I’ve seen pretty different benchmarks, and was surprised when this Phoronix article came out! May be a matter of Micheal Larabel using the 6.7.0 kernel for some reason? That’s quite an old kernel for bleeding edge hardware.
I suggest this video for far more in-depth benchmarks! Written english subs are available.
Yes, if the scheduler was not well adapted to Linux, the performance will be poor. This CPU has 3 types of cores: Performance, Efficiency and (new) Efficiency Low Power. That’s a lot for a scheduler to handle and it has to know a lot about what software the OS is running and which task to hand off to which core type. It can very easily get this wrong.
The simpler P+E cores in the 12th and 13th gen took quite some time to get right in Linux and early benchmarks were bad.
Notebookcheck has reviewed a number of 7840U and 155H laptops at this point. Performance is roughly on par between them so I think you can basically choose whichever color you like.
On the higher-end models (Core Ultra 7/Ryzen 7) the Ryzen might be the better deal. But I think between the Ryzen 7640U and the Core Ultra 5 125H the latter is the better choice as you only lose 2 P cores with the same amount of E cores and 1 CU on the iGPU, so multicore and iGPU performance should turn out to be superior to the 7640U.
With the new display, webcam and 61 Wh battery the price difference is 70,-€, which I think is really worth it.
Hence why I preordered one. I didn’t want to go with the more expensive CPU options as I feel like the next generation Intel/AMD CPUs will be such a big step that I’ll be quickly upgrading anyway, so I figured I’d save myself the 400,-€.
Despite NotebookCheck’s collection of reviews, I think that they are misleading and insufficient due to the broad differences between how OEM SKUs address heat management, CPU power delivery, and more.
I think that an in-depth review of the upcoming Framework Intel Core Ultra 1 series laptops could be a golden opportunity to finally be able to elucidate a truly in-depth, comprehensive and somewhat more scientific comparison between the Core Ultra 1 series (eg. 155H) and the AMD 7040U series (eg. 7480U).
I think that there’s been a lot of uncertainty online in the comparison between these two CPU series, especially because of the rather original approach that Intel has taken (chiplet tiles vs monolithic design) and that this Framework release will provide an opportunity for an unprecedentedly “clean” comparison given that everything aside from the chips would be the same.
In order to make the most of the in-depth comparison, the AMD Framework laptop also should use the same new 2.8k display as the new Intel one (the reviewer would need to upgrade it), in order to eliminate that as a differentiator in terms of battery life.
The article is 6 month old. Phoenix retested both AMD framework laptops, and a 155h laptop together with a 165u laptop last week:
EDIT: The article is new but data are still from Ubuntu 22.04 with kernel 6.5. Did not realized that.
I hope they’ll test in depth the Intel Ultra framework when it comes out.
Thanks everyone for the nuance 
You saying AMD Ryzen 7 784U is better than Intel Core Ultra 7 165H?
Yes it’s better, and will be even better: https://www.phoronix.com/news/AMD-VCN-Dynamic-Power-Gating
Take the AMD one
can anyone talk about which of these two CPUs offer (better) support for eGPU via USB ?
Short version: Intel for maximum compatibility, AMD for maximum performance.
The Intel Core Ultra Series 1 model will probably receive Intel Thunderbolt certification. Intel Thunderbolt certification is a certification saying that the product has passed Intel’s rigorous testing process to ensure excellent compatibility and stability.
However it is important to note that Framework doesn’t (yet) list Thunderbolt 4 certification on their specs (they do for models with 13th gen Intel CPUs and previously for 12th gen Intel CPUs), indicating that it hasn’t yet received that certification/passed that testing process but I expect it may in the future (although Framework’s 11th gen Intel models failed the testing).
The AMD Ryzen 7040 series has more USB4 PCIe tunneling bandwidth available, enabling better eGPU performance. Theoretically the bandwidth is 25% higher (ignoring overhead), however testing on eGPU.io has indicated closer to 30% extra bandwidth. However only some eGPUs (models equipped with the ASMedia ASM2464PD/ASM2464PDX chip so far such as the ADT-Link UT3G) can take advantage of that extra speed.
2 posts were merged into an existing topic: AMD announces new Ryzen AI APUs (Strix Point) in Computex 2024
I’m upgrading my mainboard soon to hopefully lessen the CPU starvation to my eGPU. Should I pick up a UT3G as well or will my R43SG (~3y older) be fine? I know it’ll work just as well on Intel/Thunderbolt, but this has me concerned whether it will perform on Linux over AMD/USB4. This is my main consideration now on whether to get the Core Ultra or the Ryzen.
I’ll be using Linux (which is a slight performance hit), dual channel memory, and a 3060 Ti either way.
