What is the "it" that I mentioned? Please be specific.
The only statement I've made is that RDNA 3 perf/watt improvement doesn't appear to be an outlier which is the premise of this thread.
RDNA3 Efficiency [Spinoff from RDNA4]
- Thread starter Arun
- Start date
What is the "it" that I mentioned? Please be specific.
The only statement I've made is that RDNA 3 perf/watt improvement doesn't appear to be an outlier which is the premise of this thread.
But by posting those numbers, one can only come to the conclusion that, you believe RDNA2 had <50% the ppw improvements of RDNA1 and RDNA3.
That means RDNA2 had below average ppw improvements and puts it in the league of Hawaii and Vega.
Edit- And that RDNA3 had similar ppw improvements as RDNA1.
I'm not drawing any conclusions about RDNA 2. I was trying to understand the claim that RDNA 3 is broken.
Even if we accept that it didn't ultimately meet AMD's internal targets and is pulling too much power at nominal voltages the fact remains that delivered performance is fine and competitive with other architectures on TSMC 5nm. So definitely not "broken" or an "outlier".
AMD aimed really high and missed is a more believable story.
That is such a cop out... the data you gathered and used to validate your opinion about RDNA3 shows RDNA2 has terrible ppw.
Are you claiming the data you collected and posted can ONLY be used for RDNA3 comparisons and not accurate for any other comparisons or conclusions?
Answer yes or no.
Does RDNA3 have 2x the ppw of RDNA2?
Does RDNA2 have terrible ppw improvements over RDNA1?
Either your data is correct/accurate and RDNA2 had terrible ppw or maybe there is something wrong with your data/comparison causing you to come to the wrong conclusions, i.e. it might be due to something already discussed in this thread.
From AMD's own marketing materials:
Thought I was pretty clear. This thread is about whether RDNA 3 was a busted arch in terms of perf/watt. It’s clearly not. The numbers I shared straight from TPU support that conclusion.
Btw what exactly about my numbers indicate RDNA 2 had terrible perf/watt? I certainly never claimed that.
Last edited:
I have answered all that previously... You have to be trolling me at this point...
Why are you constantly asking me questions (ones I have clearly answered) and not answer any of my questions?
It doesn't seem like this discussion is in good faith when you are consistently avoiding/ignoring/dismissing most of my posts.
Edit-
Are you claiming the data you collected and posted can ONLY be used for RDNA3 comparisons and not accurate for any other comparisons or conclusions?
Answer yes or no.
Does RDNA3 have 2x the ppw of RDNA2?
Does RDNA2 have terrible ppw improvements over RDNA1?
Okay... I have covered that already in my first post here.
Last edited:
Mod mode: chill out. This thread was already a turd sandwich before I had to clean house, I'll clean users next time.
Back to regular posting, I don't understand why the absolutism is necessary. I don't think RDNA3 is twice (or more) the PPW of RDNA2. I see gains across the board, which is expected of a generation leap. I don't think RDNA2 was so bad as to need some new definition of terrible, either.
I mean, at this rate we are going to end up going back to the 7500 (like, 20 years ago) to somehow conclusively prove some mysterious power efficiency thing about AMD which then polluted ATI's good name or some such.
Back to regular posting, I don't understand why the absolutism is necessary. I don't think RDNA3 is twice (or more) the PPW of RDNA2. I see gains across the board, which is expected of a generation leap. I don't think RDNA2 was so bad as to need some new definition of terrible, either.
I mean, at this rate we are going to end up going back to the 7500 (like, 20 years ago) to somehow conclusively prove some mysterious power efficiency thing about AMD which then polluted ATI's good name or some such.
Last edited:
The whole question is whether RDNA 3 was an outlier wrt RDNA 2. Looking at one review isn't going to reveal any insights; look at more then one and see if there is any consensus. @CarstenS posted results from multiple reviews above which seem to indicate a concensus on the question.
I’m ignoring your questions because they’re not related to anything I’ve said or the topic of the thread. You’re setting up random strawmen for some unknown reason. Would take Albuquerque’s advice.
This is why there is no point in having real discussions anymore.
I took issue with some of the claims people made, I attack the data.
People post slightly different data making the same claims, I attack the data.
What do I get in response for my effort?
I get attacked and accused of "setting up random strawmen."
I took issue with some of the claims people made, I attack the data.
People post slightly different data making the same claims, I attack the data.
What do I get in response for my effort?
I get attacked and accused of "setting up random strawmen."
Nothing about performance-per-watt can be absolute, anyway.
Somewhere earlier in the thread (which I believe may have been lost to my great purge), someone mentioned performance-per-watt can be traded for die area and can also be traded for competitive advantage. My 4090Ti can be insanely good at performance-per-watt when I turn down the clocks substantially and the voltage accordingly. Of course I'm speaking specifically about Folding at Home, where the output is knowable and easily measured. Even at stock configuration it's better than every other desktop card: ( https://folding.lar.systems/gpu_ppd/overall_ranks_power_to_ppd ) and when undervolted it's patently insane how good it gets.
At the same time, we have more than a few historic examples where cards were juiced pretty hard after discovering the competitor's offering was going to out-perform what was expected. Of course, perf-per-watt can only rationally discussed as a function of the delivered SKU rather than some hypothetical which never materialized. As such, much of this conversation ignores the other environmental factors which conspired to deliver the result.
Which is to say, I think it's interesting to talk about those factors, but it's useless to draw hard lines in the sand about how much it truly mattered. All of us nerds on this forum obviously care as it's part of what we talk about here. At the same time, it really is a fungible figure which can absolutely change based on workload, the rest of the PCB phy, chip die size, competition, drivers, operating systems, you name it.
