Discussion A creazy idea about AM5 platform

FAhentai

Distinguished
Jul 3, 2016
99
13
18,545
Hi guys, recently I made some research about AM5 platform, later I have realized that PCI-E 5.0 PLX is in fact there already, here is the link
https://www.broadcom.com/products/pcie-switches-bridges/expressfabric/gen5
as you can see, Broadcom has already get them.
Then, a creazy idea flash into my mind, here is what I got
https://ibb.co/kQxF1df

As we all know, AM5 cpu has total 28XPCI-E 5.0 lanes. However, the problem is, for now and maybe next 5 years, PCI-E 5.0 devices will still be rare and it may not be really neccessary. In fact, how long before we having PCI-E 4.0 10GBE NIC? Thus, PCI-E lanes are much important than PCI-E Gen. The all idea about this is convert AM5's PCI-E 5.0 to PCI-E4.0 so that we can have more PCI-E lanes, which allows AM5 as a MSDT computer has some HEDT function.

Let's back to the picture, the plan A, I would call it a little bit radical. The whole idea is get all 28XPCI-E 5.0 onto PLX, and with that PEX89088 PLX, we will get 60XPCI-E 5.0 out put. However, as we all know, AM5's B650/X670 chipset is rushly made, its m.b. chip, whose official name is Promontory 21(PROM21), is in fact th same as B550, which can only use PCI-E GEN4. Thus, if we directly link CPU to the chipset, there is a waste of total bandwideth and if there is a plx, we can use them more effectively. Besides, instead of X670's series connection of 2XPROM21, since we have enough PCI-E lanes, now we can do parallel connection on those 2XPROM21 so that the system link traffic will not go that bad comparing with X670/X670E. Fianlly, we have 52x PCI-E 5.0 to use. However, to balance the total bandwideth, we have to make sure that the rest of PCI-E's bandwideth is less or same as PCI-E 5.0 X20,which measn that we have 40X PCI-E 4.0 to use. Of course, for the future upgrading, it will be better all those PCI-E slots runs on PCI-E 5.0.

For the Plan B, it goes a little bit conservative. It is almost the same as normal X670E motherboard. The only difference is we use PEX89048 to connect its PCI-E 5.0 16X. So, in this case, we will have total 32X pci-e lane to use.

That is all my idea. I know this looks creazy, and PLX for GEN 5 is also expensive. However, this is not a motherboard for everyone, basically, its concept is close to Asus WS X570-ACE. Nevertheless, I am not really sure how much this MB could cost to build, since I have no idea about how much those PLX chips can cost. If, at the end, this MB can be sold at 800 usd, I personally believe there is still such market for it. How you think about this?
 
Last edited:
It is called Threadripper and already solves the problem at that price.
Treadripper? Come on, the cheapest one cost 2500+ AUD while R9 7950X cost only 850 aud. Not montion if the MB can be sold at 800 usd price, there is hardly any price difference between the M,B.
However, till now I still not know if it is possible.
 
  • Like
Reactions: artk2219
Treadripper? Come on, the cheapest one cost 2500+ AUD while R9 7950X cost only 850 aud. Not montion if the MB can be sold at 800 usd price, there is hardly any price difference between the M,B.
However, till now I still not know if it is possible.

But considering the very rare person who actually needs it will buy the Threadripper, what would be the purpose of selling it for a fraction of the cost, from their point of view?
 
  • Like
Reactions: artk2219 and Eximo
You don't spend development time and money where you aren't likely to have a customer base. The person that needs 50 PCIe lanes isn't someone doing something cheaply.

At some point you are crossing over into the territory of server hardware anyway, where I can already get massive PCIe bandwidth and the right hardware to facilitate a mass of PCIe connected devices.
 
  • Like
Reactions: artk2219
You don't spend development time and money where you aren't likely to have a customer base. The person that needs 50 PCIe lanes isn't someone doing something cheaply.

At some point you are crossing over into the territory of server hardware anyway, where I can already get massive PCIe bandwidth and the right hardware to facilitate a mass of PCIe connected devices.

Yup.

The problem is that a lot of people look at a product not existing as proof it hasn't been thought of or nobody knows how to do make it, when it's usually that the product makes little sense in terms of devoting resources to develop it. They certainly could make a Honda Civic that can double as a ride-on lawn mower, but nobody actually wants to buy that, at least enough people to make the development worthwhile. I like beer, and I like playing games, but I don't need a $500 beer stein that can play Fire Emblem.
 
