I can’t help but wonder how far you could take the Switch with an external VRM and proper cooling, as opposed to the baby fan and sink it has stock. This begs for a “non-portable” Switch implementation, something for me to fiddle around if I ever come across a nice faulty rope for experimentation, which I often do.
On Erista, not very, everything goes to shit after 2GHz or so on the CPU and probably on GPU after a similar overclock (didn't play around with it nearly as much because I'm not a complete dumbass). With "goes to shit" meaning you just get exponentially higher power draw with little to no impact anywhere else
It's called Copper/Calcio but it'll never reach the public eye
Even on Erista stability and performance is a question of how much power you have, how stable it is and how cool you can keep the chip. You haven’t reached anywhere near peak until the silicon experiences some level of critical stability failure and locks up. Unless by “goes to shit” you mean instantaneously crashing or freezing, there is still headroom there, provided you have the right approach (and even that can be mitigated if you want to go to the extreme). Given the fact that Mariko is effectively the same chip, just made with smaller lithography (with some differences in microcode, I’d imagine), the performance difference can probably be calculated (if it’s linear, and linear is usually a good start for predictions). I suspect that with adequate cooling and proper, stable VRM phases you could push both further beyond. Honestly, it’s not *that* uncommon to see VRM’s completely separate from the main circuit in extreme overclocking - EVGA used to manufacture boards like that for enthusiasts. The cheapskate alternative is to simply cut one out of an existing circuit and reuse as needed, or build one yourself, but that’s not cost effective. Regarding the measurement of “impact”, that’s hard to gauge on console - the games are designed for specific hardware, after all. They’ll only ever run “so well”. A proper performance difference measurement would require a dedicated benchmark, not a game. You’ll even see diminishing returns in load time measurement - the bottleneck is the storage medium itself.On Erista, not very, everything goes to shit after 2GHz or so on the CPU and probably on GPU after a similar overclock (didn't play around with it nearly as much because I'm not a complete dumbass). With "goes to shit" meaning you just get exponentially higher power draw with little to no impact anywhere else
On Mariko, who knows, I'm not aware of anyone who's done any proper testing on mariko, quite curious myself
It's called Copper/Calcio but it'll never reach the public eye
By "goes to shit" I mean you won't get anything out of it, though iirc it does hang at some point. It'll say it's running at x freq, will consume the appropriate amount of power, but won't perform any better or anything. Last time I benchmarked the CPU in Linux, 2193 and 2295 performed almost the same as 2091 and I believe 2397 immediately hanged on boot. No other limits should have been hit up to that point.
Keep in mind that you’re driving the VRM into dust at those settings - it’s very possible that the amount of power pushed through the chip isn’t actually enough, or stable enough, to maintain the clock it’s reporting. That’s not an uncommon behaviour either. For all you know the number displayed may very well be erroneous, or switching rapidly between two frequencies trying to keep up, or a number of other oddities chips do when starved. Design-wise they’re the same chip, one is just etched smaller than the other, so they should theoretically reach similar performance, with the larger chip producing more waste heat and consuming more power. You’ll always get diminishing returns as you scale, but you shouldn’t be that far off Mariko just because of different lithography - it’s not “as good”, but not to the point of creating huge differences in performance. Smaller lithography generally provides better thermal and energy consumption characteristics, which by proxy lead to better overclocking possibilities.
Mariko CPU already is at its limit with 2397. It differs per chip what you can reach. I (definitely) lost the silicon lottery. Mine gets to 2370 (checked with a homebrew application called MHz and made by KazushiMe) while KazushiMe's Switch gets to like 2390. There's not really a way to check the GPU frequency but through benchmarks, at which 1305 and up performed the same, while there was more than enough headroom (905mV while up to 1250mV was unlocked). There's only 38.4MHz for GPU to overclock above its rated frequencies and thus quite pointless. I don't have Erista.
“Unlocked” doesn’t mean actually provided. What you’re looking at is a target, you don’t know the actual value unless you externally measure. Of course that’s neither here nor there, when you get that far out of spec you’re starting to scrape for every MHz just for the sake of it, performance uplift will be minimal, unmeasurable or non-existent.
