Can I escape ThreadRipper PRO with AM5?

Notice: Page may contain affiliate links for which we may earn a small commission through services like Amazon Affiliates or Skimlinks.

MichalPL

Active Member
Feb 10, 2019
189
25
28
Anything works on Zen 2 onwards... There's a reason new HPC and DCs are using Zen 3 now...
It's faster than Ice lake and cheaper.
Meanwhile Quad socket systems are on their way out due to physically being unable to fit all the DRAM possible and inter-system level NUMA latencies

The OCP servers I work on, on a daily basis are usually 1 or 2 socket.
Agree (in most cases ;) ) and this was the reason to make experiments like this:
2x2697v2 <- first builder
4x 8895v2 <- second builder
TR3970X <- third one
4x 8894v4 <- mistake (better to use 5950X, AMD is slower but not that much)
8x 8894v4 <- fifth try (yes, faster than 7950X and 13900kf, having one for testing and having more i9)
probably the future builder will be:
2x TR 8995WX overclocked 256 cores + used ddr5 memory ;) (or 7995WX depend of the timing)


I am a kind of former game engines rendering programmer, and no experience with servers - it's just a "hobby" (and ~2% of my work), but I can design CPU's and use C++, assembler, cores and cache in the proper way.
 

DaveLTX

Active Member
Dec 5, 2021
150
34
28
Agree (in most cases ;) ) and this was the reason to make experiments like this:
2x2697v2 <- first builder
4x 8895v2 <- second builder
TR3970X <- third one
4x 8894v4 <- mistake (better to use 5950X, AMD is slower but not that much)
8x 8894v4 <- fifth try (yes, faster than 7950X and 13900kf, having one for testing and having more i9)
probably the future builder will be:
2x TR 8995WX overclocked 256 cores + used ddr5 memory ;) (or 7995WX depend of the timing)


I am a kind of former game engines rendering programmer, and no experience with servers - it's just a "hobby" (and ~2% of my work), but I can design CPU's and use C++, assembler, cores and cache in the proper way.
You need to understand that EPYC is so much faster rather than throwing strange arguments that compares apples and oranges is what I have to say. If you say EPYC is slow, then so is everything you have is even slower.
 

MichalPL

Active Member
Feb 10, 2019
189
25
28
btw. @DaveLTX please make the CPU-Z test because I am fan of the Passmark you are fan of the CB CPU-Z will be independent.

And I am not fan of anything history of my PC's that I personally was working on: AMD 286, Intel 486, AMD K6, AMD A64, Intel Core2Duo, i5, 1650v2, 10885H.

Probaly your bar is... longer thank mine in CB 23. here is mine:
1671816722777.png

and meanwhile problems with the QPI links when all 8 CPUs connected together (Diskmark)


let's do a CPU-Z test !

1671817322920.png
 
Last edited:

DaveLTX

Active Member
Dec 5, 2021
150
34
28
btw. @DaveLTX please make the CPU-Z test because I am fan of the Passmark you are fan of the CB CPU-Z will be independent.

And I am not fan of anything history of my PC's that I personally was working on: AMD 286, Intel 486, AMD K6, AMD A64, Intel Core2Duo, i5, 1650v2, 10885H.

Probaly your bar is... longer in CB 23 here is mine:
View attachment 26273

and meanwhile problems with the QPI links when all 8 CPUs connected together (Diskmark)


let's do a CPU-Z test !
I am not a fan of CB. Everyone simply uses CB as a yardstick. Nobody uses Passmark for a reason because it is extremely biased and suspicious.
1671817135936.png
 

MichalPL

Active Member
Feb 10, 2019
189
25
28
Where are the pcie lanes?
Builders don't require lot of them

4xNVMe is very good especially when PCIE 4.0
1x NIC 2x 100GbE is ok
RTX connected to PCIe 3.0x4 ;)

memory is a problem - should be supported by NVMe drives if not possible to install 512GB.

so AM4 AM5 and Z790 is ok
 

MichalPL

Active Member
Feb 10, 2019
189
25
28
I am not a fan of CB. Everyone simply uses CB as a yardstick. Nobody uses Passmark for a reason because it is extremely biased and suspicious.
OK, your bar is (slightly) longer than mine! (what about CB R23?)
but.. others have much longer bars for example gamers! equipped with 13900kf.

so our CPU's are still.... sloooow!
 

Styp

Member
Aug 1, 2018
69
21
8
Sorry guys, but all this benchmarking is just bullshit in the end.

