Death of enthusiast HEDT? Maybe not.

[spikeb]

I've often wished desktop motherboards+CPUs (or parts within 1.5x in cost) had more memory channels, max memory and PCIe. Even something simple like a home NAS with 5 x SATA, 2x10G network, and 2 NVMe start to push the available PCIe lanes of an Ryzen system.

Sure the Threadripper PRO 7955wx is an amazing chip, but costs $1,900 for the CPU and the motherboards aren't cheap.

Seems like years ago it was easy to upgrade to a CPU with 4 memory channels without a huge price premium, at least from Intel.

I noticed the AMD EPYC 8124P, compared to the Ryzen 7950x, has the same number of cores, 3x the memory bandwidth, and 4x the PCIe lanes. Sure the cores are Zen 4c instead of Zen4, so less cache. But at least for some workloads 3x the memory bandwidth will help. As a bonus, the AMD Epyc 8124P is *CHEAPER* than the Ryzen 7950x and lower power. Not to mention another 8 cores is only an extra $215 or so.

Seems like a pretty promising desktop/HEDT or small home server and price competitive with AM5 chips and crazy cheaper than the Threadripper PRO that costs $1,400 to $9,999. Think of the 8124P as 75% of the Threadripper PRO's 7955wx bandwidth, 75% of the PCIe lanes, 1/3rd the cost, and 1/3rd the power.

Thoughts?

 

[BlueFox]

It's also half the clock speed of a 7950X and will perform considerably worse under the majority of workloads.

 

[CyklonDX]

I think people forgot that x79/x99 mobo did cost like 600-1.2k usd in past, and cpu's could cost like 1k or more for broadwell. The premium was extreme especially in broadwell case (top-end cpu i7-6950x was 1700-1900 usd, while normal high-end desktop part was like 350-500 usd)

They became obtainable only decent chunk of years after in humane pricing.

 

[DaveLTX]

While there was X99 boards that came in at 250usd, they were pretty devoid of anything you might actually want in a board. I know, I had two. And they were NOT reliable
X299 wasn't particularly cheap either and the CPUs well they sucked.
I still have a X299 with 7940X
Wouldn't really be fair if you compared to Intel of old who would sell you a 600usd 6 core and it was limited to 28 pcie lanes as well

Now the boards are built to a far greater degree than they used to

TL;DR we get a lot more features and cores than we used to, surely that's got to be more expensive

 

[spikeb]

It's also half the clock speed of a 7950X and will perform considerably worse under the majority of workloads.

For cache friendly stuff, sure. Often not by much. For SpecIntRate (cache friendly) The 8124P is 76% (150 vs 195) as a 7950X despite having a base clock only 54% as fast. Adding 8 more cores with the 8224P (costing $155 more than the 7950X) wins with a score of 218.

However for more memory intensive workloads like SpecFpRate the 8124P is 1.41 times faster than the 7950X (182 vs 129). Sadly I don't have Threadripper numbers for SpecIntRate or SpecFPRate.

In any case I think the Epyc sienna is a pretty interesting option proving ryzen 7/ryzen 9 performance at Ryzen 7/9 prices, while having substantial advantages if you need more PCIe lanes or bandwidth and avoiding the power (350 watts) and cost ($1800) of threadripper pro.

 

[Bert]

Your title says HEDT but your post talks about NAS. Are you looking for HEDT or Server or workstation?

 

[bayleyw]

There are so many better used options that it's hard to justify a new 8124P for home use, especially since DDR5 is *not a feature* at the moment unless you are doing CFD. You can get 4x the cores at the same price with Rome, the same multithreaded performance and slightly less single threaded performance at 1/8 the price with Skylake, or 4x the cores with the same single-threaded performance with Milan.

 

[unwind-protect]

I've often wished desktop motherboards+CPUs (or parts within 1.5x in cost) had more memory channels, max memory and PCIe. Even something simple like a home NAS with 5 x SATA, 2x10G network, and 2 NVMe start to push the available PCIe lanes of an Ryzen system.

Sure the Threadripper PRO 7955wx is an amazing chip, but costs $1,900 for the CPU and the motherboards aren't cheap.

Seems like years ago it was easy to upgrade to a CPU with 4 memory channels without a huge price premium, at least from Intel.

I noticed the AMD EPYC 8124P, compared to the Ryzen 7950x, has the same number of cores, 3x the memory bandwidth, and 4x the PCIe lanes. Sure the cores are Zen 4c instead of Zen4, so less cache. But at least for some workloads 3x the memory bandwidth will help. As a bonus, the AMD Epyc 8124P is *CHEAPER* than the Ryzen 7950x and lower power. Not to mention another 8 cores is only an extra $215 or so.

Seems like a pretty promising desktop/HEDT or small home server and price competitive with AM5 chips and crazy cheaper than the Threadripper PRO that costs $1,400 to $9,999. Think of the 8124P as 75% of the Threadripper PRO's 7955wx bandwidth, 75% of the PCIe lanes, 1/3rd the cost, and 1/3rd the power.

Thoughts?