I wouldn’t recommend playing on these settings lol, or any other significant overclock. I normally set a safe “daily driver” overclock at around 15% (give or take) over max stock (unless the architecture is specifically keen on overclocking). Even Nintendo doesn’t push it in Boost Mode for an extended period of time, just for loading. Overclocking by beastly numbers is for testing, flexing and extreme modding, a bog standard unit isn’t made for this kind of thing 24/7.
It's not quite that simple. They have like 10 unique device binnings or smth stupid that vary from 1.4GHz all the way up to 2.2GHz on the CPU. Nvidia inspects them wherever and send them out to whoever depending on what range they fall into. Foster (2015 Shield TV) was binned for 2.2GHz CPU/1GHz GPU, Darcy (2017 Shield TV) was 1.9GHz CPU/1GHz GPU, Jetson TX1 was 1.7GHz CPU/9xxMHz GPU, Odin (switch) was 1.7GHz CPU/921MHz GPU. All run at a max of 1227mV at max CPU and 1150 or w/e at max GPU, I think porg (Jetson Nano) is the only exception.
In all fairness, those are different internal SKU’s. I’m talking specifically about the Erista spec, and that’s the figures I dug up for it. Obviously different binning will lead to different peak results, but they shouldn’t vary by wild margins, they’re simply downclocked accordingly so that the cores perform at their best with minimal failure and error rate. This kind of thinking lends credence to “factory OC” GPU’s which are, more often than not, complete snake oil.
The spec sheet is "meh" and not a great representation of what it's capable of, some SoCs may be completely incapable of running at those speeds. The A53 cores it mentions are also broken/disabled on all known hardware including dev boards (jetsons, SDEV, etc).
If I were a betting man, I would peg the manufacturing tolerance of the chip at 5%. Any non-working or under-performing cores below the number specified in the datasheet normally necessitate disqualification of a chip from its specific range. “Faulty” chips like these are normally used in stop gap “mid-range” SKU’s as a method of recycling parts that otherwise don’t meet the spec - it’s not uncommon to see desktop CPU’s with a certain number of cores disabled outright, down to the level of the lower SKU, with only the highest-performing cores being active. This is a known quantity in manufacturing - to give an example from the same industry, when Sony and Microsoft ordered their custom SOC’s from AMD, they knew out of the gate that some CU’s will be faulty or underperforming out of the box, so they designed them with more CU’s than they actually needed and disabled the least-performant ones at final stages of APU assembly. In the case of the Xbox One, you had 14 CU’s total with 2 of them disabled for the sole purpose of increasing yields - it gives the manufacturer wiggle room. In this regard the Switch is no different. Better-binned chips *will* clock higher, absolutely, but the tolerances are tight. There’s a number of reasons why NVidia would clock down chips besides manufacturing defects, even below the level specified in the spec. Mix in a good deal of silicon lottery and you can get wildly different results depending on the unit in question. That’s generally why I don’t like “chip X won’t do it” kind of thinking - you don’t know that until you coax it just right.
I think there might be a slight misunderstanding. All the different binning ranges are different SKUs (as per the value in FUSE_SKU_INFO), but they're all called Erista, Tegra210/T210, Tegra X1, or whatever, there's not much of a naming differentiation between them. They would produce a bunch of them and check them all, the ones that should be able to properly do 2GHz CPU or whatever at 1227mV or lower would go to be used in Shield devices, lower quality ones would go to be used in Switch devices, and the worst ones would go to be used in the Jetson Nanos, etc etc.
No, there’s no misunderstanding, I’m well-aware.
I’m in contact with Cooler discussing a different implementation of his package that would be compliant with the licenses of the individual components. If he fails to comply, I’ll be removing the package regardless of whether a rights holder reports it or not - I’m obligated to do so. A license means nothing unless we execute it, as pleasantly accommodating as you are.
Thank you; I do appreciate it.I’m in contact with Cooler discussing a different implementation of his package that would be compliant with the licenses the individual components. If he fails to do so, I’ll be removing the package regardless of whether a rights holder reports it or not - I’m obligated to do so. A license means nothing unless we execute it, as pleasantly accommodating as you are.
@
KenniesNewName:
@
KenniesNewName:
@
KenniesNewName:
