RDNA3: What is the Real Performance per Watt?
Вставка
- Опубліковано 4 сер 2023
- What is the real performance gain of the RDNA3 architecture in the AMD Radeon 7000 GPUs this generation? In this video, I compare RDNA3 vs RDNA2 in power scaling and determine the real performance per watt and how AMD is wrong in what they advertise.
RDNA3 Performance Playlist:
• RDNA3 - Performance
#RDNA3, #Radeon7000, #amdradeon - Наука та технологія
Tech channels often cover repetitive topics and make similar points, but your videos stand out by focusing on rarely discussed issues. That's why I love your channel.
Thank you, I really appreciate that!
AMD's whole presentation I would consider a lie. 1.7x performance vs 6950xt and 50% performance per watt are intentionlly misleading. If we didn't have independent reviewers on UA-cam and all we had were magazines like PC World , a lot of people wouldn't even know the truth.
What about 4X performance per watt of Nvidia claim.
Also games don't scale well with the FPS metric.
To be fair it was up to 1.7x in 1 game but i don't think it hit that in that game haha
Those 6.7% probably comes from the 6nm node vs 7nm and the extra 0.5gbps memory speed. RDOA3 in reality has 0% improvements.
If you mean N33 vs N23, improvement you say yeah, but it was transfered from 7nm, no new architecture, but 7900 family is new, improvement is there even 5nm vs 7nm is just 15perc faster, or 30perc more efficient.
Techpowerup say 7900xtx is 47perc faster than full 6900xt, which with 15 percent of faster node means almost 28perc design gen to gen uplift.. MCM is definitely generational typical uplift.. Whoever say whatever bad.. Check it. Or people spend stupidly 1000USD 7900xtx vs 600USD (N21 gpus)?
A small UA-cam thumbnail suggestion: having a red border alongside the bottom might make people think they've watched the video already, reducing the likelihood it's clicked on. A red line along the bottom is already used to denote how much of the video you've watched.
I do like how the red border is the same thickness as the watch-time bar, but I'd remove the border alongside the bottom edge.
Thank you for that feedback!
I like plotting the curves as it shows the most detail. Plot one for a 4080 and watch Nvidia blow AMD out of the water. Lovelace is the biggest jump in gpus in quite literally a decade
Going from Samsung 8nm to TSMC 4nm really did help out 😁 sucks that costs of the graphics cards went up significantly too...
That power scaling graph explains why the RDNA3 iGPUs are barely quicker than the RDNA2 iGPUs from last gen
iGPUs are bottlenecked by slow ram speeds anyway
What's even more concerning about the power consumption misinformation is the issue with competitive games like OW2 that OptimumTech demonstrated - basically the 7900XTX consumes 2x the power of a 4000 series card when playing games that are not even stressing the GPU at close to 100%. It's almost like there's a minimum power treshold on the Navi 31 chip that's required for it to boost above base frequency.
That's the flaw of a. Chiplet design. But manufacturing cost is saved there.
This is a good analysis video, i like this type of content, keep it up!
Thanks, will do!
I am gonna share this video. Great work dude.
Thanks for sharing! Much Appreciated!
this is a very smart video - impressive analytical skill here - maybe there is hope for rnda 3 in low power devices
It's criminal how few views you get. I'm sure your channel will blow up one day.
I appreciate that. Thank you.
The performance gains where correct at the time of testing.
But the 7000 series had really bad artifacting.
AMD fked up because they realised this too late, hence the delayed launch.
They rushed out a fix at the 11th hour, but it came at the cost of performance.
AMD hoped to fix the performance with software updates down the road.
But they can't fix one without messing up the other
Nice conspiracy theory...
@@fazlerabbi460 it was mentioned on MLID
What did AMD change at the last minute to reduce performance gains but eliminate artifacting?
@@ImaMac-PC i cant remember that bit. I'm trying to find the video where he mentions it.
Yet they priced these cards as if they were good.
Yeah I would like to mention that the RX7600 is on 6nm not 5nm like the RX7900xtx so numbers could actually be notably different with the FPS per watt.
