AMD: Speculation, Rumors, and Discussion (Archive)
- Thread starter iMacmatician
- Start date
- Status
Silent_Buddha
Legend
Well, that was pretty poor, I was hoping it would at least come between the 390 and the 390x, oh well...
I almost picked up a 390 for £199 from Overclockers this morning but decided not to, looks like that was a bad decision. I made the same mistake when the 290X's were heavily discounted the day before the 390X reviews went live
http://www.anandtech.com/show/10446/the-amd-radeon-rx-480-preview
Well, it's roughly on par with 390/390x at 1080p. It falls short at 1440p in many cases, however. I wonder if this might be at least partially due to having half the ROPs.
It's also almost half the power consumption of 390/390x.
[rest is general reply to everything else since reviews released]
I'm with Rys on this. It's somewhat disappointing considering the hype leading up to this, but sticking with comparisons to AMD's own past products it's not bad. And given the price (at least until we see what Nvidia responds with) it's the best product you can get at that price point.
It'd be good enough for me if it was more consistent at 1440p. If I was still gaming at 1080p/1200p, I'd likely buy this immediately.
Compared to Pascal, it's obviously not as good architecturally.
The most disappointing thing to me is that it appears they've likely clocked this past the knee of the power curve similar to Fury X, which while increasing the performance makes the perf/W significantly worse as well. Fiji at those clocks had terrible perf/W while Fiji at Nano clocks had decent perf/W. I think we're seeing something similar here. If there's any disappointment to be had, it's likely based around this.
At what AMD were probably hoping to target it's clockspeed, it likely did have up to 2.8x perf/W. But clocked past the knee of the curve, it's only about 2x perf/W. Either that or the results from Global Foundries 14 nm FF implementation were significantly worse than they expected. Perhaps the 2.8x perf/W came from a test sample at Samsung or simulated using characteristics from Samsung 14 nm FF. Somewhat reasonable considering GF is used shared tech. But implementation varies, and it's possible GF did a lot worse with it than Samsung did.
Either way. ~2x perf/W (compared to 390/390x) for an equivalent gaming performance is still respectable. Sure AMD could have done better. But they also could have done a lot worse.
But, yes, as Rys mentioned, too much reliance on pre-product release hype from 3rd party websites (Videocardz/WCCF/Etc.) greatly inflated expectations for a card targeted at the 199 USD price point.
I admit, I also fell for some of that hype. But I am also not a slave to the hype and can look at this somewhat more objectively. It's not what AMD wanted and it doesn't live up to the 2.8x perf/W. But it pretty much delivers on everything else that AMD promised, which was to bring VR capable performance down to mainstream prices.
And hell, what did most people on this forum say when AMD had better perf/W than Nvidia? Perf/W doesn't matter in a shipping product.
I also, am shocked at people's response to this. Not as shocked as Rys, as I was expecting some of this from certain people, but still shocked at how many people appear to be overreacting.
Disappointing certainly. But all this vitriol. Dang.
Regards,
SB
Orion
Regular
It might have no effect at present, but AMD's continued existence if they manage to compete will affect future prices.
About the up to x2.8 efficiency:
https://www.reddit.com/r/pcmasterra...at_amd_the_time_has_come_to_ama_about/d4smj2dAMD Robert Hallock said:
Man, but do I wish that vega is just a few months off and not 6+ months off... Not sure that's a realistic wish though. *shrug*
While polaris might be disappointing on a strictly implementation level, the under-the-hood improvements that sebbbi mentioned means vega could be much more interesting. It probably won't compete with NV's equivalent because AMD doesn't have the R&D resources, but if we get good feature set and good performance at a decent price rather than astronomic price like NV's high-end offerings (and expensive proprietary bullshit, like g-synch and so on); I don't care if the chip is running as hot as hawaii. The heat I can live with.
It's the frames per seconds I want. Give me that, and I'll be happy.
Ext3h
Regular
One thing I noticed about the gaming benchmarks:
Witcher 3 was always tested with Hairworks off.
Yes, that option is known for beeing troublesome in the past. But it would be interesting to see if the "discard accelerator" isn't actually solving the issue for Polaris.
That's actually something I would hope for in general - that upon the release of a new architecture, games aren't tested in the well known failsafe configurations, but rather with previously problematic features intentionally turned ON.
Witcher 3 was always tested with Hairworks off.
Yes, that option is known for beeing troublesome in the past. But it would be interesting to see if the "discard accelerator" isn't actually solving the issue for Polaris.
That's actually something I would hope for in general - that upon the release of a new architecture, games aren't tested in the well known failsafe configurations, but rather with previously problematic features intentionally turned ON.
Benetanegia
Regular
I don't believe in marketing myself. That's not the point, Rys asked for which slide was the cause of the problem and I provided it.