You build a workstation for your 'primary' purpose - full stop. If its gaming, go for an AMD latest Zen (or Intel). If you need more PCI-E etc. check what is available within your budget.

I had a coworker running a 12 core 5900x and I run an exotic LGA3647 - in 90% of the cases his machine is way faster, but as soon as I use the machine for the purpose it is build - I get the performance that I need. So stop discussing single core performance, if memory throughput and multi-core is your use-case, and wise versa. This discussion just leads nowhere, LGA 3647 is still a mighty platform but if you need to maximise on FPS - it's not the right for you!
 

DaveLTX

Active Member
Dec 5, 2021
150
34
28
Sorry guys, but all this benchmarking is just bullshit in the end.

You build a workstation for your 'primary' purpose - full stop. If its gaming, go for an AMD latest Zen (or Intel). If you need more PCI-E etc. check what is available within your budget.

I had a coworker running a 12 core 5900x and I run an exotic LGA3647 - in 90% of the cases his machine is way faster, but as soon as I use the machine for the purpose it is build - I get the performance that I need. So stop discussing single core performance, if memory throughput and multi-core is your use-case, and wise versa. This discussion just leads nowhere, LGA 3647 is still a mighty platform but if you need to maximise on FPS - it's not the right for you!
Which is what I have been trying to get through to him.
 

MichalPL

Active Member
Feb 10, 2019
189
25
28
I had a coworker running a 12 core 5900x
I skipped LGA3647 after bought LGA2066 10gen i9 and see almost same results as 1660v3 ;) (10 cores vs 8)

but you are right coworkers here are using i9 or 5950X (now first 7950X "arrived")
and when don't have to re-compile everything they are fine:

Spec is like this:

GPU slot divvied into 2x x8:

GPU typically 3070/3080 in PCIe 4.0 x8
2x Samsung PM9A1 1TB in x8 slot (configured as x4x4 on AMD possible on Intel not)
1x Samsung PM9A1 2TB in CPU NVME
1x Mellanox 40G in chipset PCIe

and slightly different between AM4 AM5 and Z790

and yes they are fast, but slower than S8, much faster than S8 and Epyc in linking
 

mattventura

Active Member
Nov 9, 2022
443
210
43
Sorry guys, but all this benchmarking is just bullshit in the end.

You build a workstation for your 'primary' purpose - full stop. If its gaming, go for an AMD latest Zen (or Intel). If you need more PCI-E etc. check what is available within your budget.
But the problem that I and others have is that we want both the gaming performance and the I/O. Previously, HEDT was perfect for this use case, but there's still no current-gen HEDT, only rumors of such.

Oh, and the fact that most of us would be fine taking the slight performance impact from using chipset lanes instead of CPU, except for the fact that you can only go up to x4 even if 8 lanes are available.
 
Last edited:
  • Like
Reactions: Pakna

CyklonDX

Well-Known Member
Nov 8, 2022
784
255
63
if you are comparing stats on passmark make sure the margin of error is not High. If it is, then it had 1-2 samples (and avges them out.)
Person who ran it, could've been in power saving etc...
 
Last edited:
  • Like
Reactions: RolloZ170

Patriot

Moderator
Apr 18, 2011
1,450
789
113
But the problem that I and others have is that we want both the gaming performance and the I/O. Previously, HEDT was perfect for this use case, but there's still no current-gen HEDT, only rumors of such.

Oh, and the fact that most of us would be fine taking the slight performance impact from using chipset lanes instead of CPU, except for the fact that you can only go up to x4 even if 8 lanes are available.
I guess I should do some fun experiments then. I have a 1680v3 @4.2ghz with 256GB of ram... and a 64c 7763 milan...
Time to compare gaming performance lol.

I honestly think @Pakna might have best use for a high frequency used Rome chip, or some ES rome overclocking.
The problem with using a server as a workstation is IO (sound and USB lacking). If it is desired... I can do some testing.
 

Pakna

Member
May 7, 2019
50
3
8
I guess I should do some fun experiments then. I have a 1680v3 @4.2ghz with 256GB of ram... and a 64c 7763 milan...
Time to compare gaming performance lol.
That would be super appreciated - I'd be really interested to see some benchmarks of overclocked E5-16XX v3/v4, especially if their core count > 6.

