Discussion in 'Frontpage news' started by Hilbert Hagedoorn, Jun 16, 2019.
All I wanted was B550, that only comes in Q1 2020...
I'll have to get B450 probably.
Funny, no one even thought about dual socket...
Quite a feaseable approach.
Desktop-grade CPUs haven't been multi-socket capable since the early 2000s. To my understanding, sockets AM4 and 1151 are physically incapable of a dual-socket config, and it doesn't make economic sense why you'd want to. But, both AMD and Intel have multi-socket offerings (for servers).
Where can i find this info?
Rumours says that AMD almost launched an high end chipset last gen, but cancelled, maybe now it makes sense for them. Who knows? They omitted threadripper on some new papers, maybe their strategy is to from 32 to 64 with thread ripper and the in between, 16 to 32 with a dual socket AM4.
Ever noticed there's no longer a "northbridge" in modern PC motherboards? That's because they're integrated into the CPU package. Back in the old days when there were external northbridges, they controlled pretty much everything, and were the central hub of communication. If you wanted a multi-socketed motherboard back then, you needed one with a northbridge that was capable of it.
That doesn't apply anymore.
Now that it's integrated into the CPU, you can't swap it out. Today's "northbridges" (for lack of a better term) has a pre-determined set of functions and you're stuck with them. For desktop parts like AM4 and 1151, none of those CPUs were designed to be capable of multi-socket setups.
You might be thinking "what about the chipset on the motherboard" but that's basically what the old southbridge used to be. It still for the most part functions the same way, maybe taking on a few extra responsibilities here and there. In modern motherboards capable of multi-socket CPUs, that chipset ("southbridge") isn't what makes that happen.
To further emphasize my point, consider these 2 products:
Both CPUs are nearly identical. They have pretty much all the same specs, they're the same generation, same everything. Both models will work in any Epyc-compatible motherboard [alone]. And yet, the non-P model is $1300 more expensive. What's the difference you might ask? The 7551 can be paired up with another 7551 in a dual-socket motherboard, whereas the P model can't. Why? because the on-die chipset is different.
What this means is whether you're going for AM4 or TR4, you can't just slap in a new/better chipset and another CPU socket. The CPUs themselves determine whether they can be paired up. I'm quite confident this is no different for Intel.
In the case of Epyc, I'm not sure if there's actually a physical difference between the P and non-P models (or if it's just a firmware difference), but my point is Epyc was built with multi-socket in mind and the CPUs themselves are now what determines whether they can be paired up or not. AM4 was never built with this capability, and therefore, it isn't physically possible to have multiple AM4 sockets interconnected on a single motherboard.
Besides, I don't really see what the reason for a multi-socket AM4. Might as well just go for Threadripper or Epyc for what would likely cost roughly the same.
A high-end chipset doesn't mean it will support multi-socket boards. I will bet you your dream PC that a dual-socket AM4 motherboard will never happen.
No dreams about it, just a possibility, and no, never said that higher end chipsets means dual socket support, but it could enable it, and seems AM4 has no limitation at all about that, the limitation would be in the CPU-Motherboard-Chipset relationship, so since we don't acctually know about all details about the 3 since 2 of them are not present right now, i don't discard it, yes i don't think this is the highest probability, but i don't think it's impossible either.
You know, in this hardware business there are a lot of secret weapons to be deployed the fact that AMD did not disclosed it does not mean that they did not had this idea in mind, afterall we would not be speculating the existence of a x590 chipset here if everything was direct, AMD would have said openly about the 16core ryzen from the getgo, the 5700xt aniversary edition would have been disclosed to the jornalists, etc...
A dual socket motherboard would be for a niche, and i believe AMD wants to show that they aren't a cheap value oriented brand, as they've said already on interviews, so overclockers with top scores, and content creators that already have an AM4 processor and don't want to go with the Threadripper or EPYC, maybe could buy a very cheap zen or zen 2 processor and have a 16 core PC right now with future support for upgrades on the AM4.
There is no immediate need for dual socket on the desktop, when a single AM4 socket holds full 16 SMT cores in a multi-chip package for Ryzen 9 3850X (and a single TR4 socket would hold up to 64 SMT cores in the upcoming Threadripper 3000 series), with fast on-die links for cache coherent NUMA access to local memory.
Multi-socket motherboard would make sense for cloud servers, which have to pack a single rack unit with as many CPU dies and gigabytes of DRAM as it is physically possible; and even then, each socket (i.e. multi chip package) has its own local memory, and far (inter-socket) memory access performance is only a fraction of near (on-die) performance.
Its possible on am4, the zen 1 die has phsyically has 32 pcie lanes, am4 has enough extra pins to wire up an xgmi interconnect between sockets, each cpu would use half of its pcie lanes to provide the interconnect. Leaving 32 lanes for io (16+16)
Epyc 2p already does this.
Now i dont know if the io die for consumer zen 2 has 32 lanes , but the same principle applies.
That said there is no reason amd would do this as threadripper exists, which makes such a config pointless for desktop, why buy a dual am4 board , when a threadripper can do the same with less latency and more io.
First: how exactly do you know what pins are deemed "extra" in AM4?
Second: Even if you're right that there are enough pins for an XGMI interconnect, the fact of the matter is none of the existing AM4 CPUs are set up that way...
Third: It takes more than just PCIe lanes to bridge a couple of CPUs; that is a tremendous oversimplification. If it were really that simple, we'd have seen PCIe CPU expansion cards for years, but, the only thing we got anywhere close to that was a the Xeon Phi from several years ago, which as we all know, doesn't at all operate the same way a second socketed CPU would.
AMD uses pcie lane phys for its Xgmi interconnect on its 2p(dual socket) epyc servers, they use the exact same dies as zen 1, it was designed to do this.
as for extra pins, you can estimate this from other sockets plus the fact zen is basically an soc, all of the pins are for either pcie , usb, power and dram, that and most if not all sockets have reserved pins for debug or future use.
There was also the z490 chipset which was cancelled, but confirmed to exist by board oems last year, that would have added more pcie lanes.
Yes... and? Their server chips are designed very differently. They share the same core architecture but there's a lot of in-between differences. Again: PCIe lanes is not the only thing used to connect 2 CPUs.
Sure, but that doesn't mean there's enough.
It is on epyc, there are no other data paths, unless you consider LPC or serial.
the Server platform is essentially identical to TR4, The chip used for all gen1 ryzen/epyc/threadripper is called Zepplin is very flexible in how it can be implemented.
I see no reason why it wouldn't work, when it is clearly within its capability. it would just be 1/4 of 2p epyc.
If you want a mathematically consistent answer , you have 4094 pins to supply 128lanes of pcie and 8channel memory and power, for 4 dies ryzen is 1 die, 4094/4= 1023pins per die, for 32lanes of pcie , power , and 2 channels of memory.
AM4 has 1331 pins, some of these pins are used for graphics, most of which will be used for power (if its needed), since display io is maybe50-75 pins pins max, so you are left with a couple hundred pins that serve no obvious purpose, and thats with 32lanes accounted for, which is well within the range needed for such an interconnect.
On Socket AM4, only 47 (forty seven) pins are RSVD (reserved - not connected) , and 22 of them are located immediately around VDDP pins (+5V power rail) - so it is unlikely these can be reassigned as additional PCIe pins.
On the contrary, Socket TR4 has 902 (nine hundred and two) RSVD pins, of which the I/O area has a total of 240 RSVD pins pre-arranged and interspersed with Vcc/GND to double the number of existing PCIe lanes from 64 to 128 on the Socket SP3 EPYC processors, and the remaining pins are used to double memory channels from 4 to 8.