RX 7600 Efficiency vs RTX 4060 and M2 Max
Вставка
- Опубліковано 31 тра 2024
- With AMD's RDNA3 architecture, can it be as efficient as the RTX 4060? How does it compare to Apple Silicon in performance and efficiency?
RDNA3 Performance Playlist:
• RDNA3 - Performance - Наука та технологія
![[柴犬ASMR]曼玉Manyu&小白Bai 毛发护理Spa asmr](http://i.ytimg.com/vi/0TsXQ7z2Dh4/mqdefault.jpg)
When an M1 Max costs $300 I'm in.
MacStudio at 2500€ is cheap for a 24/7 non-stop operation. If you hire 2 partners for using your same Mac doing shifts... in 2 years you will save enough electricity to pay for your MacStudio. Keep it 2 or 3 more and you may end with almost 4000€ more in your pocket that with a PC with a 7600 costing only "300€." But way cheaper is to buy a M2 MacMini basic model for intense work and very light gaming video processing etc. And upgrading whatever pc you have with a 600€ 4070 or a 300€ 4060 just for gaming. If you game 2 hours a day and work with your partners 22h per day / 7 days a week, by de second year you may have saved enough to buy a MacStudio...
@@InternetListener Multiply 4000 by 250 and you've earned a million!!!! The problem is no matter what machine you use, the energy costs are absolutely nowhere near as much as you are saying. I estimate a maximum of 80 euros extra per year when using it full time.
@@peanutnutter1 those are aproximated numbers of course, but more realistic than you think. You may not get rich by setting up a farm of Apple silicon render/video/etc. but mainly because server clouds are already using arm based and gpu-like optimized hardware in the form of GraceHopper and Ampere Arm and even AMD optimized data center solutions MI300X have additional hardware solutions and efficient configurations to make CDNA3 competitive enough.
For doing estimation you need no take into account not only price per kwh but taxes and full billing price but also full system power measured at the wall. a faster gpu, means also more display, ram and cpu load...and it all adds up.
In Spain or Portugal is safe to say 24/7 of 150W mean 260€/year, it's true than one office 1500W 24/7 it's not going to cost 2600€/year, because some cost may be constant and only variable part is going to increase... but you can do accurate simulation of billing cost if you want on online electric bill simulators using actual data...if you want the links for Iberic energy market...
Incredible efficiency increase without even undervolting. My 6600XT runs at a maximum of around 85W just by slightly undervoting and pulling clocks down to 2300 Mhz… cutting off the top 200-300 Mhz that cost a lot of voltage/wattage and just fluctuate to the point of uselessness.
The power limit I kept at 130W…, but it doesn’t need more than 85W. My problem with the RX 7600 is that there is nothing ITX-sized (
Clocks are basic in RX cards to go fast to the efficiency zone, if you can -10% voltage curve + -10% clock you may get easy -27% W, for just a lose of -10% of performance. But if you want to find the +54% efficiency (or -35% power, which is the same) you need to lower voltage curve and clock a little more -13%, et voilà... but there is more efficiency hidden in your RDNA3 card... instead of using game/boost clock of RX 6600 specs to save a lot you may like to try even lower and put official game/boost clocks of RX 5600 XT, yo may end with a RX 7600 doing next to 35W (to 50W depending of undervolt) for a very fast RX 7500 XT/RX 7400 faster and why more efficient than the ones they are planing to launch...differences between power of RDNA1 and RDNA3 may be over -58% power (or +138% efficiency).
In reality power is not the way to measure efficiency, but energy per operation (frame, render, laps of video encoding...): average frames per second divided by tbp of gpu (or way better total system usage, because you can't render frames withou ram, cpus, hd, display...) in W, gives Frames per Joule which is a real metric of the energy being saved. Only if you work with uncapped at full load you may save the same energy than power, but if you render a scene or process a video you'll spend more time...
in this case you may want to know that undervolting/underclocking to -73% max power, will save you -58% energy in fixed size production work. This means -35% Voltage curve, -35% max boost clock, or find the minimum combination of two. But you'll need +54% time to finish the work. if you further reduce clock over the minimum clock your silicon is able to work with you may only save energy because you are limiting max power but you're not increasing efficiency, hence your battery will last more but may reduce its capacity the same amount when the work is finished.
