Gemini’s take on this…and I gave it a skeptical perspective:
I think it’s an intentional artificial ‘limitation’ placed by Intel.
Not sure why the previous post was flagged…maybe I’ll rephrase it.
Maybe it’s an artificial limitation placed by intel to differentiate systems between soldered-on memory vs removable modules platforms. e.g. lowered end of soldered-on memory systems (say, with 16GB) will eventually have a shorter ‘useful’ life…which would push portion of them to the landfill sooner…pushing for future sales. Intel has been continuously spec out lower speed for removable memory.
Hell if you let them, they would rather have you on a subscription with processors (cough HP subscription laptops).
It’s why we’re here rooting for Framework, with [user-]replaceable memory, even with the speed ‘limitation’.
First off, people have to tamper their expectations a bit. LPCAMM2 can never match soldered memory directly next to the chip since the traces are physically longer. So one shouldn’t even be comparing GPU/CPU performance under LPCAMM2 vs soldered memory if they want to accuse the manufacturer of deliberately “knee-capping” their product. As such the 22%/11% numbers that are thrown around in this thread are deeply misleading.
Secondly, even if the LPCAMM2 module can run memory at a certain speed it doesn’t mean that every processor with LPCAMM2 support will have a controller that can support those speeds. This reminds me a bit of the early Ryzen CPUs who weren’t able to take advantage of faster memory. This is most likely the culprit for why the LPCAMM2 run only at ~1k MT/s less than that advertised maximum of an LPCAMM2 module.
In any case, I strongly suspect that this is a limitation from Intel’s CPU controller rather than the far-fetched “Framework is deliberately knee-capping their product and covering it up”.
1 Like
We’re updating the copy to “LPCAMM2 LPDDR5X 8533 memory modules are compatible with Framework Laptop 13 Pro mainboards with Intel® Core™ Ultra Series 3 processors, with support for a single module up to 96GB. These modules are rated at 8533MT/s, and will run at up to 7467MT/s in our Intel® Core™ Ultra Series 3 Mainboard.”
Our intent was always to communicate that the modules will run at 7467MT/s in this Mainboard, and we want to make sure that is dead obvious in the text. This is in general a limitation of LPCAMM2 with Intel Core Ultra Series 3, without using an esoteric PCB process that would drive up the price of the product substantially.
22 Likes
That’s totally reasonable but leaves the question: why did you go for 8533MT/s Modules then? wouldn’t it be cheaper and less confusing to go for 7467MT/s or 7500MT/s?
My thought was maybe it’s firmware that will be updated later on to enable support for faster RAM.
I think it’s generally also unfair to say something is/was misleading here, the ram is not labeled in the configurator and everywhere where it is labeled there is a note telling it only runs at 7467MT/s.
Adding that to the configurator might actually make it more confusing because technically you’d have to write both then, 8533MT/s that runs at 7467MT/s.
1 Like
My guess is availability. They probably have to go with whatever they can reliably source at necessary quantities. This may change, but with the way RAM pricing and supply is right now, they probably have to take whatever they can get. I see Lenovo doing the same thing (advertising machines as coming with 8533 modules, but running at 7467MT/s).
1 Like
My guess is Framework is more transparent than Lenovo. However since Lenovo has a much larger user base and Lenovo uses 8533MT/s to advertise, it makes the 8533MT/s more popular than 7500MT/s, which leads to Framework also stacking this RAM speed in bulk.
Guys, Intel’s own datasheet show with LPCAMM2 it supports max 7467 speeds.
It’s Intel limitation not Framework.
Here: https://www.intel.com/content/www/us/en/content-details/872188/intel-core-ultra-series-3-processor-datasheet-volume-1-of-2.html
Section 3, Page 155
Under LPCAMM2
Processor H484/404 H484/H12Xe 404 H484/H12Xe 404
Maximum Frequency
[MT/s]
Type 3:
1R/2R - 7467
Under LPDDR5/x
It goes up to 9600.
By the way, this is why I think they could have gone with SO-DIMMs as well. Because the same datasheet shows it goes up to DDR5-7200.
2 Likes
Thank you for showing that! Good to know it’s Intel’s limitation rather than it being the Framework.
1 Like
Intel only classifies their iGPU as Arc B390 with memory speeds of 7467+ Mt/s. It shows as Intel Graphics otherwise.
There are probably some data integrity issues tied with that, or whatever.
That would have defeated the “LP” point of the equation. Also actual jdec 7200 sodimms are probably even less available than lpcamm2 is currently.
2 Likes
Yes but it would have allowed current users to get it working with existing modules and upgrade to higher speeds later when the AI hype crashes and burns.
By the way power differences are not significant under actual usage scenarios. It’s 20 percent savings over what is already low. My desktop RAM is rarely seen at 1.5W, and usually stays at 1W. Laptop RAM is even lower.
Check RAM power under HWInfo.
It’s 70 percent or so savings in sleep mode but again RAM is a subset of the total.
As for the naming it’s just branding. Actual performance wise it’ll be practically zero.
You know what, I get Framework’s decision. They need it because initial impression in early reviews significantly determine sales.
That may be for the ram alone, a big deal about lpddr is that it can do some stuff like refreshes itself and uses much more efficient signaling which massively reduces the power consumption of the memory controller especially under light and medium loads. Under high loads it probably matters somewhat less.
A significant portion of the savings comes from lpcamm(and regular camm) allowing much shorter memory trace lengths which allow lower power for the same speed or higher speeds. Then again you don’t really see many manufacturers soldering ddr despite lpddr being more expensive so there must be some value to the stuff.
I have a laptop that still supports S3 sleep, when switching from S0ix to S3, its sleep time at least doubled, because in S3 the power to the CPU is cut.
A laptop that support S3 with LPCAMM2 will probably be the king of suspend
S3 basically turns everything off while S0iX has things in a lowest power state enough to wake the thing up nearly instantly.
100mW system power was the goal. That’s 700 hours with a 70WHr battery if done well, or almost a month.
Most issue with S0iX was when it woke up, not only it stayed awake but actually active, stuck in some higher power state. I really don’t need sleep updates, and it’s a potential privacy nightmare.