I honestly think @Pakna might have best use for a high frequency used Rome chip, or some ES rome overclocking.
The problem with using a server as a workstation is IO (sound and USB lacking). If it is desired... I can do some testing.
At the moment, it looks like a used i7-6950X would be about the highest-end part that I could possibly upgrade my system to so that I have 40 PCIe lanes. The lane count predicate is more for my own convenience than strict necessity - but if I am buying a new CPU from 2015/6 I'd absolutely want to min-max the heck out if it. Thus, whichever existing E3-16xx/E5-2xxx v3/v4 could give me the least amount of single-thread performance loss (say, within 10% margin?) without the loss of multi-threading performance (i.e. number of cores). I am also fine with ES/QS versions of chips, though I'd don't exactly know what am I potentially losing with these versions of chips and who is a reputable seller.

My work is coding non real-time apps in Java/Scala/Rust, not 3D rendering or AI. I prefer Linux over Windows (unless it's gaming) - if I play games it's 1600p and I am more than fine playing them at 30-60 fps.

I am attaching my existing Cinebench and CPU-Z results for reference - the CPU is running at 4.2 GHz and memory is running a DDR4-2400 XMP profile (FWIW). If we could get some benchmarks for comparison, that would be absolutely super appreciated.

single-thread_benchmark.PNG

multi-thread_cb.PNG
 

DaveLTX

Active Member
Dec 5, 2021
150
34
28
That would be super appreciated - I'd be really interested to see some benchmarks of overclocked E5-16XX v3/v4, especially if their core count > 6.



At the moment, it looks like a used i7-6950X would be about the highest-end part that I could possibly upgrade my system to so that I have 40 PCIe lanes. The lane count predicate is more for my own convenience than strict necessity - but if I am buying a new CPU from 2015/6 I'd absolutely want to min-max the heck out if it. Thus, whichever existing E3-16xx/E5-2xxx v3/v4 could give me the least amount of single-thread performance loss (say, within 10% margin?) without the loss of multi-threading performance (i.e. number of cores). I am also fine with ES/QS versions of chips, though I'd don't exactly know what am I potentially losing with these versions of chips and who is a reputable seller.

My work is coding non real-time apps in Java/Scala/Rust, not 3D rendering or AI. I prefer Linux over Windows (unless it's gaming) - if I play games it's 1600p and I am more than fine playing them at 30-60 fps.

I am attaching my existing Cinebench and CPU-Z results for reference - the CPU is running at 4.2 GHz and memory is running a DDR4-2400 XMP profile (FWIW). If we could get some benchmarks for comparison, that would be absolutely super appreciated.
From personal experience 7 years ago, its the 16xx V3/V4 non ES that will OC, the ES? won't.

Maybe you can look at a E5 2687W V4 as they are probably more abundant than 1660/80 V4 and you could use some bios hacks (Assuming you are using a X99 board) to crank a E5 2690 V4 to their max single core turbo speed for all cores. Power limit is still fused and it will stay under but you could bring up performance higher as apparently most Broadwells consistently stays far under their power draw at stock or you can disable some cores if you are bumping into power limit at the hacked frequency(i am not sure if you can, its been a while since I had experience with X99)
 

Pakna

Member
May 7, 2019
50
3
8
From personal experience 7 years ago, its the 16xx V3/V4 non ES that will OC, the ES? won't.

Maybe you can look at a E5 2687W V4 as they are probably more abundant than 1660/80 V4 and you could use some bios hacks (Assuming you are using a X99 board) to crank a E5 2690 V4 to their max single core turbo speed for all cores. Power limit is still fused and it will stay under but you could bring up performance higher as apparently most Broadwells consistently stays far under their power draw at stock or you can disable some cores if you are bumping into power limit at the hacked frequency(i am not sure if you can, its been a while since I had experience with X99)
Yes, I am using the Asus Sabertooth X99 (with latest official BIOS). E5-2687W v4 is exactly what I've been looking at for the last week as finding any E5-16xx north of E5-1650 v3 is very difficult or too expensive. Right now, I am seeing E5-2687W v4 going for 120 USD upwards on the fleaBay - is that a reasonable price or should I look for lower prices?

If you could provide some links on how to hack BIOS into overclocking such 2P CPUs with locked multiplier that would be great - any starting point would be helpful as I am completely unfamiliar with these procedures. How do you even do that since I thought not only the power limit but the multiplier is fused at the silicon?
 

DaveLTX

Active Member
Dec 5, 2021
150
34
28
Yes, I am using the Asus Sabertooth X99 (with latest official BIOS). E5-2687W v4 is exactly what I've been looking at for the last week as finding any E5-16xx north of E5-1650 v3 is very difficult or too expensive. Right now, I am seeing E5-2687W v4 going for 120 USD upwards on the fleaBay - is that a reasonable price or should I look for lower prices?