But considering the very rare person who actually needs it will buy the Threadripper, what would be the purpose of selling it for a fraction of the cost, from their point of view?
I guess I have already said, that this MB's concept some how close to Asus WS X570-ACE, here is the link for its specs, it is clear that this MB is close to "cross over". Meanwhile, if you ever check Asus, Asrockrack's website, you will also find many of those workstation/server motherboards that are using MSDT cpu. Thus, it is clear that there is such niche market.
Now the only problem is, on theory, this MB is possible, yet, I have no idea how much the PCI-E 5.0 PLX is going to cost, nor the price for the rest. That is why I am saying that if this desgin can be done and the fianl price of this M.B is about 800 usd, even we push the price to 1000 usd, there should be a market. The main purpose for me to post this thread is to ask if any one can estimate the cost for this desgin. After all, price matters
 
I'm not sure you are helping your idea out by listing existing products that fill the demand you are talking about. Specifically AM4 socket workstation boards there was a gap for Ryzen chips to fill the need of entry level Threadripper when a new Threadripper generation wasn't available outside of OEMs.

ASUS and ASRock Rack caters to high end enthusiasts and small businesses that don't need a straight up enterprise grade solution. Though they have that too. Priced accordingly and fills that niche.

Fairly easy to look up the pricing for bulk PCIe switches, but engineering a board to use them and the quality of board needed to do that with PCIe 5.0 speeds is somewhat apparent in the Threadripper boards. Not sure I have seen any Sapphire Rapids boards for sale anywhere, but presumably just as pricey.

I don't think you are going to find anyone with enough inside information to give you the financial outlook on a potential product idea.
 
I don't think you are going to find anyone with enough inside information to give you the financial outlook on a potential product idea.
:cryingcat::cryingcat::cryingcat::cryingcat::cryingcat::cryingcat:
In fact, I can not even found any exisiting PCI-E 5.0 switch. 🤣Tho there is already chip for this.
 
I made an account just to tell you its not a crazy idea. Those PLX switches are goldmines in the right applications, as they effectively load-balance the throughput to maximise effective PCIe to the functionality of a much larger system. They are becoming more popular in workstation motherboards i.e. there by default. https://www.anandtech.com/show/15821/microchips-new-pcie-40-pcie-switches-100-lanes-174-gbps

Rarely, however does a new-generation PCIe switch (gen5, maybe even gen4..) make sense, because they are so expensive, or inconvenient. Thats changing in the gen3 NVMe market right now, where people have x16 gen3 and they want all 4x4 pcie3 NVMe storage to use those lanes with a cheap $20 adapter, but shockingly (to them) their motherboard dorsnt support bifurcation, as this is essentially something a PCIe switch would do....

Now with PCIe NVMe gen4 is being out there, the demand for PCIe gen4 switches is there and growing. The limiting factor is cheap solutions, so adding those solutions (however you can) to motherboards, essentially throwing out the old model of pcie and moving over fully to "fabric" design style, would make AM4 very competitive in workstation use for another 10 years.
 
you missing the op want it on the cheap consumer level pc motherboard
No mate, I never said that. If this kind of MB can be made with Worksation MB price there is still such market.
The price between CPU is still high.
I made an account just to tell you its not a crazy idea. Those PLX switches are goldmines in the right applications, as they effectively load-balance the throughput to maximise effective PCIe to the functionality of a much larger system. They are becoming more popular in workstation motherboards i.e. there by default. https://www.anandtech.com/show/15821/microchips-new-pcie-40-pcie-switches-100-lanes-174-gbps

Rarely, however does a new-generation PCIe switch (gen5, maybe even gen4..) make sense, because they are so expensive, or inconvenient. Thats changing in the gen3 NVMe market right now, where people have x16 gen3 and they want all 4x4 pcie3 NVMe storage to use those lanes with a cheap $20 adapter, but shockingly (to them) their motherboard dorsnt support bifurcation, as this is essentially something a PCIe switch would do....

Now with PCIe NVMe gen4 is being out there, the demand for PCIe gen4 switches is there and growing. The limiting factor is cheap solutions, so adding those solutions (however you can) to motherboards, essentially throwing out the old model of pcie and moving over fully to "fabric" design style, would make AM4 very competitive in workstation use for another 10 years.
Look mate, in fact, for this MB, you don't even need to that creazy. You don't even need full PCI-E 5.0 PLX to do the job. Just 28X up link PCI-E 5.0 to 56X down link PCI-E 4.0 PLX would do the job. However, I can not found this kind of product, the only thing I can found is PEX890XX family
 
What would the average person do with 60 PCIe lanes? Multi video card setups are pretty much dead for gaming and there's no reason to have a dozen M.2 drives in the system unless you bought a bunch of 250GB ones over time (RAID is also impractical/unnecessary for a variety of reasons)