this is also why servers and chips that run for a long long time (think servers that aren’t decommissioned for years) don’t run such high voltages and clock speeds.
What I would want to know is what happens to these good binned chips when you lower their voltages and frequencies and run them at a more reasonable 3.5GHz or something for a long time. Is the price / power / performance ratio better than a server chip or is it worse ? Would love some data on that. But no one is buying these for it and server chips have different things going on for them like being able to replace the chip for years etc. still would be interesting to find out if these the best ‘quality’ chips out there
> What I would want to know is what happens to these good binned chips when you lower their voltages and frequencies and run them at a more reasonable 3.5GHz or something for a long time
I'm not sure about intel chips, but amd's last couple generations of chips have an "eco mode" which underclocks the chip for you. You get about 85-100% of the performance while consuming 60% as much power. The chips probably last way longer like that too. NVidia's GPUs have similar options. 3090/4090 cards can have a totally reasonable power budget if you're happy to lose ~10% of your framerate.
Where are you getting that ECO mode cause a 15% performance drop?
In my pretty unscientific testing, I saw ECO mode have zero impact on single core workloads and ~4% on multicore with a big drop in cpu temperature.
If I am reading the “Gaming CPU” chart correctly from your link, the AMD 170w performance was 114.8 FPS, but the 105w ECO performance was 111.4 FPS. Effectively nothing for a huge drop in power.
> Where are you getting that ECO mode cause a 15% performance drop?
It depends on the test. In that link I posted, 3dmark’s cpu test shows eco mode drop the score from 19000 to 14000. Which I guess is more like a 25% perf drop. In other tests, performance increased in eco mode.
Was this on a laptop or desktop? If a huge drop in power results in no meaningful drop in performance I would say that is a sign that previously you were hitting a point where heat was more of a problem than power.
My testing was on a desktop with a way over specced cooler. Regardless, the above link shows that they were getting huge power drops for basically no loss in performance.
With makes sense to me. We have seemingly run out of cheap architectural wins. CPUs and GPUs keep cheating by amping up the power draw to improve the performance numbers for vanishingly small returns. To get an extra 5% performance on these chips can require 10s of watts.
Thanks. The 7950X doesnt have a non X variant. So it seems the non X variant chips are basically doing the same - giving you lower clocks for lower wattage with an option to tinker at your own risk for pushing the chip as those are not as good as the X variants.
I design chips in similar technologies. What these binned parts are fast-fast parts, meaning both PMOS and NMOS transistors have lower threshold voltage than typical due to stochastic process od production or intentional skewing of the process during fabrication. You can run them at faster clock speeds. If you stack higher voltage on top you can run quite a bit faster. For a while.. The transistors age. Aging is an exponential function of voltage level and temperature, and lunear function of clock frequency. With aging the device threshold voltages increase rapidly, so the devices get slower. On top of this there is electromigration (EM). With EM the interconnect resistance increases. These effects combined, you have a horrible product lifetime at those cobditions. After failung to work at 5-6 GHz at 1.4V, it would most likely still work at a lower clock frequecy, because in essence it becomes a slower chip.
Now the answer your question: A fast binned part operating at nominal conditions will perform exactly the same as the same sdries typical chip. The power consumption would be like 10% or so higher due to higher dynamic power and leakage. The lifetime would be muuuch more longer than a typical device though. So, in short, wouldn't be outperforming anything.
What I would want to know is what happens to these good binned chips when you lower their voltages and frequencies and run them at a more reasonable 3.5GHz or something for a long time. Is the price / power / performance ratio better than a server chip or is it worse ? Would love some data on that. But no one is buying these for it and server chips have different things going on for them like being able to replace the chip for years etc. still would be interesting to find out if these the best ‘quality’ chips out there