You'd think AMD would pass the savings from 6nm to the customer, but they just are being as greedy as Nvidia.
6nm is efficient, 6nm area is 36units , when we look into intel on 10nm CPUs, consuming 150W on laptop CPU. While 10nm is 100units. Roughly 3X area.
@@vmafarah9473 The cpu nm process measurement has not actually represented the exact size for quite some time. Going from 6nm to 3nm or 4nm is really not that much of a size difference as you'd think. NM is just a buzz word for chip makers.
@@bansheezs 4x4=16, 5x5=25, thats 56% less area roughly, they can pack more transistors than before in that area, clocks are bit different, they can get better with using fine tuning older nodes, but effficiency wont increase much,
7945hx on 5nm performance in cinbenchr23 at 70w is barely matched by intel at biggest laptop chip 13980HX at 150w .
That ays alot about efficiency.
stcking to the same old 10nm+++++ fine tuned chip will give you diminishing returns , only apps that rely only on single core can benefit from that .
lower node plus chiplet design can make more chips out of a single wafer than a large wafer with more defective parts .
the 8 core AMD 7940hs 4nm can match perform better than 14 core 13700H, and matches 16 core 13700HX .
Im currently using an intel system , but I should give props to amds achievements using tsmc advanced node used in phones .
I vote for my wallet and performance and buildQ inside to outside, then battery life, is what I always look in laptops,
@@vmafarah9473 The nm is supposed to represent the size, but it actually doesn't anymore. Its just a catch phrase to represent the next generation of chips. The make improvements here and there, but there are somethings they can't shrink
was the clock speeds the same between the cards?
This is like investigative journalism for hardware. Bravo!
so the real change in RDNA 3 is really addition of AV1, slightly better power consumption (from smaller node) and more CU from MCM layout?
Good informative video. Thx
Glad it was helpful!
They have a quiet setting... Does that limit the power?
Right on the money. I think where I'm settled at is, since I'm not really happy with any of the current GPU market offerings, I'll see what's available used, or just buy literally the best bang for buck if it offers more performance than what I have. So for me, that's probably as 6700XT or a 6800 if it makes sense due to the 16GB of VRAM. Only way to get that much from Nvidia is a 3090, ti, 4080 or 4090 all of which I am not that interested in, just the cost really.
I love my Powercolor Hellhound 7900xtx. Going to go play some games with it. TTYL :)
How are you limiting past -10%?
Just look at 7900gre that is 260w but it is same gpu as 6900xt 300w or 6950xt 335w so in the end its 50w less for same performance same money and same hardware confi........Did it improve anything? Maybe a bit RT but thats it, it does have 2x transistor count and 2x TFlops but it does not deliver it in FPS....Maybe that 54% uplift is like 54% more transistors at same or a bit less wattage but thats only 6900/6950 26.8B vs 7900gre 54B not 6600xt 11.1B vs 7600 13.3B.
5:15 explains everything.
This makes a lot of sense. Intel did something similar when they said a 13900k at 65W can match a 12900k at 242W, although thankfully they didn't say it was a 300% improvement in performance per watt.
Could you check the curve for the 5700 XT vs 6700 XT?
For whatever reason, AMD MASSIVELY overvolted the 5700 XT out of the box for no benefit and the 6700 XT was a lot more reasonable. You could literally drop nearly all 5700 XT's voltages by nearly 10% and gain performance from it. It was absurd. The 6700 XT genuinely may be 40% more efficient out of the box.
to be fair rdna has seen little to no ipc increase since rdna 1. hub compared a 5700xt to 6700xt at same clocks and it performed pretty much the same and as we can see now rdna 3 performs like rdna 2 with th same cu and power draw. really rdna 3's only reason for better performance at all is they made a bigger die. honestly rdna 2 on 5nm with the same cu counts at 7900xtx would probably be the same performance
While I agree that when looking at "up to", you will never actually get that much, this video is simply misleading. The 7600 is a monolithic GPU on a 6 nm node, which is completely different from the rest, which are 5nm gpus with chiplets. The 6nm node is nothing more than a more optimised 7nm node, which should only really give you 5-10 more% performance per watt. It's the 5nm node which is new, it's the node that makes zen 4 so efficient and would surely get better power efficiency compared to RDNA2.