If you could provide some links on how to hack BIOS into overclocking such 2P CPUs with locked multiplier that would be great - any starting point would be helpful as I am completely unfamiliar with these procedures. How do you even do that since I thought not only the power limit but the multiplier is fused at the silicon?
X99 Sabertooth? Oh my lord... I have bad stories about that one. Asus sent me one back in the day and it was DOA, turned out to be a bad bios chip but putting a copied file (using a external flasher) worked for all of few days before it quit working again

120USD is reasonable. Grab one.

Unfortunately I don't. I only know it from watching videos of them doing 2P CPUs with locked multis on the chinese motherboards.

The ratios are fused at silicon level but it does not prevent you from cranking the max single core clockspeed on all cores, so in other words 4GHz all core given power limits allow as the UEFI usually respects the all core boost fused into silicon so basically like "overclocking" a 12700 in any board except for H610 because they often have these days an option to ignore power limits and go to max all core clocks (which they still have to, intel happens)
In this case, its ignoring max all core clocks and try to ignore power limits but that cannot be overridden unlike Alder Lake S
 

Pakna

Member
May 7, 2019
50
3
8
X99 Sabertooth? Oh my lord... I have bad stories about that one. Asus sent me one back in the day and it was DOA, turned out to be a bad bios chip but putting a copied file (using a external flasher) worked for all of few days before it quit working again
I have had zero problems in seven years I've had it - running a 1.26 V i7-5820k overclocked daily, without a problem. It touted "TUF" stability and had beefy heatsinks which I thought will help in the long run and also with the system being as quiet as possible. So far so good.

1672089271689.png1672089293484.png

120USD is reasonable. Grab one.
Sounds good - in case this unit gets sold before I pull the trigger, what price range is reasonable? Ideally, I'd like some benchmarks at least as a sanity check....I'd hate to have a sluggish system once I do the swap. Do you happen to have such a CPU lying around? Or could you point me to a trustworthy benchmark elsewhere?

Unfortunately I don't. I only know it from watching videos of them doing 2P CPUs with locked multis on the chinese motherboards.

The ratios are fused at silicon level but it does not prevent you from cranking the max single core clockspeed on all cores, so in other words 4GHz all core given power limits allow as the UEFI usually respects the all core boost fused into silicon so basically like "overclocking" a 12700 in any board except for H610 because they often have these days an option to ignore power limits and go to max all core clocks (which they still have to, intel happens)
In this case, its ignoring max all core clocks and try to ignore power limits but that cannot be overridden unlike Alder Lake S
So, even though the CPU's multiplier is locked, you can still change the frequency multiplier say, from 30 to 40, and get a 4 GHz chip with a BCLK of 100 MHz? If that's how it works, then what is actually multiplier locking? I thought it would prevent the user changing this factor in the first place?
 

DaveLTX

Active Member
Dec 5, 2021
150
34
28
I have had zero problems in seven years I've had it - running a 1.26 V i7-5820k overclocked daily, without a problem. It touted "TUF" stability and had beefy heatsinks which I thought will help in the long run and also with the system being as quiet as possible. So far so good.


Sounds good - in case this unit gets sold before I pull the trigger, what price range is reasonable? Ideally, I'd like some benchmarks at least as a sanity check....I'd hate to have a sluggish system once I do the swap. Do you happen to have such a CPU lying around? Or could you point me to a trustworthy benchmark elsewhere?



So, even though the CPU's multiplier is locked, you can still change the frequency multiplier say, from 30 to 40, and get a 4 GHz chip with a BCLK of 100 MHz? If that's how it works, then what is actually multiplier locking? I thought it would prevent the user changing this factor in the first place?
Mine was broken on day one. I also had a X99-A that did a error "LED dance" sometimes

I long sold my X99 platforms entirely to move to Ryzen back in 2017

Multiplier locking is locking the max multipliers to 40. But stock UEFI respecting the all core limits set out by Intel.
 

Pakna

Member
May 7, 2019
50
3
8
Mine was broken on day one. I also had a X99-A that did a error "LED dance" sometimes

I long sold my X99 platforms entirely to move to Ryzen back in 2017
That is unfortunate. What are you running now and what are you considering for upgrades?
Any thoughts on upcoming AMD Storm Peak?

Multiplier locking is locking the max multipliers to 40. But stock UEFI respecting the all core limits set out by Intel.
Do I understand this correctly that multiplier locking actuallly _doesn't_ lock the multipliers completely - it just locks step increases over 40?