Here is an AMD SP5 socket (LGA6096) for EPYC Zen4 Genoa processors - (updated with photo)

#6008292 Posted on: 04/10/2022 12:30 AM

---

all manufacturing businesses start at the high end as that's where the profit margin is.

it takes time to scale up volume enough to make money off of a 15% margin which is where a low end mass market part is at (for the most part & yeah exceptions to every rule).

which is why yield has a direct relationship to profitability - AMD has gone to chiplets which are high yield by design and also have a lower cost to produce. Intel decided to flood the market with AL and eat a lot of the production costs, which has to be measured in years not just dollars as every chip is monolithic and a lower yield, albeit at a newer more efficient process that's finally ready for "primetime".

be not afraid (of socket size) :p
these are the products that make everything on the internet less expensive and more feature rich.
there are services just waiting to be invented as well as services that blow your socks off.
aka "the cloud" 3.0

Yes! The number of pins the CPU uses directly relates to the number of cores/chiplets the CPU support...but note that these motherboards have to be able to take the massive 128-core Zen 4 beasties, all the way down to 32-cores, maybe. You cannot very well have the motherboards customized for each CPU because then you'd have to replace the mboard every time you upgraded the CPU. One of the huge attractions to the AMD designs are the drop-in-replacement characteristics, imo. I very much doubt that anyone buying a 128-core AMD CPU is going to balk at the number of socket pins. "I know this is the fastest Enterprise CPU available on planet Earth right now, hands down, but it's got too many pins in the socket so I don't want it."--I don't consider that a likely scenario at all...

Yes, things are getting very complex, I agree. But like tunejunky says, it's nothing to get your panties in a wad over, imo...

#6008301 Posted on: 04/10/2022 02:32 AM

---

Low end CPUs are not gonna be made off a same wafers so I don't expect them to get more expensive, at least not in relation to specialized CPUs like these. On the other hand I can't think of any reason to buy this CPU for regular use. Unless you are operating a large scale data center you don't need it. And more and more organizations which had their own data centers are eliminating them and moving to cloud. And cloud operators will buy servers with these instead.
Nevertheless it is very large CPU for sure and I thought the electronics are getting smaller, boy was I wrong .
I'm aware. That's not what I'm concerned about. My concern is the cost of the CPU package but more importantly the motherboard traces. A socket that huge cannot scale down well in cost.

Just looked up the prices of a 5950x and a Epyc 7282, both 16 cores and the price difference is just 150€ higher for the Epyc.
It is one of the slowest 16 core Epyc, but still super cheap for 128 pci lanes, 8 channel memory on servergrade hardware.
That isn't quite an apples to apples comparison, since the 7282 is comprised of many low-binned chiplets, whereas the 5950X has a pair of pristine chiplets. As mentioned before, it's not the CPU itself, it's the motherboard. Motherboards will have to be bigger and far more complex to manufacture. So, while the cost of the CPU itself likely won't go up too dramatically (maybe another 25-50) the cost of the motherboards will most likely go way up.

#6008376 Posted on: 04/10/2022 01:28 PM

---

It is possible to buy a Tyan Barebone server kit with everything except the CPU, memory and storage for around 1000€.
It has dual powersupply for redundancy and 10 hot swap 2.5" drive bays.

Putting together a Alderlake system with the same components would end up in more then 700€ too, so the Price difference for a business server compared to a hobby pc is not that great.
The Epyc memory and PCI is more complex then regular hobby computers and that part cost more, I agree yes,
but look at the Alderlake VRM compared to a Epyc VRM, the Alderlake VRM is 2-4 times the size and complexity compared to a Epyc board.
Routing DDR5 is also much harder to do then DDR4 apparently, because 16 sticks of DDR4 looks to be no problem on epyc, but 4 sticks of DDR5 on Alderlake is risky.

#6008391 Posted on: 04/10/2022 02:50 PM

---

no need.
the number of non-functional chiplets is so low as to be a footnote. the uArch was designed for the process, squaring the circle (literally, as in a wafer) while minimizing e-waste.
this is where the partnership with TSMC comes in handy from all of their ARM lithography. the chiplets aren't much smaller than a Snapdragon and TSMC has honed the yield of tiny chips over a long time.

Intel on the other hand ironically makes tiny chips from mature (i.e. huge) processes, so they aren't "tiny tiny" like TSMC but run from 55n to 14n - like your wifi or router. those have more e-waste but they're cheap to produce as the fabs have long been paid for.

#6008578 Posted on: 04/11/2022 12:05 PM

---

Good heavens, 6,096 pins is enormous. I don't recall seeing any chip with that many.

For my purposes I am not sure if I would like the 3D cache or higher clock rates. I suspect if its only a 10 or 20% drop in frequency then the larger cache will likely make up for it. My application really benefits from larger cache as I have seen comparing it on an EPYC and a Threadripper with the same number of cores and similar clock rates. The EPYC is considerably faster, likely due to the cache but possibly also because more memory channels were used.

Higher clock rate only brings limited performance benefits as the main bottleneck for any processor is memory bandwidth. It doesn't matter how fast you clock a processor, if the code is waiting for data to arrive from the memory subsystem. Improving the performance of the memory subsystem through larger caches and a better cache controller is more beneficial than squeezing another few percentage points out of the core computing hardware. That is why artificial benchmarks are so deceptive, those tend to run completely in L1 cache, so the real bottlenecks of a system are never exposed.

You see it with the Ryzen 7 5800X3D. The artificial benchmarks that target execution speed all show a deficit in performance, while real world applications do show benefits from the larger L3 cache.