Can I escape ThreadRipper PRO with AM5?

[CyklonDX]

If you must go with intel I would advise going with xeon platinum 8124M instead (at least you'll have 2 avx-512 pipes)
(socket 3647), its not that expensive. (18c @3GHz@240W for around 200 usd) performance wise it should be around 5950x ~ but the pcie lanes on that... wohoo 48 lanes.
Price for full build i think it'll come to same as full ryzen 5900x build.

(mobo will be expensive, but you'll save a ton on cpu)

 

[acquacow]

I'm waiting for the 6950K to drop in price for my old x99 currently running a 1620 v3 may just throw a 2667 V3 in there...

I paid $165 for mine on ebay, how low are you looking for it to go?

 

[T_Minus]

I paid $165 for mine on ebay, how low are you looking for it to go?

Cheapest right now are $200 it looks like :/

Way too much IMO when the last gen AMD are priced what they are now used.
$200-250 for CPU\RAM\MOBO with +1000 passmark single core

Of course not the PCIE lanes...

I'd pay up to 90$ i'm cheap when it comes to OLD desktop gear where single core has made such improvements lately.

 

[acquacow]

Cheapest right now are $200 it looks like :/

Way too much IMO when the last gen AMD are priced what they are now used.
$200-250 for CPU\RAM\MOBO with +1000 passmark single core

Of course not the PCIE lanes...

I'd pay up to 90$ i'm cheap when it comes to OLD desktop gear where single core has made such improvements lately.

Yup, I get that.

The most CPU intensive thing I really run is Lightroom, but it's slow no matter what I've run it on. I'm about to pull the trigger on an AMD 7950x build, and the IPC gain alone over this i7-6950x is pretty ridiculous. Add in DDR5 and some pci-e 5.0 storage... and I'm sure Lightroom will still keep me waiting, lol!

I held out a few generations, but I really like the numbering of going from an intel 6950x to an amd 7950x... seems like it was meant to be =)

 

[Pakna]

I'm waiting for the 6950K to drop in price for my old x99 currently running a 1620 v3 may just throw a 2667 V3 in there...

How about something like a E5-2689 v4 or E5-2687W v4? I can't really say if I'd percieve the single-core clock drops (my 5820k is running at 4.2 GHz from day one).

 

[MichalPL]

How about something like a E5-2689 v4 or E5-2687W v4? I can't really say if I'd percieve the single-core clock drops (my 5820k is running at 4.2 GHz from day one).

Can you overclock v4 ?

My experience is:
1620 <- 4.2GHz all core without increase voltage
1650 v2 <- 4.1GHz all core without increase voltage (some models 4.2)
1680 v2 <- 4.1GHz all core without increase voltage
1660 <- 4.8GHz all core at ~1.45V (yes it is 32nm CPU!)
1650 v2 <- 4.8GHz all core at ~1.45V
1660 v3 <- 5.1GHz single core / 4.9GHz all core at ~1.45V

btw. 8894 v4 in multicore (only) is very similar to the 13900kf stock (single core twice speed).

 

[MichalPL]

(A workstation from 2013 can have 2x v2 14c, 768GB of ram, lot pcie lanes, you could get 6-8 pcie gen3 x8...)

We were considering something like this ~3.5-4 years ago and problem was:
-10c v2 are almost the same in multicore, and better single core
-8c was just a bit slower (and boosting 4.0GHz)
-1660v3 after overclocking was much faster as a workstation
-1year later Ryzen 5000 arrived

Still having few old servers (3 years ago worth maybe ~$150) some with 3 dimms per channel (768) some with 2 dimms (512 max) - they are good ZFS backuper when 2x 2687 v2 inserted and boosting to 4.0GHz, -100GbE partly make sense. (4GHz v2/v3/v4 can achieve about 7GB/s compressed ZFS - 13900kf will be much faster , overclocked 1650 v2 can almost saturate 100GbE)

 

[MichalPL]

CMIIW, but in Mellanox-land, I'd have to go all the way up to connectX-6 to get gen4. Shame, because 3-series dual 40GbEs are down to $25, but tossing that in a x4 slot turns it into a 32gb NIC.