And this video showed RDNA3 in the 7600 is more power efficient than RDNA2. How is demonstrating your claim misleading?
@@ImaMac-PC If you used the higher SKUs that use a better node, the power efficiency would be far higher as compared to last gen, and the claims of power efficiency by AMD wouldn't be so misleading, even if they'd still be.
Just undervolt an AMD Gpu,s and you good to go!👍
yayyyyy
Thank you for doing your test. 3:13 I've been claiming that AMD promises where made at optimal specs that their gpus should be runing instead of wasting watts for no reason...other than the ones you point out. RX 7600 needs to be limited to RX 7600M XT specs of 105-120W. Or somewhere between 90 or 100W to make a better W7500 RAdeon Pro (not limited to 70W but delivering energy efficiency equal to that, for half the price but also being much faster), W7600 RAdeon Pro is another gimmick, becasue you can tune a regular RX 7600 to do better and cheaper as shown in your video. They came too overclocked at stock spec...
Your welcome. I do wonder why Radeon is not more successful in Laptops.
@@ImaMac-PC RX 6600M was so successful at laptops that you can count with your fingers latops and minipc with them on sale that arrived to the market. 6600M at 100W delivered so much better work/Joule that asian cryptominers bought'em all...where do you think all that aliexpress rx6600m reporpused as discrete gpus come from?
in your tests 6600@84W delivers 1F/J in 3dmark the same as 7600@96W... but some years ago... no way Samsung node RTX 30 could achieve that energy efficiency, the same as RTX 40 is better if reclocked revolted to a optimimal zone... at stock (or slower max TDP limited below optimal) energy consumption goes up at 0.6 F/J, similar to RTX 30. FRame capped scenerarios are better for energy savings for this reason... over uncapped...
there's no need to check. when "up to" is included, it's never a coincidence. it's code for "this is the best case scenario possible"
Yea, I'll be looking for those two words in the future. Those two little words need to be emphasized and not ignored. Do you see how they use small font and different colors on those words in the presentations so they are not that noticeable?
It didn't feel like a 50%+ gain before I saw your nice analysis. Now it really does feel like a lie. The sad part is that I'm not surprised as I doubt there is a single tech company that actually reports truthfully to the consumer. I'm not even sure they report truthfully to their investors.
The only way to really calculate true efficiency would be to lower each card down to their peek and compare.
can anybody tell me why my mem clocks stay at 2487 all the time on 7900 xtx . my first amd gpu in over 15 years?
Have you activated variable refresh rate in win11? how many monitors do you have? whats their refresh rate?
@@samserious1337 using 2 monitors and im still on win10 for the time being , will upgrade in a amonth or two when i upgrade cpu
@@samserious1337 new adrenalin software . can i do it here?
@@samserious1337 MAIN MONITOR IS 165HZ AND SECOND IS 60 HZ
@@peterpeter5666 You may wanna read upon "Enabling VRR may lead to substantial reduction in Radeon GPU idle power consumption with high refresh monitors" and update your driver. GL
If anything this video is the misinformation, the slides shown talk about the 7900 xt/xtx and the testing is done with the 7600 which isn't even on the same node. The 6950xt reports 335w usage with the red devil 6950xt but in reality is pulling 400-500w as RDNA2 does not display full board power, meanwhile the 7900 xtx displays 400w while actually using 400w and being considerably faster. Secondly, It's already been covered that when those slides were released, a bug had been found days prior and the cards were downtuned slightly, they're working to undo that downtuning through driver updates.
Please list what you believe to be misinformation. Also, what bug? 8months later and they haven't solved "the bug"?
@@ImaMac-PC Comparing RDNA3 vs RDNA2 with the 7600 while showing slides from the 7900 cards which aren't even on the same process node. That's just as bad or worse than what you're accusing of. A 7900 xtx uses about 100-150w less power on average vs a 6950xt while being a solid bit faster, it might not be what the launch slides showed, but it's a lot more than 10% increase, probably somewhere in the 30-40% increase in frames per watt. I've had both top end cards. And that's just on average, not all games take advantage of the RDNA3 shaders, some do, a lot don't.
