Blender Eyes Raising Its CPU Requirements

[piotrj3]

No.


No, no, no! Some M.2 slots can operate in AHCI mode (i.e. those connected to the chipset, rather than the CPU), if you put a SATA M.2 drive in it. That's a totally different issue, and irrelevant to the discussion because we're talking about boards that don't even have M.2 slots!


No, that's not how it works -- there's no "some action". The SSD does a PCIe DMA transfer of its result, and that's automatically hardware-routed to the CPU hosting the memory address targeted by the DMA.


Sandybridge upgraded QPI to 8.0 GHz and each link is 64 (data) bits per direction. The E7 CPUs supporting 4P configurations have 3 links, enabling a point-to-point topology. So, between each pair of CPUs, there's a dedicated 64 GB/s link in each direction. That's plenty, when PCIe 3.0 x4 maxes out at about 4 GB/s.


Explain what you mean by "Custom SSD working as PCI-E". You can't because that's not a real thing! That's some arbitrary distinction you imagined, just like that crap about audio dropouts and network lags.


Some workloads scale well to 2P. From my own experience, I can say that builds are among them.

Another use case that can work well on NUMA systems is VMs that fit on a single CPU.


You have no idea why it didn't scale better, because you don't know how bottlenecked it was on storage. Michael's storage configurations are pretty basic, which is why he doesn't run any database benchmarks. He doesn't have enough fast, write-optimized PCIe 4.0 SSDs.

However, if you look at older 2P benchmarks, which are more applicable to a 32-core system, we see that builds indeed scale pretty well to a system of that size.


And why isn't that good enough? If you're waiting for a render to complete, having it run 71.5 to 89% faster is sure going to sound like a good idea!


Depends on what you're doing.


The first thing I said about it, in this entire thread, was to point out the noise and power issues. Those are the main reasons why I don't have an old 2U rackmount server.

However, what's awesome about servers is stuff like:

• Full ECC memory, with advanced features like patrol scrub
• Large memory support
• Battery-backed hardware RAID controller
• Hotswap backplane for HDDs and SSDs
• Fast networking
• BMC for remote management

If you don't care about any of those things, then yeah, maybe don't get a full-fledged rackmount server.


LOL. If you're using such a slow IDE that how it's threaded is an issue, you have only yourself to blame.


I expect people don't buy an old server without having some idea of what they're getting into (or some willingness to find out). The internet is full of information and advice, if they go looking.

On the flip side, you shouldn't say reckless things about NVMe that you don't actually know to be true. It's perfectly fine for people to use them in old machines, as long as they understand you can't boot from it. You can even still mount your root filesystem and install the OS there, as I've done.

FIRST SSDs by default work in AHCI mode. all M2 non nvm-e want that, and all SATA cable connected want that. but AHCI is still inferior to NVMe.

AHCI is not NVME. NVME is diffrent protocol to AHCI, and requires hardware support. You won't be able to use NVME drive in NVME way if you don't have required hardware for it. Haswell refresh is minimum required hardware to suppport NVMe. and NVMe specifies a communication SSD <-> CPU (or chipset) and it has special hardware inside to accelerate operating on such protocols. And this is why you will have performance penalty because even if you get top of the line SSD it won't be able to communicate in NVME mode, and that will hurt stuff like IOPS a lot.

it is not imaginary thing dropouts, I seen LinusTechTips as well Level1 techs talking about it in few cases when it was limited.. When in normal way of working such platform rarerly hits such limits like QPI links, they do exist and when designing a workload that will scale well to 4P you can do that. Why do you think super computers go for Mellanox etc. and Intel all the time was upgrading QPI and UPI with way bigger speeds then QPI links - that is right because it is needed for them. Now for most workloads it won't happen, but you could imagine easly a case when CPU 2 is talking to GPU connected to CPU1, at the same time CPU1 wants something from RAM of CPU2, and viola you pretty much put maximum limit on QPI link cpu1 <-> cpu2. This is why server or multinode software puts such a heavy task of making sure algorithms are NUMA aware, and this is why (again not suprising) best scalability you achieve on openssl or NAS performance benchmark or NAMD - those were literally designed to. But typical render engine sometimes needs to load something from RAM of another CPU and that slows it down, that is why you achieve still good scallability but not close to 100% like 3 ones i mentioned.