No, ConnectX-5 support PCIe 4.0, if your model don't you can flash them to support it.
For 40GbE 4 lanes PCIe 3.0 is almost ok ~3700MB/s instead of 4420MB/s

 

[MichalPL]

Octal socket is a bad meme, please folks don't buy an 8S system.

Why? they have lot of PCIe lanes and are quite fast (comparing to dual Epyc)

 

[Pakna]

Can you overclock v4 ?

My experience is:
1620 <- 4.2GHz all core without increase voltage
1650 v2 <- 4.1GHz all core without increase voltage (some models 4.2)
1680 v2 <- 4.1GHz all core without increase voltage
1660 <- 4.8GHz all core at ~1.45V (yes it is 32nm CPU!)
1650 v2 <- 4.8GHz all core at ~1.45V
1660 v3 <- 5.1GHz single core / 4.9GHz all core at ~1.45V

btw. 8894 v4 in multicore (only) is very similar to the 13900kf stock (single core twice speed).

I was unaware E3/E5 CPUs were overclockable at all - don't they all have their multipliers locked? Also, from what I remember, BCLK can move up a bit but doing so yields very little gain and compromises stability?

 

[zer0sum]

Awesome topic!!

Does anyone have any recommendations for a Micro ATX board?
I have Asrock Rack X470D4U right now and it's running a nice little Proxmox node with a 5900X, 10G NIC, and a few nvme drives, but I'm always looking for moar powerrrr

It has a pretty weird allocation of PCIe lanes:

• SLOT6: PCIe3.0 x16 (splittable in x4/4/4/4), auto-switch to x8 when SLOT4 is occupied)
• SLOT4: PCIe x16 (PCIe3.0 x8)
• PCIe x 8 SLOT5: PCIe3.0 x4 link

M.2 slot 2 x SATA 6Gb/s or PCIe3.0 x2 and supports PCIe2.0 x4 only)

 

[MichalPL]

I was unaware E3/E5 CPUs were overclockable at all - don't they all have their multipliers locked? Also, from what I remember, BCLK can move up a bit but doing so yields very little gain and compromises stability?

Some of them (only single socket one) was unlocked, now (after Ryzen 5000) it is a historical info, because all of them are slower even than 5600X.

short list from my head and more less default boost frequency of the unlocked E5 (all of them 40 PCIe lanes):
E5 1620 (4 core ~3.9GHz)
E5 1650 (6 core ~3.8GHz)
E5 1660 (6 core ~4.0GHz)
E5 1650 v2 (6 core ~3.9GHz)
E5 1660 v2 (8 core ~3.9GHz)
E5 1650 v3 (6 core ~3.9GHz)
E5 1660 v3 (8 core ~3.9GHz)
E5 1680 v3 (8 core ~3.9GHz)
E5 1681 v3 (10 core ~3.9GHz)
E5 1683 v3 (12 core ~3.8GHz) <- maybe this one will be faster (after overclocking) than 5600X multicore only
E5 1689 v3 (14 core ~3.8GHz) <- and also maybe this one will be faster (after crazy overclocking) than 5800X multicore only

All of them you can overclock to almost 5GHz with the proper cooling (~3000 passmark points in v3 case)
Edit: 1620 was good, but 1650/1660 (without v2 or v3) was slow and don't /support/ this number of cores properly (v2 and v3 are ok and scale properly)

 

[MichalPL]

BCLK can move up a bit but doing so yields very little gain and compromises stability?

I didn't try this, "too risky". Overclocked Xeons was ok working weeks without crash (and years as a server). What was on the edge was overclocking memory to 2133MHz (Samsung DDR3 ECC) on 1650 v2 if I remember corectly E5 v3 was limited to 2400 or 2666MHz DDR4 - so no memory overclocking

 

[bayleyw]

Why? they have lot of PCIe lanes and are quite fast (comparing to dual Epyc)

Because they only perform well for select benchmarks (e.g. SPECrate) that require zero interprocess communication. You can get crazy high SPECINTrate, Passmark, or Geekbench scores out of them, but otherwise performance is a crapshoot and you end up with anomalous behavior where your application doesn't perform very well at all.