Voltage clock Power Energy Time
100% 115% 115% 100% 87%
100% 100% 100% 100% 100%
95% 100% 90% 90% 100%
95% 95% 86% 90% 105%
90% 90% 73% 81% 111%
85% 85% 61% 72% 118%
75% 75% 42% 56% 133%
70% 70% 34% 49% 143%
65% 65% 27% 42% 154%
65% 50% 21% 42% 200%
M1 Max is great. The thing is, besides the price, that you can't swap any of the components for future upgrades or trouble shooting. You're stuck with whatever configuration you bought in the first place.
Apple does this on purpose bc they are an awful company and want you to rebuy everything again
It also should be noted that power consumption is only relevant in the context of heat dissipation and power delivery. What determines your electricity bill is energy consumption. If a processor uses 30% more power than the previous generation but is in return completes the task you gave it 30% faster, you're not going to see an increase in your bill. If it uses 50% more power but is 2x faster, the energy it consumes will be 50% lower.
Only in time bound workloads. In something like gaming, absolute efficiency matters
I also dont understand where is the 54% perf/watt uplift? It consumes ~20% more power than the rx 6600. So it should be at least 60-65% faster then the rx 6600. Is there a way we can launch a court case against AMD where they will be fined for a few billion and learn their lesson?
it doesn't work this way, it's not proportional or inversely proportional related perfromance and power, neither variables afecting perfromance and power... AMD and NVIDIA so overclocked, overvolted and out of the optimal efficiency area (if exist for some specific application taht may perform also non-linear with raw perfromance) that it may be there somewhere but only used in server, studio or if limiting yourself by testing specific game/app voltage and clock. It's the same as overclocking and overvolting but now they already come stock almost at the limit, delivering almost max perfromance when you ask for it... you may want to test 1080p medium at 60 fps in RX 6600 vs 7600 or vs 5700 and may found the +54 perf/watt, remember that power is inversely proportional so you have to expect 1-1/1.54=0.35 saving on max power. Where a 6600 needs 120E your 7600 may need 80W, if uses 165W for +30% perfromance is expected behaivour... it's like offers claiming 2x1, you are getting -50% discount not one 100% free, 3x2 hence mean -33% discount, not 1 free, you may need to pay 2 full at asking price...
you must do the same test with the rtx 4060 ti lower its consumption until it consumes the same as an M2 Max and see how it performs with the same consumption of an M2 max
Exactly, more cores running at lower power will increase the efficiency. I'd be curious about how a 4080 at 145W compares to the M2 Ultra as well.
@@ehtasam806 Only in a fanboy's dreams. Yes there are scenarios where the 4090 is ~4x faster than the M2 Ultra such as 3D animation performance in Blender, but that is the best case scenario for the 4090 since Blender's renderer makes use of Nvidia's hardware ray tracing. In pure rasterization performance, in Metal-optimized cross platform benchmarks like 3DMark and GFXBench we see the M2 Ultra being anywhere from 16% behind to about as fast as a 4080. If Apple adopts some form of hardware ray tracing acceleration with the next generations I can see them closing 50% of the gap and being 2x slower than the 4090 at significantly less power, and the rest of the gap being closed by the 4-die M Max chip should that come to market.
Great stuff. Keep on this comparisons they are really great.
Comparing to MAC chips, RX7600 is not tunned to be efficient at low power. Better comparison would be M2 against 780M.
Thanks. The RX 7600 desktop does have a couple of counterparts in the laptop space as RX 7600M XT and RX 7700S. They all have 32 Compute Units. However, the laptop chips are spec'ed as low as 75W and can go as high as 100W for the 7600S and 125W for the 7600M XT. Seems to be the "sweet spot".
I've wanted to see testing like this for a very long time. So many GPU's have just meme tier efficiency out of the box, sacrificing power, temps, noise and even performance in some cases while ramping the power up.
5700 XT is disgustingly overvolted out of the box, with a voltage that's too high even for stock clocks. You probably could cut power in half and lose like 15%.
I agree, these practices are disgusting. I can run about 15% less performance for 60 or more watts less, just by reducing the gpu clock, hence the voltage. I already undervolted heavily from stock 1.2 to 1.03 by using morepowertool.