If you claim such IDEs are bad, then sorry that industry standard InteliJ, Visual studio, all will have a minor hiccup there and there. And when you debug something you mostly will debug one part of it and again single threaded performance wins.

[coder]

FIRST SSDs by default work in AHCI mode.

That's simply not true, or else you could boot from them!

AHCI is not NVME.

Correct.

NVME is diffrent protocol to AHCI, and requires hardware support.

Nope. The only special support needed for NVMe is in BIOS/UEFI, if you want to be able to boot from it. Otherwise, it's just a PCIe card -- not unlike a GPU. The device driver inside the OS is what talks to it -- not any special hardware in your system or CPU.

Honestly, I can understand where you got some of this, due to the fact that there are some M.2 SATA drives, but most of it you're just pulling out of your ass. Please stop and read up on the subject, so you don't just spread more misinformation.

You won't be able to use NVME drive in NVME way if you don't have required hardware for it.

That's Bull FUCKING Shit.

I'm speaking from actual experience of having used NVMe drives in these old-ass systems and they have the proper device mapping, they're recognized as NVMe drives, and they perform like NVMe drives.

When you have a M.2 SATA drive, it only gives you 600 Gbps of performance, in spite of the fact that PCIe is faster than that. And, it shows up in the OS as a SATA drive.

Haswell refresh is minimum required hardware to suppport NVMe. and NVMe specifies a communication SSD <-> CPU (or chipset) and it has special hardware inside to accelerate operating on such protocols.

Lies.

If you really believe that, then tell us exactly what this hardware is doing and why it's required.

And this is why you will have performance penalty because even if you get top of the line SSD it won't be able to communicate in NVME mode, and that will hurt stuff like IOPS a lot.

How much is the performance penalty, specifically what is disabled, and what independent evidence do you have to support these claims?

I seen LinusTechTips as well Level1 techs talking about it in few cases when it was limited..

Well, you need to go back and reread that shit, because they weren't talking about what you thought they were.

If you claim such IDEs are bad, then sorry that industry standard InteliJ, Visual studio, all will have a minor hiccup there and there.

I don't use any of that shit. If you do, that's your problem.

And when you debug something you mostly will debug one part of it and again single threaded performance wins.

I've been debugging multithreaded code for long enough to know that it's not only the thread your debugging that's running. And when you're stepping through code, that thread is hardly doing any work -- it's limited by interaction with the user.

[piotrj3]

That's simply not true, or else you could boot from them!


Correct.


Nope. The only special support needed for NVMe is in BIOS/UEFI, if you want to be able to boot from it. Otherwise, it's just a PCIe card -- not unlike a GPU. The device driver inside the OS is what talks to it -- not any special hardware in your system or CPU.

Honestly, I can understand where you got some of this, due to the fact that there are some M.2 SATA drives, but most of it you're just pulling out of your ass. Please stop and read up on the subject, so you don't just spread more misinformation.


That's Bull FUCKING Shit.

I'm speaking from actual experience of having used NVMe drives in these old-ass systems and they have the proper device mapping, they're recognized as NVMe drives, and they perform like NVMe drives.

When you have a M.2 SATA drive, it only gives you 600 Gbps of performance, in spite of the fact that PCIe is faster than that. And, it shows up in the OS as a SATA drive.


Lies.

If you really believe that, then tell us exactly what this hardware is doing and why it's required.


How much is the performance penalty, specifically what is disabled, and what independent evidence do you have to support these claims?


Well, you need to go back and reread that shit, because they weren't talking about what you thought they were.


I don't use any of that shit. If you do, that's your problem.


