done over the last couple of generations. GCN is quite a potent Uarch OVERALL, if you look at all sides and not just pick out little bits and pieces and call it a "fail"

    Nv is really great at appearing all slutty dressed or being a big show off at the party compared to being the "real date" ^.^

    ------------------------------------------
    ------------------

    Time will tell, but it absolutely would not be the first time AMD/Radeon has caught competition off guard, if anyone can whip the RTG into shape it will be through Lisa Su leadership (those she has hired and shifted around to give AMD the best chance) these things take many years to "design" and put into place afterall.

    had Nv (or TSMC) screwed up with chopping things away with Pascal coming from Maxwell it would be a VERY different landscape, next to no power use and also very little real performance to show from it, AMD has very good yields on wafer use compared to Nv (last I heard) the last couple of generations.
     
  22. _mockingbird

    _mockingbird Gawd

    Messages:
    612
    Joined:
    Feb 20, 2017
    He was talking to the toilet.
     
    DF-1 likes this.
  23. mesyn191

    mesyn191 2[H]4U

    Messages:
    2,917
    Joined:
    Jun 28, 2004
    No one knows any details because whatever AMD is going to have in 2020-ish is probably still mostly in white board planning stages.

    BUT, its a good thing and a big deal.

    GCN is old and has some fundamental limitations that prevent it from scaling much anymore. It really should've been replaced by 2017, or perhaps early 2018 like AMD seemed to have been planning years ago. That didn't pan out so now they're stuck because it takes years to design and implement a new clean sheet GPU architecture. Sorta like how it is for CPU's since they're so complex these days.

    Until they come out with something new AMD will struggle to compete even in the mid range and they're probably going to have to compete on price even there. Basically kinda like the Bulldozer days all over again but for the GPU market and not for as long as a stretch of time.

    Pretty much. Maybe they'll do a multi die MCM like they've done with TR/Epyc with it but they've been very quiet on that front for a long time now.

    Usually if a tech company is quiet on a given issue it means they screwed up big time and can't deliver on their past statements and they're hoping it blows over or gets forgotten about when something else comes along to dominate the news cycle.
     
  24. N4CR

    N4CR 2[H]4U

    Messages:
    2,630
    Joined:
    Oct 17, 2011
    You forgot the article link, bud!
    https://www.hardocp.com/news/2018/06/12/amd_graphics_inside_story
     
  25. N4CR

    N4CR 2[H]4U

    Messages:
    2,630
    Joined:
    Oct 17, 2011
    Early roadmaps showed nexgen, not next gen. It was because of some development they'd been doing with nexgen for SSG type cards, which they did release with Polaris a year or so back and of great benefit to oil&gas types. AMD has been super quiet since Raja got the boot (gpu wise) and also CPU wise. Lisa has a tight grip on the grapevine these days which is good, AMD needs to knuckle down and bring results, not words.

    They won't talk about MCM because it's a massive jump for them performance wise and only way they can definitely beat Nvidia without a new uarch. Imagine AMD bringing a dual rx580 type thing on 7nm, with 1.6-1.8Ghz clocks, smaller total die area, less defects and less power than a 1080ti with the same performance? That's what they might be sitting on in future. Vega already has 500gb/sec interconnects and we see zero products taking advantage of them, why do you think they wasted that silicon space and have not used them? Because they are using their one-die-fits all strategy to also test for future products.

    Also there has been talk that they may just improve on the geometry limitations (the main issue with GCN) for next revisions of it, without replacing it. It can be done. That could mean something even more competitive going forwards without massive cost.
     
    Zuul and Darth Kyrie like this.
  26. N4CR

    N4CR 2[H]4U

    Messages:
    2,630
    Joined:
    Oct 17, 2011
    If that article is right I may as well keep my eyes peeled for a Vega or cheap 1080/1070.
    Problem is, Nvidia ain't going to do shit and AMD isn't going to do shit. So used GPU prices have no reason to drop >_>
    God damn perfect storm. Mining, then mining crashed out a bit for the northern hemisphere going into summer, now nothing coming out..
    Wonder if the 7nm vegas can be used with gaming drivers...
     
  27. mesyn191

    mesyn191 2[H]4U

    Messages:
    2,917
    Joined:
    Jun 28, 2004
    A distinction without difference.