What is that based on? The read speeds are pretty much identical between modern Intel and AMD CPU-integrated USB4 controllers.
It seems that with eGPUs we are seeing reduced H2D bandwidth on Intel systems. But given that D2H has the full expected performance I would think that this is probably not an actual bandwidth issue, but caused by some other difference. Maybe its latency and it still affects GPU performance. But I have not seen good benchmarks for that.
And Meteor Lake is newer. The one benchmark I found seems it has more H2D bandwidth than before, though still not as much as AMD in the CUDA benchmark. But without apples-to-apples game benchmarks would not apply this directly to game benchmarks. Too many bugs with ASM2464 solutions etc.
Pretty sure even the new the intel platforms don’t get the same bandwidth increase from the asmedia chipset as the amd ones do, on amd the difference is quite big and may be worth it. Though since you already got something working may as well wait for better ones to come out at this point, ut3g feels more like a proof of concept than an actual product.
What bugs are you talking about? So far the asm has behaved pretty much the same as may gen 1 and gen 2 tb3 devices, appart from not usb4 hotplugging on linux with one speciffic ssd.
When I got my Satechi ASM2464 enclosure, it fell back to USB2 on my Titan Ridge TB3 host.
Newer (inofficial) firmware solved this.
It still makes x2 PCIe connections if behind a TB4 hub or TB3 dock. On a TB4 Maple Ridge host as well as my 12th gen Framework.
And with the UT3G as the only eGPU board with it I am aware of so far, there seem to have been many revisions trying to solve overheating issues etc. But here I am relying on others reports. So I may not have an accurate picture of how well it works for people on average.
Also mess like PR that it is now supposedly TB4-certified. But no way to find out for which firmware this would be true for and seemingly no manufacturer claiming the product to actually be certified…
It certainly doesn’t have the polish that comes with certified tb stuff, I’ll give you that.
I also thought I had a kernel bug there because it reported running at 1x but turns out I just performance tested wrong and single threaded dd was not able to fully load the link and when I retested I got numbers indicating full bandwidth anyway. Link reporting on usb4 is a bit weird rn.
That thing certainly feels more like a proof of concept than an actual product and I do hope a nicer one with 65 or 100W pd comes out at some point. Also don’t think the current asmedia chipsets support daisy chaining which was pretty neat on the TH3P4 but I am fine trading that for the extra usable bandwith.
Vendors lying about certification sucks but I doubt that’ll get fixed soon. Not sure what you mean by PR.
Parts of the connection are virtual. That is were I have seen “x1” links from Intel controllers and where it seems not to matter at all. But what is reported on the actually physical connection between TB/USB4 controller and PCIe equipment such as NVMe or GPU has been accurate in all cases. And bandwidth tests match, while TB3 equipment or TB4 equipment makes x4 connections on the same ports. So that is some kind of issue involving the ASM2464.
Sadly I do not know enough about debugging PCIe under Linux to find why it downgrades that, because it still reports the ports capabilities accurately.
Yeah. Not sure if no other manufacturer has brought a competitor because of the lack of PD support from the controller or because they have problems getting it to work for off-label use. (ASM2464PDX is supposed to allow external PD controllers to allow PD output. But we have not seen any in the wild yet).
It was specifically designed for NVMe enclosures. It is completely geared towards that. Including fully integrated handling of PD, which prevents any kind of PD power output etc. Adding a downstream USB4 port would just force them to add more capabilities like PCIe switch etc. that make it more complicated and make very little sense for the NVMe enclosure market.
Press Release. I do not doubt that they had there reference board certified. But if no manufacturer using the chip gets their product certified or provides a firmware update to achieve “TB4 certification” like Framework did for 12th gen, its not really useful. Because with the original firmware on my enclosure not even showing up in TB control center (my Anker 566 USB4 endpoint still does not), its probably not certifiable on any of the firmwares I have tried (basically all that I found starting with the October 23’ version.
So it either seems not easy to make that controller TB4-certifieable or there are other complications with it…
Interesting, while them having a reference board certified may not help you immediately it does prove that it is possible and their implementation is/can be complete and it’ll massively increase the probability of manufacturers using that chip attempting to get certified.
I’d say it’s probably not as hard as it is expensive and tedious.
If big laptop manufactures usually can’t be bothered to even use the tb4 ports that come for pretty much free (you do of course need a fancier re-driver and need to do more careful signal routing but it is certainly a lot cheaper than needing a separate controller for it), let alone certify it I doubt most nvme enclosure and especially egpu manufacturers will even attempt it, especially for very low volume and low margin stuff like ssd enclosures or diy egpus.