I've been debugging multithreaded code for long enough to know that it's not only the thread your debugging that's running. And when you're stepping through code, that thread is hardly doing any work -- it's limited by interaction with the user.

https://www.itechguides.com/pcie-vs-nvme/

Therefore, if you come across an NVMe SSD, it is a PCIe SSD. However, a PCIe SSD might not always be an NVMe SSD.

https://www.crucial.com/articles/abo...h-pcie-or-sata

However, you can also get PCIe compatible SSDs that are non-NVMe.

Now i don't remember the name but before 2013 there were quite a few mostly enterprise SSDs that worked in PCI-E slot but predated NVME protocol and had nothing to do with it. They had custom drivers. Yes you might have NVMe storage in hardware that didn't have NVMe hardware support (or probably you weren't aware of it). But they weren't working using NVMe protocol.

Now in details....

There is one thing semi incorrect i said that NVMe won't work in older. It can work, but...
a) it requires early load of NVMe driver in software, you can even boot with it like with clover efi (that will load nvme driver, consequence of that is longer booting time because it needs now to load in software. When you have compatible hardware from start, it will boot with it straight away which kind of suggests there is some hardware support for offloading it, but it can be supported in software as well,
b) it still requires NVMe controller on SSD itself, but not on motherboard side. (this seems logical as driver speaks to device that speaks in NVMe protocol)
c) many Samsung NVMe SSDs has for some reason compability mode with older mobos and somehow can boot even in x58 chipset using pci-e adapter, but other nvme drives won't be detected by very same motherboard with very same. Look at this : https://h30434.www3.hp.com/t5/Deskto...t/true#M532720

Very curious case, Samsung has drives NVMe that can work in AHCI mode but still over PCI-E interface. Also SSD itself needs to present legacy oproms that mobo can recognize. And in that case i am 99% sure you will have performance penalty.

Also in post below

Benchmarks are a little slower than expected bur overall everything reacts very fast.

And that machine does support PCI-E gen 3. And that guy is using NVMe version of drive. Which indicates there is some software overhead most likely.


Also on last page with newest driver they achieved 950MB/s write. However in benchmarks done in 2015 (before that new driver by samsung was released) the same drive is capable of 1400MB/s write and with older driver one user had 1200MB/s reads (and samsung new driver went 2050MB/s) but benchmarks from start shown 2100MB/s or even higher for read.

[Danielsan]

Studios (marketing agencies, architectural and engineers firms, 3D studios, etc...) are the least ones that would upgrade their hardware... If were for them you would still work on 486 and WooDoo 3DFX...

I have plenty of experience working for small firms to mega corp, and the age of the HW I worked on it was, on average, ten years old...

[Svyatko]

'It's not BS. NVMe defines specific functionality which that system doesn't have. Yes, you can physically plug in a PCIe SSD (or an M.2 nvme drive with an adapter), but they're not going to operate as an NVMe drive. They're going to act like legacy add-on storage cards that require an OS-specific driver to operate once booted."

Is my full statement. Performance is not mentioned nor implied in any possible way. I also never once implied it wouldn't work at all, hence the "they're going to act like legacy add-on storage cards that require an OS-specific driver to operate once booted."

While this is true for some old SSDs, now ordinary NVME drives just work with simple adapter and PCIe 1.x+. Yes, you may encounter incompatibilities with some configuration.
No need for UEFI, it works with old BIOS. NVME support is in Linux kernel, so...

Options:
1. Put "/boot" at BIOS-visible drive, for instance - USB flash. Other stuff - on NVME drive. I've done it just for fun.
2. Add NVME drivers into BIOS/UEFI.
3. Load NVME drivers during UEFI boot.
4. Use OS loader which include these NVME drivers.
5. Use NVME with "option ROM" - some old Samsung PRO & old Plextor.

[coder]

https://www.itechguides.com/pcie-vs-nvme/

"Therefore, if you come across an NVMe SSD, it is a PCIe SSD. However, a PCIe SSD might not always be an NVMe SSD."

This explainer is for noobs. It doesn't contradict anything I said.

The quote is not inaccurate, but if we're talking about PCIe add-in-cards, then the only non-NVMe drives you'd find would be those from 2014 and earlier, which pre-date the NVMe standard.