    There are old roadmaps from like 2015 or whatever that showed their game plan and its clearly changed a whole lot from what they expected to happen back then. Releasing SSG high end compute cards wouldn't have effected their desktop roadmap.

    Huh? No they've blabbing about Zen2 and 7nm CPU's and APU's non-stop. They know they've got a winner there and are understandably vocal about it.

    Their silence about using MCM and IF to link multiple GPU dies is not at all a good sign for how things are turning out for that approach and in the past were quite vocal about it too.

    Any changes to GCN are going to be like all the past changes to GCN: iterative and evolutionary with minor and/or situational performance boosts. Same thing happened with Bulldozer. They're stuck for a long time until a new clean sheet design is finished. There is just no way around this.
     
  28. OrangeKhrush

    OrangeKhrush [H]ard|Gawd

    Messages:
    1,320
    Joined:
    Dec 15, 2016
    I agree, NV has basically become just about GPU's and cut themselves out of many markets by abandoning x86 progression. Ultrabooks and handhelds were a market I though NV would aggressively pursue with custom designs but ultimately may be just down to a niche market of mobile gamers that want a dedicated graphics card to pay through the roof for. The squeeze on Nv future in mobile is looking bleak with APU's and INTEL/AMD collaborations netting very good and cheaper designs.

    this is why some f the white collars suggest that long term investment in NV is a massive risk to be done with caution as should AMD and Intel get competitive in the graphics space which is plausible, NV's market could fall like a deck of cards.
     
  29. lostin3d

    lostin3d [H]ard|Gawd

    Messages:
    1,233
    Joined:
    Oct 13, 2016
    I totally agree. I really want to jump ship back to AMD, especially since the free-sync/4k/hdr t.v.'s are about to hit the scene. Unfortunately, kind of useless without a proper gpu to drive them.
     
  30. N4CR

    N4CR 2[H]4U

    Messages:
    2,630
    Joined:
    Oct 17, 2011
    Yeah the roadmaps have changed for sure, I was looking at one today in my documents and laughed seeing Navi was supposed to be out soon.
    But if you keep that in mind, no wonder they've been quiet about MCM, it's likely a large part of why they are moving things around. Because they need the right core to base it on. Vega/Polaris both have IF links that are unused bar a very niche product lineup, they are not there for a niche, they can't afford that. They are there to R&D for future products.

    We all know that process tech is nearly at the end of the free ride. The only way out from here, just like CPUs, is new tech or more cores/parallelising. You're right it takes a long time and RTG has not exactly been floating smoothly, so perhaps now things are more set, these timelines are realistic.. we are a while away from MCM if it ever happens.

    Yes they have been noisy about 7nm. Both vega 7nm and CPU 7nm, because they are both coming out soonish (industry speaking). Graphics seems to be a dead street for the next while so they will of course be quiet about that, outside of enterprise vega 7nm coming.
    As you said it takes time to come out, I think they are still unsure given the recent changes and likely NDA'd to the balls with Sony for navi, so have to be more quiet than usual.

    P.s. Nexgen is a company AMD aquired in 90s..

    From wiki.
    Quite a coincidental 'slip' in wording on the roadmap don't you think?
     
    Sulphademus likes this.
  31. mesyn191

    mesyn191 2[H]4U

    Messages:
    2,917
    Joined:
    Jun 28, 2004
    If they could scale up MCM'd GPU's with IF they'd have something that would allow their current GPU's to compete with 1080Ti's and at least some of NV's next gen high end products and they wouldn't be quiet about it at all. Something isn't right here. Exactly what I can't know but they wouldn't sit quietly on that sort of thing at all. Heck they've already announced 32GB HBM 7nm Vega products to tout their ability to compete in HPC GPGPU markets for that same reason.

    No not really. They've mispelled things before. No point reaching for conspiracies.

    Personally I'd like to see MCM'd GPU's come to fruition because it'd allow AMD to scale up mid/low end GPU's into high(er) end products in a more affordable fashion and I'm tired of the BS prices on graphics cards these days. There just isn't much reason to believe they'll pull it off unfortunately.
     