Problem here is that it is quite a bit more than a notch. I don't think you can just dismiss everything with "it's marketing", or should we just allow them to make ridiculous claims with no responsability afterwards? There's always some expectation of at least some correlation to reality. It's hard to see what correlation to reality exists here. From 2.8x last gen, to the Fury and Nano actually offer better perf/watt there's a world of a difference. I don't think that is something we should allow to happen without proper criticism, but then that's my opinion.
Anyway, I didn't want to enter the discussion, I just posted the relevant slide to "prove" that the 2.8x figure wasn't a fabrication by the likes of wccftech, fudzilla, etc. And that AMD did actually claim big architectural improvements.
Orion
Regular
I was immersed in the hype, at [edited] boost clocks I expected around 390x in some apps, though I considered it possible and likely that the 390x might be slightly faster in several perhaps many apps(10-15%). in some apps reviewed at 4k the 480 is within 1-3~fps of the 390x, doesn't seem to be significant difference.
But in certain apps such as ashes I expected significantly above 980 performance, given what we've seen of 390x in some benches this year(30% faster than a 980ti) and how close the 480 can get to it.
Overall I find it hasn't disappointed my hype, performance wise, around 390x perf in several apps, even at 4k, 30% faster than a 980 in some apps at some (1440p)settings. The power issue does seem like a pickle, can the manufacturing process improve and resolve it in the months to come, or is it unfixable?
Agreed.
One similar area showing improvements is the benchmarks using Fallout 4, which I understand the lighting-shadows/AO/Godrays were responsible for heavily impacting performance on AMD.
The 480 is a good improvement even over the 390X, and close to the Fury X.
And that result seems pretty similar to other reviews using Fallout 4.
Although would you expect from a DX12 AoTS game perspective the 480 to be slightly slower than a 390 (which some reviews measured)?
Cheers
https://translate.google.de/transla...iagramm-the-witcher-3-tessellation&edit-text=
It's definitely helping, although Nvidia is still leading by a noticeable margin.
Raja's "2.8x" slide was an oblique comparison to NVidia's similar improvement=process+architecture claim. Raja showing perf/watt, and NVidia's was showing raw performance. An NVidia perf/watt graph would be pretty similar.
Yeah, 32 ROPs feels like a mistake, especially considering the marketing angle they took with it (bringing high-end to maintstream). Probably should have gone 64 or at least 48. And 40 CUs. At 64 ROPs, 40CUs you have a 390 in base units, and you know that anything else (clocks, arch. improvements, etc) is gravy. Of course this means +10 - 12% die size. But then you just focus on beating the old high-end on performance and performance/$, and you do not talk about power. OTOH, the $199 price point is really important, so I think the card could end up doing much better (for AMD) than many people here expect. And if the additional die size pushed the price up to 220 - 250, then maybe it wouldn't have been worth it.
Something else I am wondering is what this means for the next-gen consoles (especially MS).
Something else I am wondering is what this means for the next-gen consoles (especially MS).
Love_In_Rio
Veteran
MS will launch with whatever AMD chip at 6 Tflops but made at TSMC.Sony should do the same.
monstercameron
Newcomer
Yet most of the analysis -save for sebbi- has only been around process tech or perf/W. I a little disappointed in the snr around here lately.
Ext3h
Regular
And one more thing, regarding the perf/W number published by AMD:
I strongly believe that using FP16 over FP32 is already factored in into that number.
That is not of relevance for already published titles, but it certainly is for future ones. That should then also put Polaris clearly ahead of the 390, as marketed.
Oh, and something else:
The "real time" capabilities. (See https://en.wikipedia.org/wiki/Real-time_computing for what that term refers too, it's not the same "real time" as in "real time graphics".)
AMD has yet to publish which latency guarantees the card can actually make, but it's a new one for any GPU so far to make these at all. There is a good chance this makes Polaris not only attractive for gaming, but also for industrial applications where providing real time characteristics is a must have in many domains.
The driver can't mess up the latencies this time, as far as I understood. For tasks scheduled to run at a fixed time, the command processor has a builtin hardware clock and is acting independently from any possible host side interference.
One word of caution though: On inquiry, AMD has stated that they are not literally making guarantees regarding latencies.
I'm not sure whether this is due to not wanting to take the legal risk, being unable to guarantee this for the driver side, or if there are still some caveats in the hardware.
In case the hardware is capable, as I expect, we might see a certified driver for the professional series GPUs.
(Btw.: Real time constraints also apply to automated driving. One of the reasons why you won't e.g. see Nvidias Drive PX system in any regular series, there is no way that hardware can pass the regulations for a safety critical component.)
I strongly believe that using FP16 over FP32 is already factored in into that number.
That is not of relevance for already published titles, but it certainly is for future ones. That should then also put Polaris clearly ahead of the 390, as marketed.
Oh, and something else:
The "real time" capabilities. (See https://en.wikipedia.org/wiki/Real-time_computing for what that term refers too, it's not the same "real time" as in "real time graphics".)