@@ImaMac-PC Also, the issue that they ran into was an artifact issue that occurred during extended gameplay, it was leaked that there was a delay added into the driver that hurt performance and that the slides were made before this "fix". This is already known information, so yes, you're either an idiot who doesn't have any insight into the issues or a clickbaiter. Take your pick.
Famous math: v1 driver = 80 fps, v2 driver = 81 fps, v3 driver = 84 fps ===> v3 brings a 300% improvement vs. v2 (in small letters: when v2 is compared to v1)
👍
So glad I've found this channel which is based on honesty, facts and truths, unlike Moore's Law Is Dead, Ancient Gameplays or Red Tech Gaming who are living by pumping copium for AMD fanboys. Been saying this all the time, AMD has been cocky since the release of RX 6500 XT and overconfident and RDeadOnArrival 3's messy release and the cancel of high end products in RDeadOnArrival 4 lineup is the consequences of AMD Copeon Terrible Group's own undoing. They've been trying to charge premium money for their underdog products as soon as possible while providing no premium features or siginificant performance improvements at all, which makes them no better than Nvidia at this point. Under such circumstances people should just go buy GeForce GPUs instead. RIP AMD🎉
Thanks. Data tells the story. AMD execs have a much higher opinion of their products than the general buying public. AMD doesn't appear to be fooling anybody if you look at market share and Steam surveys. People can try to discount both of those however, it's a long term trend on both and there is no other data available to counter those.
Your investigation work is refreshing. I knew the 54% performance-per-watt RDNA III claims were a lie, but with the RX 7600 vs 6600XT at stock settings, 6.7% is a joke. This is why people can't take AMD seriously. Do better, AMD.
Thanks and yes, AMD needs to do better. They have in the past. Not sure what has happened with their leadership. Makes me wonder why Robert Hallock left AMD last year...
@@ImaMac-PC I think the corporate suits have accepted NVIDIA's market conditions and will not try to overthrow NVIDIA's dominance in the market. Also, I wonder if this is why Raja Koduri, a controversial character, left Radeon because he knew that AMD's top management surrendered to NVIDIA, and the underwhelming RDNA III was a disaster.
On the topic of Robert Hallock, I guess he knew that AMD peaked, and it's a tough road ahead with a resurgent Intel and an angry NVIDIA after RDNA II caught up to Ampere. I don't know if you heard, but on August 8, 2023, Hallock joined Intel as Senior Director of Technical Marketing to help market their AI products and CPUs.
to be fair. ppl who care about ppw undervolt and that 300w 7900xtx isnt unrealistic. what you should do is find the sweet spot wattage of RNDA 2 & 3 and then compare them.
Isnt that RX 7600 on 6nm ?
and this is why I waited for 8 years to build my new computer
It is impressive tbh, just they sold it really wrong, if they sold it at 100w rdna 3 would have looked way better n thats on 6nm not 5nm
At what point do these people get fined for fraud
When investors lose enough money. Unfortunately, the sales volume of discrete GPUs from Radeon is a very small percentage of AMD's overall sales figures.
I don't think that's a big issue, we all know it's a bunch of lies, the performance claims were much worse. Also although no one sells a card with those configurations u can buy one and lower the power target to match the more efficient one. I did that with my 3090, it boosts to 1800-1900Mhz at stock but I lower the voltages and max out at 1785Mhz so it ends up using 250w most of the time instead of 350w
Bravo on that 3090. My 3090 is also a power hog and most of the time, I run it at 1800MHz (using the 5% rule, that performance drop won't be noticed...and it hasn't) and under volted and it now runs around 300W and under during most games.
rx6800 is AMDs perf. per Watt king... we will se how rx7700/800 will do against it..