I know it's silicon lottery but i'd like to see undervolt results too. Gonna watch the 7600 vs 6600xt video for sure! 😊
After the 6600, chart out the 7900xt/6950xt power please
Main problems with this video are the node difference, silicon tier, single source of comparison being benchmarks and voltage curves
M2 Max node is 4nm while its competitors are 7600 node with 6nm and 4060 with TSMC 4N or 5N* not even apples to apples comparison or iso-equal comparison there's no way to do apple to apples in node comparison but you can achieve almost iso-equal comparisons
Another problem from the node is the node scaling silicon tier
M2 Max is second tier from the M2 series (M2 Ultra>Max>Pro>Base) from the ultra while 7600 and 4060(4050) are all the bottom tier from the respective series thus only way to make equivalent comparison in tier would be to compare against 7900xt(7800XT) and 4080(4070) thus video compares the 4nm 2nd highest tier silicon vs 5nm/6nm* with lowest tier silicon same could be said about silicon transistor count
Single source of benchmark is not representative or does represent entire performance thus real life tests of actual games/applications are required this only amounts to cherry picking if you only do single source
Another problem is that Apple silicon run their hardware below 1.1-1.2 V realistically 0.5-.09 rarely above 1.2V meanwhile the x86 hardware typical run their hardware in 1.1-1.4V ranges if you lower the voltage for the GPU the FPS curve will shift at higher values compared to the norm even for the 7600
Make a new video with higher tier GPU like 7800XT/7900XT or 4080/4080ti with power and voltage limit to same as apple 0.5-0.9 V against M2 Max then you will find that there's not outrageous claim of x5.2 difference rather within margin of error values
The 4060 looks so efficient because this is literally the 4050 on a tiny 107 die just put inside the 4060 box.
The problem is that the top-tier RDNA 3 cards competes only with a 103 die (not 102), 7900xt competes only with 104 (not 102) die, and 7600 competes only with a 107 die (not 106).
RDNA 3 is a massive flop in performance because these cards should be 30-40% faster than they are with the same power consumption.
7600 has exactly the same specs as the 6650xt and is 3% faster in raster AND in RT so even RT wasn't improved and is using like 5% less power. This architecture is a disaster.
I wonder if the M3 Max will be like the 4070ti or 4080.
You focused on the power !!
I do read the comments and after thinking about it and running a couple of tests, I had to make this video.
@@ImaMac-PC there is probably a much faster and proper 7500 XT hidding on the RX 7600 silicon that may consume at max power less than a RX 6400, look for it, I'm some other comments I may offer some tips...
First hidded GPU you may find with little undervolting is a 7600M XT, 100% performance at only 120W, I'm guessing you play the silicon lottery to -14% Voltage curve, and reduce -2% clock...
54% per "what?" is great.
You'd have to look at some of the mobile GPUs like the RTX 4070 mobile, and RX7600M - if not the m1/m2 max are going to have a much bigger spread when looking at fps per watt. The other thing is that game engines are watering down for the apple silicone (much like they do for consoles) so even the exact same game on both systems is not tit for tat.
Wouldn't an APU be more fitting to compare to than a solo GPU?
APUs typically do not have very many GPU cores e.g., Radeon 780M has 12 RDNA3 cores. M2 Max has 30 or 38 GPU cores. APUs will need to have much higher GPU core counts to be comparable.
Apple's APU doesn't have to drive signals across eight GDDR6 interfaces and eight PCIe lanes. Unified memory and coherent caches also eliminate redundant bus traffic. Did that benchmark tell us anything about latency? That's another thing that can be traded for efficiency. By competing against consoles the Apple silicon would lose those advantages because consoles have also made those sacrifices in flexibility.
Interesting video but as far as claiming Apple is generations ahead u could not be more wrong. The silicon isn't engineered to run at those power limits in the first place and if you gave an apple gpu 160 watts you would be getting 0 fps cause it would just fry itself. If AMD made a lip-sticked integrated gpu and stuck it in a system for 4 grand I'm sure it would have similar performance metrics to apple's. Same thing goes for Nvidia.
What the point of comparing massive die like m2 max against low tier products? try it against undervolted 4090 and chech how much it needs to consume to give same amount of performance
5nm vs 6nm, id love to see the mac vs 4060
Stay tuned. There is one LP 4060 that, if they offer at MSRP, I am getting for my SFF build.
apple silicon is good but its arm not x86.
its pretty much a beefy smartphone chip, and windows apps are not compatible with smarphone hardware
Just looks like amd invidia are pulling prices out their bum
Same with Apple. And?
This comparison is pretty useless. You could compare the power draw of a similiar products like the ryzen z1 or a similiar mobile apu (or something like the 5700G). Keep in mind that in silicon design you have to juggle with several competing goals like die size, power draw or perfomance target. Obviously the low-end dedicated gpus are not designed with power efficiency as main goal.
Apple has allways better small less core system that runs better, i think its the software-driver thing.
