Reply to thread

Message: [QUOTE="Bombastus, post: 91511, member: 2178"] Minisforum lists one PCIe 4.0 x4 SSD slot and one PCIe 4.0 x1 slot in the MS-S1 MAX-P495. Both accept NVMe drives, but the second connection has only a quarter of the first slot’s bandwidth. Combining them in RAID 0 spreads file data across both drives. For the [B][URL='https://goldmidi.com/community/threads/minisforum-has-introduced-the-ms-s1-max-p495-workstation.76847/']MS-S1 MAX-P495 local AI workstation[/URL][/B], that can make reading a large model slower than using one capable SSD. Buying two identical drives won’t fix the unequal connections. [HEADING=2]The slower slot limits striped transfers[/HEADING] PCIe 4.0 provides roughly 1.97GB/s per lane, or 7.88GB/s across four lanes, before protocol overhead. These are connection limits rather than promised SSD speeds, and actual transfers will fall below them. With two equal-sized drives and conventional equal striping, each SSD must supply half the data during a long sequential read. The drive in the x1 slot cannot deliver its half any faster than its connection allows. The x4 drive cannot supply its partner’s missing data. Double the slower link’s 1.97GB/s ceiling and the striped pair reaches roughly 3.94GB/s before further losses. Adding both links together to claim nearly 10GB/s would ignore how the data is distributed. For an illustrative 80GB file, those limits imply a minimum transfer time of about 20.3 seconds across the pair. Reading the same file through x4 alone has a theoretical minimum near 10.2 seconds. Neither figure is a measured P495 result. A sufficiently fast SSD in the x4 slot can therefore outperform the array on large sequential reads. Striping makes that SSD carry half the workload without letting it use its spare bandwidth to finish the other half. The result changes when the SSD itself is slow enough to become the limit. Two modest drives could still benefit from striping even across unequal connections. A blanket claim that RAID 0 always loses would be wrong. [HEADING=2]Mirroring gives up capacity and write bandwidth[/HEADING] Minisforum lists support for an 8TB SSD in each slot, with RAID 0 and RAID 1 advertised alongside 16TB of storage. You need to separate installed capacity from space available for your files. Two 8TB drives provide roughly 16TB before formatting when combined in RAID 0. A RAID 1 mirror holds the same data on both drives, leaving roughly 8TB before formatting. The second drive stores another copy rather than another collection of models. Mirroring also makes every write pass through the slower connection because both SSDs receive the complete data. For sustained writes beyond any buffering, the x1 link imposes a ceiling near 1.97GB/s before overhead and drive limitations. Reads work differently because either healthy copy can satisfy a request. The RAID implementation decides how to distribute those reads. You cannot apply the mirror’s write ceiling to every read workload, or promise that reads will always use the faster slot. With RAID 0, losing either SSD can make the entire volume unavailable because neither drive holds a complete copy. Keeping your only copy of an edited dataset there means accepting that failure exposure, even when the original model is downloadable again. A healthy mirror can preserve access after one SSD fails, but it does not preserve yesterday’s files when you overwrite them. Both copies receive the change. Recovering an earlier version requires a separate backup or another version-retention mechanism. [HEADING=2]Separate drives keep active models on the faster slot[/HEADING] Leaving the drives as independent volumes lets you reserve the x4 connection for models you open frequently and files you repeatedly rewrite. Older model downloads and less frequently used datasets can occupy the x1 drive without forcing active files into an evenly striped layout. Suppose you use 1TB of model files regularly and keep another 5TB for occasional work. A 2TB drive in x4 and an 8TB drive in x1 could hold those collections separately. Independent volumes do not require matching SSD capacities. A failed archive drive would not, by itself, destroy files stored wholly on the separate working drive. Whole-system failures can still affect both. You still pay for the slower connection when you read those archived files. Copying an 80GB model from x1 storage to x4 storage has a link-only minimum of about 40.6 seconds. The destination’s extra bandwidth cannot make the source deliver bytes faster. Once you have copied a model to the faster drive, later disk reads can come from x4 instead of x1. Retaining the archive copy consumes space twice, reducing how much distinct data the two drives can hold. Choosing [B]separate drives for active and archived models[/B] means managing two distinct pools of free space. Unused capacity on the archive drive does not automatically extend a full working volume. A download aimed at the full drive still needs room made there or a destination on the other SSD. [/QUOTE]

Name