It's roughly the same situation that Naples wound up in - 2S Naples is logically an 8S system, and it performed well in throughput-type enterprise applications and virtualization and poorly in HPC/workstation scenarios that required a lot of syncing between threads.

Honestly the real deakbreaker is the fact that 8S Broadwell is a solid *two horsepower* of CPUs, not counting the fans and accessories. When you're looking at a 2.5KW total system draw things start to get real hairy, you run into nonsense like "my room wasn't big enough so if I'm in the same room as my computer, one of us will overheat" and "99% of American families can't use this computer because it trips their 110V breakers", not to mention "I'm in the hospital because I tried to lift my computer and it weighs 100kg".

I've owned my fair share of 4S+ systems including a X9QRi (barely made the cut as a workstation, and only because DDR4 was $10/GB back then!), an 8S SGI (hot garbage, only ran a modified version of SLES) and, long ago, the utterly absurd X3950 M2 which was a sixteen socket Dunnington system offering 96 cores in 2009. All of them were very bad.

 

[CyklonDX]

We were considering something like this ~3.5-4 years ago and problem was:
-10c v2 are almost the same in multicore, and better single core
-8c was just a bit slower (and boosting 4.0GHz)
-1660v3 after overclocking was much faster as a workstation
-1year later Ryzen 5000 arrived

Still having few old servers (3 years ago worth maybe ~$150) some with 3 dimms per channel (768) some with 2 dimms (512 max) - they are good ZFS backuper when 2x 2687 v2 inserted and boosting to 4.0GHz, -100GbE partly make sense. (4GHz v2/v3/v4 can achieve about 7GB/s compressed ZFS - 13900kf will be much faster , overclocked 1650 v2 can almost saturate 100GbE)

In terms of populating 100GbE, its more possible if you have more worker threads on your driver.
Overclocking is a stock way to get there, proper way is to utilize many more cores instead. Its only a question of tunning, by default on rhel as example comes in intel driver compiled only using 4 worker threads instead, and thus yes, your single thread performance matters a lot.

(An e5-2687W v2 8c would be as fast as Platinum 8124M in such setup)



but if you actually recompile the driver, and mount the module with more workers, and disable irqloadbalancer service, pin those workers to specific threads you can easily saturate 100GbE without OC without stress... (as long as you can saturate it in first place, memory, etc... you are still limited by storage, and memory speed.)




in ref.

Intel Xeon E5-2697 v2 @ 2.70GHz vs Intel Xeon E5-2687W v2 @ 3.40GHz vs Intel Xeon E5-2690 v4 @ 2.60GHz vs Intel Xeon Platinum 8124M @ 3.00GHz vs AMD Ryzen 9 5900X [cpubenchmark.net] by PassMark Software

www.cpubenchmark.net

(in case of just having compiled driver to use 4 cores, a intel i7 4820K can easily overclock to 5GHz without touching fsb clocks... and that would do for you.)

its possible - done that before (screenshot from 7 years ago)

 

[DaveLTX]

Can you overclock v4 ?

My experience is:
1620 <- 4.2GHz all core without increase voltage
1650 v2 <- 4.1GHz all core without increase voltage (some models 4.2)
1680 v2 <- 4.1GHz all core without increase voltage
1660 <- 4.8GHz all core at ~1.45V (yes it is 32nm CPU!)
1650 v2 <- 4.8GHz all core at ~1.45V
1660 v3 <- 5.1GHz single core / 4.9GHz all core at ~1.45V

btw. 8894 v4 in multicore (only) is very similar to the 13900kf stock (single core twice speed).

Not 26xx v4. Again... passmark is terrible.
You need TWO 8894v4 to be faster than a single 5950X which is usually power limited when all cores fire up anywhere but in Cinebench R15,
Two 8894v4 scores about 7734 in CB R15 MT
A single 5950X scores 4483 without any overclocking in R15 MT!
And also cross CPU latency on 8 CPUs is shocking

Why? they have lot of PCIe lanes and are quite fast (comparing to dual Epyc)

A dual EPYC 7702 let alone 7763 will eat a 8 socket system with a ton of NUMAs and far less power draw as well

If you must go with intel I would advise going with xeon platinum 8124M instead (at least you'll have 2 avx-512 pipes)
(socket 3647), its not that expensive. (18c @3GHz@240W for around 200 usd) performance wise it should be around 5950x ~ but the pcie lanes on that... wohoo 48 lanes.
Price for full build i think it'll come to same as full ryzen 5900x build.