Somewhere earlier in the thread (which I believe may have been lost to my great purge), someone mentioned performance-per-watt can be traded for die area and can also be traded for competitive advantage. My 4090Ti can be insanely good at performance-per-watt when I turn down the clocks substantially and the voltage accordingly. Of course I'm speaking specifically about Folding at Home, where the output is knowable and easily measured. Even at stock configuration it's better than every other desktop card: ( https://folding.lar.systems/gpu_ppd/overall_ranks_power_to_ppd ) and when undervolted it's patently insane how good it gets.
At the same time, we have more than a few historic examples where cards were juiced pretty hard after discovering the competitor's offering was going to out-perform what was expected. Of course, perf-per-watt can only rationally discussed as a function of the delivered SKU rather than some hypothetical which never materialized. As such, much of this conversation ignores the other environmental factors which conspired to deliver the result.
Which is to say, I think it's interesting to talk about those factors, but it's useless to draw hard lines in the sand about how much it truly mattered. All of us nerds on this forum obviously care as it's part of what we talk about here. At the same time, it really is a fungible figure which can absolutely change based on workload, the rest of the PCB phy, chip die size, competition, drivers, operating systems, you name it.
You also set up these absolutist "IS THIS TRUE OR FALSE, ANSWER MEEEEE" which doesn't help anything at all, ever. Only sith deal in absolutes; perf-per-watt is not absolute and it very much varies on a multitude of factors. These variables are called nuance, and it's worth talking about.
At least be self-aware enough to own your part of it. Or don't, and then storm out of the thread in a huff I guess.
Because that was part of the data they provide. The data is either accurate and true, or it is not.
There was also multiple times in this thread that certain people DEMANDED others to answer their question/s.
I made the request to prevent the discussion from circling back on itself since that is a no-no here.
I'm pretty sure I made my opinion perfectly clear about the variables involved with ppw when this all started.
Edit- To go back to the start of when I entered the thread, the absolutist claim is that RDNA3 ppw improvements were not an outlier.
Last edited:
Yep that’s the question. “Not an outlier” isn’t an absolutist position as clearly there is a wide range of values that would fall into that category.
The evidence provided in this thread is that RDNA 3’s perf/watt improvement is in a reasonable range relative to prior generations and also relative to its contemporary competition . Hence refuting the hypothesis. If you want to prove RDNA 3 is an outlier you would need to show that it’s significantly worse of an improvement than other generations.
It is an absolutist position when you aren't allowed to disagree with that position.
SOME of the evidence, not ALL the evidence.
I did.
Last edited:
It is not an absolutist position, ever. You're welcome to disagree; the mods make the rules, and none of the rules say you cannot disagree. The rules are: you need to disagree in a professional and courteous manner.
Further, as evidenced by the retorts in this thread, it has not been conclusively proven that RDNA3 is an outlier. Communication is a two way street - just because one person states a bunch of things in an authoritarian way doesn't mean everyone needs to agree with that one person. Weren't you JUST complaining about "we're not allowed to disagree!!" Turns out, we are allowed to disagree!
As another interesting datapoint, I wanted to see what Folding at Home could show us in these regards. Here's what I found with about five minutes of searching:
VEGA 10 - Radeon RX Vega 56/64: 71,504 cards in the sample pool, averaging 5.14 work units per day, consuming 7.1kWh per day, averaging 1.38kWh per work unit.
NAVI 10 - Radeon RX 5600/5600XT/5700/5700XT: 16,969 cards in the sample pool, averaging 6.67 work units per day, consuming 5.4kWh per day, averaging 0.81kWh per work unit. (This is a ~71% uplift in perf per watt versus VEGA)
NAVI 20 - Radeon RX 6700/6700XT/6800M: 40,130 cards in the sample pool, averaging 8.17 work units per day, consuming 5.5kWh per day, averaging 0.67kWh per work unit. (This is a ~20% uplift in perf per watt vs NAVI 10)
NAVI 30 - Radeon RX 7700XT/7800XT: 2,281 cards in the sample pool, averaging 11.9 work units per day, consuming 5.9kWh per day, averaging 0.50kWh per work unit. (This is a ~35% uplift in perf per watt vs NAVI 20)
I picked the x700 series of cards because there wasn't a 5900 series of NAVI 10. Feel free to rerun the numbers with your own SKUs here: https://folding.lar.systems/gpu_ppd/overall_ranks Also, I went with work units rather than points, because there's some fluidity with how points are assigned versus how work units are doled out.
The comparo I ran above also gets a little sideways because F@H blends several of the SKUs together, which can serve to slant the power and performance in various directions. At the same time, it's more than 130,000 cards in the sample pool, and it shows that (in my interpretation) NAVI 10 was going to be hard to live up to. We could point out that NAVI 20 somehow missed the mark comparatively, however I feel like that's a bit unfair given the predecessor. And to me, NAVI 30 did better than NAVI 20 if we're talking about performance uplift from prior gen.
If I wanted to be a shit, I could say my data has far more datapoints than ANYTHING anyone else has put up, so my data >>>>> all y'alls data. Who else is gonna come back with a dataset with an n > 130,000?
But that's not what we're about here. Instead, what I will respectfully point out is this additional (not penultimate) data shows us there's a lot of nuance going on, which seems to suggest there isn't one absolutist position we can take.
Well...
Did the aggregate performance-per-watt-hour analysis of 130,000 machine-collected, anonymized datapoints finally answer this question?
Kinda looks like RDNA1 was the actual outlier to me
Did the aggregate performance-per-watt-hour analysis of 130,000 machine-collected, anonymized datapoints finally answer this question?
Kinda looks like RDNA1 was the actual outlier to me
I didn't know this (i wager I did at some point in the distant past) so that seems even more interesting!