AMD graphics needs constant efficiency gains with agressive price like how ryzen did since 2017,thats how you can climb and hold your share at 35 to 40% despite your competitor is giving in everything ,like how ryzen is doing, we need constant rdna 2 moments,nvidia is becoming haswell generation intel
RDNA2 seems impressive because nvidia are using inferior node from samsung. but once nvidia using the same node as AMD (TSMC 5nm) AMD no longer hold advantage against nvidia. and Ryzen are winning because of TSMC node as well that is superior than intel node. nvidia Ada Lovelace if anything is the repeat of pascal without pascal great price. if push ada the same way as pascal (aggressive in both tech, power efficiency and price) 95% market share will be a reality for nvidia.
@@arenzricodexd4409 then why zen 3 destroyed both 10 and 11th gen despite using the same 7nm node as zen 2?
@@pham3383 because intel still stuck with their 14nm+++. Zen 2 to Zen 3 is more like intel old tick tock strategy. Zen 2 is the first transition to 7nm which can be a lot of trial on newer node. So they usually did not go all out. When they got the experience on 7nm with Zen2, the next step is try to push the architecture as far as they can on the same node.
@@arenzricodexd4409 seems like justifying for the makers to cut cost by using inferior node,with more heat
@@pham3383 it is not about cut cost. intel simply does not have anything better making them stuck at 14nm for a very long time. in the past the always ahead. for example if TSMC latest node is 40nm then intel already at 32nm.
truthfully they should have just shrunk rdna 2 to 5nm and made minor tweaks to that. my 6900xt clocked at 2700mhz stable, im sure on 5nm with some minor tweaks, faster memory and a bigger die they could have matched rdna 3 with way less work. rnda 2 was great rdna 3 is trast. and i literally own a 6900xt and a 7900xtx this isnt a hater talking. i run all amd builds and have for quite some time. rdna 3 is a let down. it was definitely supposed to be a 4090 competitor but due to power draw issues they cant get it to clock high enough. at 3ghz all of these cards would look impressive compared to last gen but at the same clocks as last gen we got what we got
I agree, they should have shrunk RDNA2 onto 5nm. I do wonder if it would have clocked higher more consistently. I can get the 6nm 7600 to 3GHz, but only under "lightly loaded" conditions while my 7nm 6600XT maxes out mid 2.7xGHz.
@@ImaMac-PC yup im sure a cheap rdna 2 refresh would have been fine to compete with rtx 4000 honestly. the only thing i could think is amd needed a stepping stone to get multi die gpu's running so even tho rdna 3 is shit they just are saying fuck it? idk really. i have a 7900xtx red devil with a block and it will do 3ghz in some games at dam near 500 watts but it needed to hit 3ghz in every game at 350 watts and it would have been competative but as previously stated rdna 3 didnt hit power or clock speed targets and thats why it looks so bad. rdna 3 kinda deserves a rdna 3+ with hardware tweaks and maybe 4nm but honestly i dont see why they would do that since consumer gpu sales are down overall. rdna 3 per sku really needed to be 15% faster minimum than what we got. but as it stand the only thing that makes them faster at all is that they have more power at their disposal the rdna 2 parts at the same power arent far off
And this is why Moore's law is dead.
RIP 1929 - 2023
ampere is hyper efficient, 3060ti vs 1060 at 150w isnt even close its stupid lol
I mean on one hand I trust neither company. But I know the 7900xtx is selling below $1000 and AIB 4080's are over $1200 and not as strong. I have gotten AMD recently because they're better value.
It's the best performing, under $1,000 GPU, in the world!
It's like they'd rather lie than try, as long as investors look at them and think they are making a product competitive with Nvidia but just a little bit cheaper, I don't think they care. AMD is really good at making at seem like Radeon only exist to make it appear like they're a player in the gpu market. I know people who work there work hard and make a decent product but AMD and their marketing suck at this.
AMD and their Marketing has absolutely no idea what they are doing with Radeon.
