So I’m in Batch 5 for the Framework 13 Pro with the Intel 358H and X7 (B390) iGPU. My system was downgraded from 64GB to 32GB, but with the latest update from Framework, I have an option to upgrade back to 64GB for $1,200. IIRC, the original pre-order price was $850.
Normally, I think 32GB is plenty for a laptop. My main concern is that the X7 uses shared system RAM rather than dedicated VRAM, so graphics-intensive workloads like gaming and solid modeling can eat into that 32GB. While testing, I noticed games like Final Fantasy VII Rebirth and Beast of Reincarnation were using nearly 11GB of system memory, which doesn’t leave a ton of headroom once Windows and everything else are factored in.
Windows can obviously use the pagefile when needed, and modern NVMe drives make that much less painful, but they’re still nowhere near as fast as DDR memory.
Even with that concern, I’m leaning toward sticking with 32GB. For my workload, I don’t think an extra 32GB is worth $1,200. If I ever need more memory, I can always remote into my desktop that has 64GB of memory. Although, a lot of places I’ve stayed have absolutely terrible WiFi, so I’m trying to find a good balance with local performance too.
I also managed to get a Gigabyte AORUS 5060 Ti AI Box eGPU before the price increase, for $699. That should alleviate most of my concerns about the shared memory while giving me significantly better gaming and solid modeling performance. The downside is obviously having a much heavier and bulkier setup if I travel frequently. Although I could probably put everything else (the power brick and the cables) in checked luggage and keep the eGPU with me.
What do you guys think? Also, for those of you who use a similar setup, what are you using to remote into your desktop?
I’m mainly using:
Sunshine/Moonlight for game streaming.
AnyDesk as a backup for general remote access.
Bitvise SSH Server/Client so I can use Windows’ native RDP.
32GB is enough for most purposes - I agree swapfile is fast these days. I got 64GB DDR5 because I use VMs heavily, and the way I use them is quite unusual. Most folks don’t need this.
Curveball - if you swap from Intel to AMD, then you’ll only lose a bit of CPU performance (Intel Core Ultra X7 358H to AMD Ryzen AI 7 350) but you will get 64GB for roughly the same price as 32GB (they are both £800 in the UK). The downsides are you lose the increased memory speed and battery performance, so you’d have to decide if they matter to you.
I think Windows, with fairly conservative optimization can fit in an 8GB envelope, leaving plenty of room for applications. Also, this test system, what GPU was in use and how much VRAM was used at what settings? The Core Ultra are powerful for iGPUs but they aren’t that powerful.
They’ll compete with entry level dGPUs and maybe lower midrange and that’s it. Those GPUs ship with 8-12GB of VRAM. So let’s say 8GB of RAM for Windows, 12GB as VRAM and another 12 used as RAM for applications. That gets you right at 32GB.
I don’t think you’ll need all that tho. That’s just worst case scenario and when using my FW13 w/32GB of RAM I never came close to using it all. Albeit that was with a much weaker iGPU that prevented me from playing super demanding games…
Still, I think if you just drop to the 3:2 equivalent of 1080P with medium/high settings you’ll have the best balance of appearance, memory usage and framerate.
I think a not-so-unusual workload nowadays is to have claude/codex in a VM with unrestricted access to everything and have it working in the background. Meanwhile one might run some foreground programs like a bunch of browser tabs and games.
With games RAM usage can easily reach 20GB, then if your dev work inside VM needs 16GB you will start to swap. I can imagine 64GB being useful for such use cases…
I wonder if I have achieved the same in a different way. I’ve jailed Claude in a container that sits in an application sidecar, in a local env only. It has access to backend tests against a RAM-disk database, a Playwright container, and a separate app/database for exclusive robot use. I’ve been using this configuration for about five weeks and am very pleased with it - it can do much of what it likes and is locked down against risky stuff. I don’t see the efficiency value of MCP, so I’ve not explored giving Claude that capability yet.
I thought about containers too, but then I decided to install the dev environment on an expansion card so I can either boot it as VM, boot it with real hardware or boot it on another machine as VM/physical, just for the sake of it
And a fair amount of GPU performance, the Arc b390 seems to outperform the 890m watt for watt, and I imagine there is still some driver enhancements to be had on the Intel side.