  32. N4CR

    N4CR 2[H]4U

    Messages:
    2,630
    Joined:
    Oct 17, 2011
    Fair call, AMD slides have terrible QC aka spell checking over the years xD

    AMD doesn't want to compete with Tis, because they are a pathetic 2-3 maybe 4% of the market at most. It's peanuts for all that R&D $. The only thing that competing with Tis does is get them good marketing. I'd much rather see them bring out solid mid range and then bring a beasty Navi 20/40 MCM GPU as you were saying, for affordability. MCM could offer that affordability, bit like Threadripper vs intels 9897369xx whatever the HEDTs are called this time around. We all know know intel is hurting selling those massive, monolithic chips for 20% of what the uncut xeons go for.
    MCM should be a logical step if the interlinks are there, GPU loads are easily parallelised as we already see them now, massively parallel. Razor1 wrote a really good post about this a while back which explained the interlink bandwidth is really the main the limiting factor, at that time we only had polaris. Vega is enough to begin experimenting with it at lower resolutions now. The interlink bandwidth is an issue because of resource scheduling, he certainly didn't say it's not possible with the right hardware, just that it's hard and a few big hurdles to overcome. Nvidia is also looking into it and they don't have provisions for interlinks on their lower end GPUs at all. It's likely going to be the future for GPUs as we are running out of places to go as process advantage dries up and the easier/lower hanging fruit on the uarch stuff is already done.
     
  33. Cobra

    Cobra 2[H]4U

    Messages:
    2,639
    Joined:
    Sep 7, 2000
    If they are smart, they will use the time in the slowdown to do a full ground up architecture rethink for a totally new generation of graphics. If they are dumb, they will pull a classic Intel and use the time to twiddle their thumbs while the profit rolls in, and then be blindsided when the competition releases a product that is much faster.
     
  34. DigitalGriffin

    DigitalGriffin [H]ardness Supreme

    Messages:
    4,528
    Joined:
    Oct 14, 2004
    Yes and No. Navi has GCN elements, but it's focus will be scaling, like in multi chip scaling. (And we aren't talking SLI/Crossfire)

    GCN operates very much like a RISC chip with a lot of ALU units. There's a lot of complex overhead leading to excess number of transistors. I think AMD was hoping console games would carry over to desktop games, and the optimizations there would benefit AMD as the architectures are similar. The PS4/XBone market is much bigger than the PC market in terms of AAA titles. (Unfortunately it's not playing out that way) It takes some considerable work to wring the performance out of GCN. (ie: DOOM) And when you do, it's a beautiful thing. A RX580 can take on a 1070 easily. But that's time people don't want to invest (getting down to the metal). Therefore I predict AMD will look at the most used GCN circuits and keep those, and remove the lesser used ones (when possible) They will also use multichip packaging with infinity fabric and large caches on each chip. I bet they learned a lot from the XBones dedicated small memory buffer and how much it granted them in terms of performance/bottlenecks.

    AMD has a tremendous amount of data on small optimized caches that NVIDIA simply doesnt have because they never built those kind of hybrid architectures before.

    Whether AMD pulls it off or not is another thing.
     
    Last edited: Jun 13, 2018
    Sulphademus likes this.
  35. DigitalGriffin

    DigitalGriffin [H]ardness Supreme

    Messages:
    4,528
    Joined:
    Oct 14, 2004
    Didn't I say "Not any time soon" (Out of Smug Mugs mouth) was completely relative to what "Soon" means?
     
  36. Stoly

    Stoly [H]ardness Supreme

    Messages:
    5,725
    Joined:
    Jul 26, 2005
    That's odd I've heard the exact opposite.
     
  37. Stoly

    Stoly [H]ardness Supreme

    Messages:
    5,725
    Joined:
    Jul 26, 2005
    Charlie... is that you? :D:D
     
  38. DigitalGriffin

    DigitalGriffin [H]ardness Supreme

    Messages:
    4,528
    Joined:
    Oct 14, 2004
    Actually their yield was good, but it wouldn't scale GHz wise for the grahics chips. Bulldozer scaled well due GHz wise due to it's very long pipeline which kept traces short.
     
  39. Zion Halcyon

    Zion Halcyon 2[H]4U

    Messages:
    2,099
    Joined:
    Dec 28, 2007
    History tells us that most companies are dumb. Rather than forward-thinking, they want to ride out profits for minimal cost.

    You would think companies would learn that this is how competitors end up biting you in the ass as Intel's learning from AMD right now.
     
  40. Meeho

    Meeho 2[H]4U

    Messages:
    3,801
    Joined:
    Aug 16, 2010
    Any data on how it would work? Latencies seem like a possible issue for high fps crunching.