AMD has yet to publish which latency guarantees the card can actually make, but it's a new one for any GPU so far to make these at all. There is a good chance this makes Polaris not only attractive for gaming, but also for industrial applications where providing real time characteristics is a must have in many domains.
The driver can't mess up the latencies this time, as far as I understood. For tasks scheduled to run at a fixed time, the command processor has a builtin hardware clock and is acting independently from any possible host side interference.
One word of caution though: On inquiry, AMD has stated that they are not literally making guarantees regarding latencies.
I'm not sure whether this is due to not wanting to take the legal risk, being unable to guarantee this for the driver side, or if there are still some caveats in the hardware.
In case the hardware is capable, as I expect, we might see a certified driver for the professional series GPUs.
(Btw.: Real time constraints also apply to automated driving. One of the reasons why you won't e.g. see Nvidias Drive PX system in any regular series, there is no way that hardware can pass the regulations for a safety critical component.)
Silent_Buddha
Legend
That makes sense as they aren't likely to have to clock Polaris 11 as aggressively as they've had to clock Polaris 10. What would be amusing is if Polaris 11 ended up being manufactured at Samsung. I doubt that'll be the case, but it would be amusing.
But at the price point (199 USD) how many people are going to be using a monitor with higher resolution than 1080p/1200p? Especially when you consider that most gamers tend to spend more on their graphics cards than they do on their monitors. I'd say it's a reasonable compromise for the intended market segment. The performance falloff in some games at 1440p and higher is disappointing certainly, but the reasoning behind the design choice is solid.
As for consoles. They are semi-custom designs. They console holders can request whatever they want. If a console maker wanted a Polaris based GPU in their SOC with 40 CUs and 256 ROPs (ridiculous configuration to illustrate the point), AMD would do it.
Looking at things though, I think it becomes more likely that Microsoft in particular is looking at something closer to Vega than Polaris. If we assume that Vega features 64 (or more?) ROPs and is targeted at 4K gaming (something Polaris with 32 ROPs is going to be questionable at doing) that seems to reinforce the idea that Vega will be closer to what they need. Even if everything else (CUs for example) is just a very small evolution or even identical to Polaris (IE - Vega using the same CUs as Polaris which would make sense).
Regards,
SB
Polaris 11 is a good product but I'm let with the feeling that AMD tried to pushed its architecture too far. I don't know where the bottleneck is but there is a bottleneck that prevent their GPU to scale as they should, I don't know the reason.
It shows on some testing done by the "techreport" wrt the texture filtering rate, the RX480 peaks at the same level as Tonga. The Fury GPU goes past that level.
I would think that the L2+RBE partition is the limiting factor, the ring / interconnect, (whatever it is) seems to be able to move enough data for up to 16 CUs though it needs to be fed properly.4 RBE partition (8 ROPS) is possibly already limited at 32CUs, looking at their line of GPU I suspect the Bonaire hit the sweet spot with 14 CUs per RBE partition.
Another thing that bothers me is that AM still try to compete with Nvidia in the geometry throughput department. They made neat architectural advancements but they are still not where Nvidia stands. Polaris 11 is not meant to be an high end I think they could have gone with a less complicated "inner layout" as I suspect that a fixed function hardware burn its fair share of power.
Long story I would have like to see Polaris 11 go with only two Shader engines and less CUs (wherever the limit is most likely below 32 for four RBE partition). Performance difference in "real" games might have been marginal and it was worse the saved Watts (less CUs, less functional units), still close enough from the GTX 970. The amount of RAM 4 GB is right. Offering a 8GB is a nice commercial move. I believe the base clock and turbo clock are also a little too high (hence the power consumption).
Overall whereas the pitfall is power consumption I think that AMD "gives" a little too much for the price of those cards.
It shows on some testing done by the "techreport" wrt the texture filtering rate, the RX480 peaks at the same level as Tonga. The Fury GPU goes past that level.
I would think that the L2+RBE partition is the limiting factor, the ring / interconnect, (whatever it is) seems to be able to move enough data for up to 16 CUs though it needs to be fed properly.4 RBE partition (8 ROPS) is possibly already limited at 32CUs, looking at their line of GPU I suspect the Bonaire hit the sweet spot with 14 CUs per RBE partition.
Another thing that bothers me is that AM still try to compete with Nvidia in the geometry throughput department. They made neat architectural advancements but they are still not where Nvidia stands. Polaris 11 is not meant to be an high end I think they could have gone with a less complicated "inner layout" as I suspect that a fixed function hardware burn its fair share of power.
Long story I would have like to see Polaris 11 go with only two Shader engines and less CUs (wherever the limit is most likely below 32 for four RBE partition). Performance difference in "real" games might have been marginal and it was worse the saved Watts (less CUs, less functional units), still close enough from the GTX 970. The amount of RAM 4 GB is right. Offering a 8GB is a nice commercial move. I believe the base clock and turbo clock are also a little too high (hence the power consumption).
Overall whereas the pitfall is power consumption I think that AMD "gives" a little too much for the price of those cards.
- Status