So they straight up lied... wow
at 101 FPS the RDNA 3 is using about 112 watts and the RDNA 2 is using 190 Watts which is a change of 69.6429% or at 102 FPS the RDNA 3 is at 113 wats and the RDNA 2 is at 197 Watts which is a 74% increase power usage for the same performance so the RDNA 3 is much better with performance per watt. Thats the max for the RX 6600. Thats about the best it can do and it takes 74% more power to reach that point than the RX 7600.
RDNA3 is a failure
2 years - 0 progress
That's what happens when you spend all your R&D on chiplets 👍🏻
@@100500daniel Which Navi 33 _isn't._
RDNA3 would've gone down as the worst architecture ever, but AMD already bested that with Vega. I'm noticing a common trend here...
At least with Vega, AMD didn't present misleading information 6 weeks prior to launch. With Vega, they launched Vega Frontier 6 weeks before RX Vega and we already knew how that was going to go...but yea, Vega, Vega II (aka Radeon 7)
I know the technical term is misinformation, but I call it a lie.
I think everybody know at this point, that AMD isn't better than Nvidia, heck they probably know it as well and they shouldn't be, they just need to be cheaper to win gamers heart, since in most cases people probably disregard power efficiency and go with cost to performance, sadly though with RDNA3, they opted the Ngreedia route by upselling their gpu way above actual MSRP, the recent price cut of 7600 just further proved how greedy they also becomes
how dare you lie to me AMD it's almost like you're on of those corporations we've all heard about 😢
The web of lies! It's all a lie. Also the pricing is a joke and a living meme, so they are lying clowns.
Let's just hope that Intel can bring something great to the table with Battlemage. 4070 performance at 400 usd and 4080 performance at 700 usd would mix the market and poison Nvidia's and AMD's marketing. We need proper competition on the market, to drive innovation and better pricing.
So AMD is Misleading and you do not trust them , then you Go buy Nvidia 😒 which is also Misleading 05:55 ... Broken logic
RDNA3 is just a massive flop.
7600 with exactly the same specs as the 6650xt and 3% faster memory clocks is 3% faster.
7900gre with exactly the same specs as the 6900xt performance almost identical
Soon we will get 7800xt that will perform as the 6800xt. They made nothing over the last 3 years and just put RDNA2 into smaller dies thanks to a smaller process and that's it.
AMD fucked up with RDNA 3 and they might as well just have skipped it and only made the 7900 XTX as the 7800 XT to compete against the 4080. That way their lower end cards wouldn't suck so bad because CU to CU there's almost no improvement whatsoever,and the TDP has been increased across the board. The 7800 XT would've been actually decent as a 7700 XT, and the 7900 XT should have been the 7800.
That's probably what they've figured out now and why there are rumors that RDNA 4 wouldn't compete with Nvidia's high end.
It's actually incredible how bad of a jump RDNA3 is compared to how good RDNA2 was. There's barely any improvement CU to CU, so you'd expect them to up the specs in order to make a generational jump between cards. NOPE.
RT is still bad (ironically they've just figured out how to make it better AFTER RDNA 3 launched) you get AI cores that 9 months later still aren't being used for anything, FSR3 isn't out yet,
In all honesty,the Radeon department is at it's bottom since Vega ATM.
It's probably due to them moving away from the monolithic design or what not, it's just sad honestly.
Best value gpu will be 4080 20gb + new connector
The 7700 xt is one of the worst AMD GPU launches of all time. The card is literally only 10% more efficient than Ampere and costs $450 for what is essentially a 3070 Ti.
I'm sitting here with my undervolted 3060 Ti with practically no options that realistically make sense for my needs. I have to choose between the only sensical RDNA 3 card, the 7800 XT, which has a 265W TDP (??????????????), and a $600 RTX 4070 with trash memory bandwidth. This generation is offering NOTHING to people who want to upgrade to an affordable, small and power efficient card like the old Maxwell/Pascal 70 series cards.
RTX 5000 is rumored to launch as late as 2025. We are stuck with these absolute disgraces of GPU launches from both AMD and Nvidia for potentially another TWO YEARS.
Your performance per watt calculation is flawed. I knew that When I saw Ur CPU ppw video .
Is AMDs performance per watt calculation flawed?
amd should be sued for misinformation