Isn't M1 max on 4nm? The 7600 is on 6nm. No pun intended but you cant do an apples to apples comparison with a discrete gpu. Although it might be interesting to compare strix if it gets 16 cu rDNA 3.5.
Interesting video. Then again running a GPU at a power profile it was not designed to is not very useful. Usually these are tuned to operate optimally at specific power ranges. We see this exact behaviour when comparing the Steam Deck with the ASUS ROG Ally at around the 9-10 Watt mark. The Ally performs better than the Steam Deck at 15 W but the Deck outperforms the Ally at 9 W.
Consider this - how would the M1 Max perform if we pumped it with power to the 100 W range? My guess is that it would flatten out very quickly beyond its normal power limits. That's of course a hypothetical as we can't actually overclock the Apple Silicon chips to those extremes.
Overall I think the Apple Silicon chips are simply designed for a very different market and are priced accordingly. Apple trades cost for efficiency by manufacturing these at smaller nodes. The 7600 in comparison is still on TSMC 6nm I think. Pure speculation - but I wouldn't be surprised if Apple would need to charge 2x-3x of the 7600 if they ever made the M1 Max as a discrete card. At least back when it was launched - maybe TSMC 5nm has come down in price now.
Apple silicon seems to not go much above 3600 MHz, but there is no need to do so, because they showed perfromance/power curves in release press info available for everybody on their website... they know how much more FPS per watt can deliver and how much more perfromance in general with the chosing of cache, unified memory bus, bandwith, architecture, software optimization, etc... the only thing is they have price evrything not being a MacMini basic at very high prices not affordable for most of the people all over the world, MacStudio at 2500€ vs MacMini 729€..., but reality is that there are no 32 GB vram pc available, 24GB vram mean 1000 to 1600€ (and 3 to 4 times faster gpu, all thing being said) if not you're on the RTX professional well above 4000€ add and 8 or 16 cores system, with same moemory bandwith as the mac ones you would need also +5000€ for server like parts and octa channel or dodecachannel memory ddr5... they are in another league, no direct comparison in general possible...
Comparing an iGPU to a discrete GPU is inherently flawed. Things like GDDR VRAM draw a lot of power.
The 8 CU RDNA2 iGPU in the Steam Deck is pretty close to the 8 core M1 GPU in both performance and power draw.
this comparison would indeed be more fair.
idk how it would pan out tho. Seeing as the chip in the steam deck is 15W and appearantly the M1 max is 45W
would we see the same story where the lower wattage designed chip dominates?
Love your videos
Thank you!
40 generation is a disappointment, much like 20 series with their demo rtx lineup. Everybody knew 20 series were not made for true rtx usage and yet nvidea took a pretty pennny for pretty much useless feature. 40 series is just another money grab as even today most games can not be ran using rtx as the cards are not powerful enough. rasterization also is a disgrace. Nvidea went all out for high end and destroyed mid - low range cards. I myself see no reason to buy 40 series cards and will skip this generation. Tho Nvidea has now to deal with overstock of their terrible cards and will most likely delay new gen cards.
The efficiency calculation is flawed, since, when I looked into cinebenchr23 graph 12700h vs 6800h done by hardware unboxed it seems 12700h is efficient as 6800h at 35w . that doesnt make 12700h efficient as 6800h, by further lowering the wattage say 20w intel cant keep up with AMd and also at higher wattage more cored 10nm intel is efficient but if ita a ryzen 12 core cpu vs 14 core intel at higher wattage AMD can dominate using the 4nm if we exclude the idling wattage due to chiplet design .in 12700h vs 6800 more cores will be always efficient in lower watt than less core in the same watt. if possible if we take 6800h and 12700h at 20w, the ryzen will win . same number of p core is to be done in such calculation to get better perspective.
Nice review, I'm a Mac! Yep at 9:06, you summarized it really well in wattage to power ratio. Apple Silicon is next level.
Much appreciated!
mb
they are not comparable, not the same architecture. mac systems are using arm (as most android devices for example). WIndows x86-64. Also mac is a closed system, so obviously his own hardware is going to work better. An example of this is Video Game Consoles as xbox or playstation they are more efficient than a pc.
They are comparable when it comes to finishing the same task. For example video editing, I got a project that with lots of Prores 4444 footage. It would be more beneficial for me to buy a M2 Ultra than a windows machine.
Who cares about the m1 and m2 if you need to buy an overpriced system and deal with an OS worse then windows 8.
I wholeheartedly disagree. I buy Mac because of Mac OS. But your comment is way off topic.