Are GPUs ever gonna get energy efficient again or will we have 500w mid range cards soon

>>105975715 (OP)
FIRST WARNING TROLL

>>105975715 (OP)
They are efficient, you retard. Ultimate energy consumption and efficiency can go hand in hand.

>>105975715 (OP)
they're the most efficient they've ever been
lower power consumption doesn't mean more efficient

>>105975715 (OP)
only if someone nukes epic games HQ

>>105975715 (OP)
just power limit your card to 100w (if it's nvidia anyway, amd makes this a pain in the ass to do because they despise you) and it'll still give you far more power than any old 100w card.

>>105977867
>pay premium for the top end gpu
>cripple it so it runs like a budget card

>>105975715 (OP)
How do you evaluate energy efficiency OP?. Because if its by operations per KW new cards even at the same 75w would absolutely destroy old cards. Or are you a jeet and just have the average indian IQ of 75?

>>105977867
Works on my 9070 machine.

>>105978314
>crippling your high end card for literally no reason
>voltage spikes, lolz
What's the point of you?

>>105976014
Space heaters are 100% efficient.

>>105977867
>amd makes this a pain in the ass to do because they despise you
Care to elaborate?

>>105978314
What software, and how's the experience? I was looking at the 9060 to do just that in my small-ish everyday and light gaming build.

>>105978915
Not sure if it's the same thing but a while back the amdgpu module for the linux kernel was changed because AMD decided that it should not be permitted to reduce the power draw of a card lower than around 95% of the rated TDP.
Pretty sure they removed that artificial limit in newer versions though so I don't think it's an issue anymore.

>>105979019
https://www.phoronix.com/news/AMDGPU-Lower-Power-Limit
>matching the behavior of the AMD Radeon Windows driver
finally admitting that all linuxtards want is to match what windows does

>>105978932
LACT. I use 245w with -100mV though, just showing it's possible to go way low. Slider goes all the way down to 0w actually.

>>105979346
>Slider goes all the way down to 0w actually.
w-what happens if you do that...

>>105979518
GPU disappears

If I want to optimize for performance-per-watt what are my best options for a local LLM rig? I was thinking CPU maxxing at first, but I notice that the Epycs recommended are actually really high power draw and they lack the idle power efficiency that GPUs and consumer CPUs have, so they seem like a bad idea for my use case. Should I go for Macs or something?

>>105979990
>optimize for performance-per-watt
If I recall right, a post mentioned the 7800X3D being the most power efficient CPU at the time. If so, that might be a starting point.

>>105979990
Under volt an AMD X3D CPU and an RTX 4090 or 5090 if you can find one in stock
>but how do I undervolt
Watch a youtube video

>>105975715 (OP)
Pushing power limits will ALWAYS be a thing.
It doesn't matter how efficient something gets, the manufacturer will push it to the limit to get as much power out as possible.

The idea they should be energy efficient at all has never made any sense, it's always been a lie to sell you less silicon

>>105975715 (OP)
> Across gens, efficiency (TimeSpy score ÷ TDP) steadily climbs.
> GTX750Ti: ~1.5k score @60W 25score/W
> GTX960 dips to ~2.6k @120W 22score/W outlier
> GTX1060, 1660, 2070S, RTX3060 climb up (~35, ~?, ~63, ~48)
> RTX4070: ~17.8k @200W ~89score/W big jump
> AMD RX580: ~4.4k @185W ~24score/W
> RX6600 & RX7900XTX legit hit ~68 & ~86score/W
> Outliers: GTX960 underwhelms vs 750Ti, RX580 drags (~24), and HD7850 is abysmal at ~12score/W
3dmark

>>105975715 (OP)
moreover
> Flagship efficiency using 3DMark Time Spy Extreme:
> Kepler/Maxwell (GTX780Ti): ~4.58k score @250W ~18.3 score/W :contentReference[oaicite:0]{index=0}
> Pascal (GTX1080Ti): ~4.66k @250W ~18.7 score/W :contentReference[oaicite:1]{index=1}
> Turing (RTX2080Ti): ~7.02k @260W ~27.0 score/W :contentReference[oaicite:2]{index=2}
> Ampere (RTX3090Ti): ~11.3k @350W ~32.3 score/W :contentReference[oaicite:3]{index=3}
> Ada (RTX4090): ~19.5k @450W ~43.3 score/W :contentReference[oaicite:4]{index=4}
> Blackwell (RTX5090): ~25.7k @550W ~46.8 score/W :contentReference[oaicite:5]{index=5}
AMD flagships:
> RXVega64: ~4.70k @295W ~15.9 score/W :contentReference[oaicite:6]{index=6}
> RX7900XTX: ~14.93k @355W ~42.1 score/W :contentReference[oaicite:7]{index=7}
> Trend: steady efficiency climb each gen (~18 ~47 score/W), with major gains starting from Turing onward.
> Outliers: early flagships (780Ti/1080Ti and Vega64) severely lag, RTX3090Ti ups only moderate (~32), while RTX5090 sees diminishing ROI vs 4090.

>>105981524 >>105981548
Even AI thinks you're stupid, OP.