As for M.2 drives, there's also the potential for M.2 SATA drives. Again, I mentioned that and it has no bearing on the discussion because we're talking about machines that don't even have M.2 slots.

https://www.crucial.com/articles/abo...h-pcie-or-sata

"However, you can also get PCIe compatible SSDs that are non-NVMe."

Same. This article is mostly focused on M.2 SATA drives. Something people should be aware of, but irrelevant to the discussion.

Also, the quote refers to legacy, pre-NVMe PCIe SSDs as above.

There is one thing semi incorrect i said that NVMe won't work in older. It can work, but...
b) it still requires NVMe controller on SSD itself, but not on motherboard side.

That's the very definition of a NVMe drive! Motherboards never have a NVMe controller (unless they have some sort of embedded SSD). The NVMe interface is always part of the SSD, itself.

[ssokolow]

Everything I've read says that, more or less...

• NVMe M.2 drives are just PCI-Express cards with a funny form factor (Which is amusing because I remember hearing somewhere that the earliest forms of IDE/PATA started out as "let's extend the ISA bus down a ribbon cable and bolt the important parts of the expansion card onto the drive so we can guarantee no tech support complaints or warranty returns due to controller-drive mismatches"... like a less drastic alternative to the old "hard card" concept.)
• If a motherboard is too old to boot off an NVMe drive, it's because the manufacturer of the controller chip decided BIOS was too legacy to include an option ROM for it... akin to when the Win9x/NT era had progressed enough for manufacturers to stop providing MS-DOS drivers. (But that, while they can't be used as boot drives, OSes like Linux or FreeBSD or Windows which include standard NVMHCI drivers akin to what they do with OHCI/UHCI/EHCI/xHCI USB support will still be able to access them... so similar to upgrading a pre-USB machine with a USB PCI card. The OS can use your USB keyboard/mouse, but the BIOS can't.)

...meaning that any previous knowledge about those conditions for traditional drives and controller cards in the BIOS era should still apply.

[piotrj3]

This explainer is for noobs. It doesn't contradict anything I said.

The quote is not inaccurate, but if we're talking about PCIe add-in-cards, then the only non-NVMe drives you'd find would be those from 2014 and earlier, which pre-date the NVMe standard.

As for M.2 drives, there's also the potential for M.2 SATA drives. Again, I mentioned that and it has no bearing on the discussion because we're talking about machines that don't even have M.2 slots.


Same. This article is mostly focused on M.2 SATA drives. Something people should be aware of, but irrelevant to the discussion.

Also, the quote refers to legacy, pre-NVMe PCIe SSDs as above.


That's the very definition of a NVMe drive! Motherboards never have a NVMe controller (unless they have some sort of embedded SSD). The NVMe interface is always part of the SSD, itself.

M2 sata is not PCIE. There are PCIE SSDs that weren't compatible with NVMe. That was my entire point. You argue that i think about M2 Sata but i never a single time said about SATA or M2. I talk explicitly PCIE SSDs that never fit into M2 slot.

[coder]

There are PCIE SSDs that weren't compatible with NVMe. That was my entire point.

No, it wasn't. You made lots of false and misleading claims which can't be explained by that. Crap like "pre-Haswell Refresh CPUs can't use NVMe, because the CPU has some magic engine that lets it talk to NVMe drives".

You argue that i think about M2 Sata but i never a single time said about SATA or M2.

I was reaching to find an explanation for your confusion, and it was a mere suggestion rather than a definitive assertion.

[Eirikr1848]

Except if the "shit" fully satisfies their needs and they are tight on a budget then why should they?

And lets not forget, Blender has GPU acceleration. You don't need the strongest CPU out there and anyone can slot a 3060 or a 6700XT or something with an Ivy Bridge CPU and get perfecrly acceptable performance

And if budget is a concern, an Arc A380 or Radeon 6400 could be slotted into a machine. I have an RX 550 in an Athlon XII machine.

I paid for Crowdrender but haven’t yet figured out how to set up distributed rendering to my legacy hardware that acts as space heaters in the winter