I don't disagree with what you said. However, if you are willing to accept lower per-core speed and want to bet on memory bandwidth then older, used systems are also a way to go.

 

[BlueFox]

For cache friendly stuff, sure. Often not by much. For SpecIntRate (cache friendly) The 8124P is 76% (150 vs 195) as a 7950X despite having a base clock only 54% as fast. Adding 8 more cores with the 8224P (costing $155 more than the 7950X) wins with a score of 218.

However for more memory intensive workloads like SpecFpRate the 8124P is 1.41 times faster than the 7950X (182 vs 129). Sadly I don't have Threadripper numbers for SpecIntRate or SpecFPRate.

In any case I think the Epyc sienna is a pretty interesting option proving ryzen 7/ryzen 9 performance at Ryzen 7/9 prices, while having substantial advantages if you need more PCIe lanes or bandwidth and avoiding the power (350 watts) and cost ($1800) of threadripper pro.

Synthetic benchmarks are not necessarily representative of actual usage. It is a desktop after all and I still stand by my statement that in the majority of workloads, the Ryzen CPU will be considerably faster.

 

[Bert]

Synthetic benchmarks are not necessarily representative of actual usage. It is a desktop after all and I still stand by my statement that in the majority of workloads, the Ryzen CPU will be considerably faster.

I believe this is true. Most consumer applications cannot take advantage of the high memory bandwidth. Unless the application needs extra connections such as gpu or storage intensive workloads. Theae are workstation workloads mote than hedt.

I am running 10980xe, amazing compute capacity if you can cool it but way behind 10900k for responsiveness on the desktop applications.

 

[CyklonDX]

(sorry for longpost / rant)

The actual performance is based on what is being done by this system.
The high-end desktop chips have always outshined server/workstation parts in single core performance, and now that GPU does most work the single core performance matters even more;

Part1

I think due to intel marketing there's been a split in understanding what is what. Maybe its also mine misunderstanding as all those are just marketing terms - and they change always to sell 'better'.

(as example i'll use sandy bridge)
There were Desktop parts the going up-to i5 (like 2500k), and you had premium for HT i7 which were branded as High End Desktop part, then as next tier we had gotten workstation systems, one aimed at enthusiast gamers with a single server socket LGA2011 and cpu's with unlocked multipliers so one could overclock them, and true workstation system that had dual server cpu's. The only thing that differed those workstation systems with servers was its form factor, noise, and its reliance on video card. All in all, it was/still are servers just boxed as a desktop.
(ryzen platform threw a wench into intel's ecosystem, and broke it all - with its support of ecc, as well as being physically ~2 "cpu's" on single chip.)

Part2

Now going back to original point i was making,
Most often in past the we relied on cpu to do all the workload. The system, rendering, calculation whatever... all was done on cpu. There was no real GPU offload until our recent history. Those systems had to have many cores, as person would want to lets say render his scene in 3dsmax, do some photoshop/corel draw work, and maybe watch a movie all at the same time. They needed more cpu, and control over them.

Today all those needs for that have been eliminated. The cpu's having 8+ cores is enough for everything as its mostly about your GPU/s now, and there's nothing as easy even on cheaper motherboards to add 2nd, even 3rd gpu, with your 8c system to accomplish all your needs without cry, and choking as cpu isn't all that important anymore for high-end desktop as such. (all those synthetic benchmarks like blender's test on cpu, is just that... a test. As blender itself wants you to use your gpu for rendering, and no1 serious would use cpu's anymore. A compilation work, another benchmark often presented to show how good the cpu's are... are the so useful in real life? Lets be clear... 20sec didn't kill a cat, only killed a tea break when ceased to take couple minute. No real benefit here... except maybe for Linus Tarvolds who compiles kernels, and rages when he has to recompile, because it takes too much time out of his already short life left.

Part3

People who say they need 24+ cores, 256G of ram or more, more memory bandwidth most often don't really use more than 10c, and rarely even get to 60G of ram, as they are often just playing games, maybe watching a movie on 2nd screen. Sometimes boinc or cryptomining.

The current real needs that come into workstation territory is almost always solely for kvm/vm setups where 1 os is just not enough. The seperation of powers is needed. (and ryzen is stealing the show in that area, as chips most often are already physically divided) A lot of people moved from their dual sandy/ivy servers/workstations to ryzen/threadripper platforms as there was seperations they wanted for cranking up vms.

End

Obviously if you were to crank up load to lets say 4 equal vm's without shared resources on those today high-end desktop systems like ryzen 7900x or intel's equilivant they would choke; and what you should have gotten was a workstation system instead, that has 2 sockets, 8+ ram slots, much larger memory bandwidth, many more gpu's, and many more io devices.

Its not the intel/amd not making those desktop grade systems like threadripper, its the use case of them, and price -
The reality is that they aren't really needed anymore, and very few buys them... and its mosty just for show rather than actual use. Most often at that point in life person who is smart enough would buy a server/workstation system with green pcb's because he hates the world, and useless fancy expensive rpg-flashing desktop parts.