No Intel client platform (including Xeons* based on client architecture) to date supports 4 way bifurcation natively (has been 8/8 or 8/4/4).
Sounds plausible, I used them on dual socket Xeons and still have one on a Xeon D 1541, which offers 4/4/4/4.
I don't think I saw that restriction on my Zen 3 and 4 systems, but those aren't as tight with PCIe lanes as Intel used to be.
They list the same switch for all of their PCIe 5.0 NVMe cards so I'm just guessing whatever volume rate they get is better than buying the 24 or 32 lane versions.
Highpoint-Tech controllers were never cheap for what little hardware the often had. But given their small scale and the fact that they offered in niches nobody else served (or much more expensively), I've used their controllers for many years.
Yet at €999 sales price, the Broadcomm chip must be the biggest chunk of their price and switch prices tend to scale at least linearly with the number of ports. And I might actually be tempted to get a variant based on the 24 port chip with x8 or x4 host side for €400 at PCIe v5, especially if it could switch from PCIe v3/v4 to PCIe v5 during transfer.
So I am really puzzled by their choice of the 48-port ASIC on this SKU, programmability and tools would be all the same across the 24-144 port PEX 89000 product range and 32 ports would certainly be enough for this use case: wasting 16 PCIe v5 lanes is a brutal waste and hard to explain, especially since Broadcomm simply isn't known to go easy on prices nor does HighPoint-Tech have enough scale to obtain significant discounts.
I don't know how many die variants Broadcomm makes, perhaps it all about blowing fuses below 48 ports while yields are great, still I don't see them relaxing on price: it's just not who Broadcomm is and why we love them so much.
I can appreciate the management advantage these switches offer, you basically gain a full level of "physical" virtualization, but again Broadcomm doesn't give that sort of thing away for free: you have to buy the hardware and then still license the software to operate that layer, likely with a yearly renewal. It addresses that shrinking on-premise niche between clouds and desktops.
Broadcomm ate so many startups that only SAN vendors seemed able to afford PCIe switch chips for the last decade. Hyperscalers use their own fabrics or co-develop with Broadcomm, but some new startups seem to appear, although also typically at the high-end and with CXL in mind.
Wherever this is going, cheap home-lab or hobbyist use doesn't seem to be part of that... even if all those lesser capacity, still good-enough M.2 drives keep piling up in my drawers, much more difficult to re-purpose than those SATA-SSDs.
Facebook
Twitter
Instagram