• Re: Killer price on a mini PC

    From Paul@3:633/10 to All on Wednesday, September 16, 2026 07:24:07
    On Tue, 9/15/2026 9:48 PM, Joel W. Crump wrote:
    https://www.amazon.com/dp/B0GXTCSJY6/


    What makes this one interesting, at that price, is it has the eight-thread CPU with 16 GB RAM 512 GB SSD.ÿ Best deal I've seen, to date.


    I think it's four threads. Compared to three other CPU
    for fun. It's similar to an N150, and how much of the
    multithread, might depend on the cooling provided.

    https://www.techpowerup.com/cpu-specs/ryzen-5-3501u.c4442

    3501u
    Zen+ (Picasso) 12nm 2018 GlobalFoundaries
    4C 4T 2.1/3.7GHz
    DDR4-2400, dual channel
    TDP 15W
    Vega 8 iGPU
    N150 N95 6600H 5700G=Mine 5950X
    Single thread CPUMark 1805 1893 1890 3142 3281 3476
    Multithread CPUMark 6195 5354 5285 18050 24218 45252

    1805*4=7220 with perfect scaling, 6195 is a bit off (3.4 threads).

    My 4930K has a similar problem. 6 cores, but depending on
    the bench, only gives 4 to 5 cores of performance, and
    that's because of the ring bus. The Zens might be a hypercube
    or something. Intel put a bus-mesh in some of theirs,
    like say five buses running horizontal, six buses running
    vertical, which performs better than three coupled-on-the-ends
    of their ring buses (a technique from previous generations, nasty).

    Summary - a weak sauce, bound to annoy, and make you buy another.
    - silicon too old, for safety reasons, like "buying a refurb" computer
    - chances are, fine on Linux (browsers behave better on Linux)
    - nice aluminum box

    - whereas, a good single-thread, helps on Windows a bit. I would
    move to the next tier up for Windows.

    - measure twice, cut once, try to select the right mini-PC before
    splashing the cash. I've made similar mistakes, back in the Core2 era,
    always buying the wrong one, having regrets later. Lots of mistakes.

    I don't expect the refurb-market can do much better. At one time, the
    refurb market was a valid choice, especially when quad cores started
    showing up there.

    Paul

    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)
  • From Paul@3:633/10 to All on Wednesday, September 16, 2026 10:11:35
    On Wed, 9/16/2026 7:36 AM, Andy Burns wrote:
    Paul wrote:

    On Tue, 9/15/2026 9:48 PM, Joel W. Crump wrote:
    https://www.amazon.com/dp/B0GXTCSJY6/


    What makes this one interesting, at that price, is it has the eight-thread CPU with 16 GB RAM 512 GB SSD.ÿ Best deal I've seen, to date.


    I think it's four threads.
    AMD do say 4C/8T ...

    <https://www.amd.com/en/products/processors/laptop/ryzen/3000-series/amd-ryzen-5-3501u.html>

    I'm confused, as Zen+ is from 2018, and this is a re-released part that still claims to be Zen+ architecture.
    This article, claims a re-released part (to consume DDR4), from *June 2026*. So it's not new-old stock,
    it is new (dreadfully bad) stock.

    https://wccftech.com/amd-quietly-revives-7-year-old-zen-architecture-with-new-ryzen-3-3100u-and-ryzen-5-3501u-laptop-chips/

    That part number is listed here as 4C 4T. That's where I usually check these, as the AMD
    site isn't all that easy to use.

    https://www.techpowerup.com/cpu-specs/ryzen-5-3501u.c4442

    The part number is not listed here, but the family trend is 4C 8T in the table (3500u).

    https://en.wikipedia.org/wiki/List_of_AMD_Ryzen_processors#Picasso_%28Zen+/GCN5_based%29

    # Similar, seen here, a 3500u listed as 4C 8T in CPUZ. Still can't find a Task Manager picture.

    https://www.reddit.com/r/pcmasterrace/comments/1sx19bt/ryzen_5_mobile_3500u_on_new_pc_incredibly_slow/

    # Listed as 4C 8T there. Bench between 3500u and 3501u about the same, may depend on cooling
    # or power dissipation on the newer one.

    # A lot of Reddit threads start out hopeful, but no illumination on why it is slow.

    https://www.reddit.com/r/linuxhardware/comments/1ay90c4/debian_12_with_amd_ryzen_5_3500u_8gb_ssd_256gb/

    For an 8 thread processor, you expect performance on 4 cores to be
    better than 3.4 cores worth. And the CPUMark isn't showing a good ratio.

    It's the same trend on the 5950X, which is Zen3. And Jim Kellers
    comment at one time was "you're testing it wrong". The ratio can
    be partially accounted for by the 5GHz single core clock rate,
    versus about 4.4GHz at 200W on all cores (to go from 16 to 13 worth),
    but that doesn't account for the hyperthreading being worthless.

    5950X 16C 32T, ratio is 13.02x multi to single
    Single thread CPUMark 3476
    Multithread CPUMark 45252

    Hyperthreading works, when locality of reference is poor, a CPU core is
    blocked waiting for a memory access to complete, and the alternate thread starts to run as it's not blocked on memory access (yet). If memory is
    not slowing down one core, then there is no reason for the other thread to run. The scheduler may cause the other thread to run, eventually.

    Historically, hyperthreading started out pretty poor. The right thrashing pattern, and it was better to turn off hyperthreading. People were
    turning it off for gaming, just as some sort of move to get better
    scheduling going.

    The comments so far then, make it out to be slow hardware, but with no particular reason discovered for whether that was a defect with the
    provided NVMe, or something else. The device is a PCIe Rev3 generation
    device. Depending on where the NVMe is connected on it, it can't go
    faster than PCIe Rev3 x4. There is also some deal about the NVMe being
    the SATA type. That still shouldn't stop it from working well, if that
    was the case.

    I also suspect, based on a review comment, that these boxes ship
    with a single DIMM in one slot, or, with the RAM being soldered
    to the motherboard, on one channel. The iGPU will use some of the
    bandwidth, and the memory is DDR4-2400. Normal for modern DDR4 might
    be 3200 or 3600.

    Paul

    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)