(mobo will be expensive, but you'll save a ton on cpu)

I got my tyan epyc S8030 motherboard for 200USD and 16 core EPYC 7002s can be had for 200usd or so.
my problem with 3647 boards are that they are usually quite expensive for a usable form factor, not those giant boards that amazon used...
I have 3 6138s with no board to go in

 

[CyklonDX]

I got my tyan epyc S8030 motherboard for 200USD and 16 core EPYC 7002s can be had for 200usd or so.

in that case i wouldn't look back at 3647 at all. The AMD system is superior even in single core perf.

 

[DaveLTX]

in that case i wouldn't look back at 3647 at all. The AMD system is superior even in single core perf.

Absolutely, way higher IPC, more cores, more PCIe

 

[MichalPL]

Because they only perform well for select benchmarks (e.g. SPECrate) that require zero interprocess communication. You can get crazy high SPECINTrate, Passmark, or Geekbench scores out of them, but otherwise performance is a crapshoot and you end up with anomalous behavior where your application doesn't perform very well at all.

It's roughly the same situation that Naples wound up in - 2S Naples is logically an 8S system, and it performed well in throughput-type enterprise applications and virtualization and poorly in HPC/workstation scenarios that required a lot of syncing between threads.

Hmm.... agree - in the most cases.
Yes for the (home!) workstation 100kg computer will be a killer. I also prefer i5 13gen (or 7700X) instead of any S2 v2 /v3 Xeon - it is much faster

But they are not that slow as you are saing... when you need to compile lot of C++ code (independent tasks) 8S really works, if the single core is not super slow. But now Epyc is still super slow in terms of the single core performance. So no difference between 2S or 8S.

Ideally will be just one TR 7995WX overclocked to 5.8GHz all 96cores 128 PCIe 5.0 lanes. Plus GF 4090 to play CoD, Age IV, Terminator and NFS

Honestly the real deakbreaker is the fact that 8S Broadwell is a solid *two horsepower* of CPUs, not counting the fans and accessories. When you're looking at a 2.5KW total system draw things start to get real hairy, you run into nonsense like "my room wasn't big enough so if I'm in the same room as my computer, one of us will overheat" and "99% of American families can't use this computer because it trips their 110V breakers", not to mention "I'm in the hospital because I tried to lift my computer and it weighs 100kg".

I've owned my fair share of 4S+ systems including a X9QRi (barely made the cut as a workstation, and only because DDR4 was $10/GB back then!), an 8S SGI (hot garbage, only ran a modified version of SLES) and, long ago, the utterly absurd X3950 M2 which was a sixteen socket Dunnington system offering 96 cores in 2009. All of them were very bad.

Actually I bought 8S Broadwell (not for the home, at home I am currently using ~250W Laptop. i9 with RTX) good for comiling C++ code.


But they are not that power hungry, ~690W if not compiling, yes it can do 2KW in peaks from time to time.
but i9 13gen + 2x GF 3090 can do 1KW

"my room wasn't big enough so if I'm in the same room as my computer, one of us will overheat" and "99% of American families can't use this computer because it trips their 110V breakers",

lol - true. I have added ~50EUR extra for a valves to be able to put all heat from the computers in hobby room into water to heat floors And can use heat pump to freeze it a bit (I hope) in summer or use home ground heat exchanger (in progress, to have ~10C in pipes "for free").

Here we have 230V (single plug is 3.5KW max) and in sunny days... 253V because in this area lot of solar panels.

 

[DaveLTX]

Hmm.... agree - in the most cases.
Yes for the (home!) workstation 100kg computer will be a killer. I also prefer i5 13gen (or 7700X) instead of any S2 v2 /v3 Xeon - it is much faster

But they are not that slow as you are saing... when you need to compile lot of C++ code (independent tasks) 8S really works, if the single core is not super slow. But now Epyc is still super slow in terms of the single core performance. So no difference between 2S or 8S.

Epyc is super slow? are you trolling on purpose now?