The mozilla/rr debugger doesn't work on Ryzen CPUs - https://github.com/mozilla/rr/issues/2034

(rr stands for Record and Replay Framework, it's a gdb extension which allows for reverse execution)

Interesting that GDB had its own "time travel" support for longer that afaik works on AMD. Will have to see what rr brings to the table.

GDB's recording is extremely limited. It doesn't even work on multi threaded programs.

It's also about 1000x slowdown during recording, where rr is < 2x.

in ZEN2:

"The key highlight improvement for floating point performance is full AVX2 support. AMD has increased the execution unit width from 128-bit to 256-bit, allowing for single-cycle AVX2 calculations, rather than cracking the calculation into two instructions and two cycles. This is enhanced by giving 256-bit loads and stores, so the FMA units can be continuously fed. AMD states that due to its energy aware scheduling, there is no predefined frequency drop when using AVX2 instructions (however frequency may be reduced dependent on temperature and voltage requirements, but that’s automatic regardless of instructions used)"

"In the floating point unit, the queues accept up to four micro-ops per cycle from the dispatch unit which feed into a 160-entry physical register file. This moves into four execution units, which can be fed with 256b data in the load and store mechanism.

Other tweaks have been made to the FMA units than beyond doubling the size – AMD states that they have increased raw performance in memory allocations, for repetitive physics calculations, and certain audio processing techniques.

Another key update is decreasing the FP multiplication latency from 4 cycles to 3 cycles. That is quite a significant improvement. AMD has stated that it is keeping a lot of the detail under wraps, as it wants to present it at Hot Chips is August. We’ll be running a full instruction analysis for our reviews on July 7th."

https://www.anandtech.com/print/14525/amd-zen-2-microarchite...

> Any significant differences in regards to scientific applications or machine learning?

The PCIe4.0 is good for Data Science ( NVMe Disk / GPU machine learning )

"Corsair MP600 SSDs will offer up to 4950 MB/s sequential read speed as well as up to 4250 MB/s sequential write speed when used with a PCIe 4.0 x4 interface, which is substantially faster when compared to modern PCIe 3.0 x4 drives."

https://www.anandtech.com/show/14584/corsair-mp600-pcie-4-ss...

Darn, I just bought a Samsung 970 Pro 1TB NVMe drive for my new Ryzen zen2 build. I probably should have waited to buy this PCI4.0 drive instead. Thanks for sharing

> Corsair MP600 SSDs will offer up to 4950 MB/s sequential read speed as well as up to 4250 MB/s sequential write speed when used with a PCIe 4.0 x4 interface

WOW! Looks like they haven't announced pricing yet. I wonder what those drives will cost.

$250 for the 1 TB version and $450 for the 2 TB model, Corsai accidentally put the listings up early on Amazon a few days ago and then took them down

Yes official and guaranteed support for ECC but it was an i3-8100 instead of an AMD equivalent. For a FreeNAS build. I will upgrade it to a Xeon E-21XX or whatever supports my Supermicro C246 motherboard in the future. AMD doesn’t test ECC on Ryzen but doesn’t turn it off. Server grade components are not as common either.

Threadrippers are tempting but I like to play with EXSI and Intel’s virtualization is handy. My next desktop might be a Threadripper but my i7-7700K and Intel QSV is good enough for Plex and 4K.

I would like have a 10gb backbone with SFP+ so I can play with Fiber, DAC, and 10gb Ethernet.

What precludes you from using ESXi and virtualization on AMD's platform? The virtualization extensions to x86 are common to both AMD and Intel. Even direct IO and IOMMU. Is there something missing?

The virtualization extensions are actually different, but yeah, VMware supports both across all platforms. I think it's only AMD Hackintosh where you might have an issue?

I use the TSC for high precision timing, and turning TSC ticks into nanoseconds requires interpreting model specific registers, so my timing code would need AMD specific patches to work properly on those processors. Not necessarily a deal-breaker, but it would definitely be an annoyance for me if I did decide to switch. Probably not super relevant to scientific computing or ML though.

How do you use the TSC for timing when the frequency is changing unpredictably?

Recent processors have invariant TSC.

There might be meaningful differences in the hardware encode/decode for video. In the Plex community for example it seems like Intel CPUs are the best bet. They support a lot of codecs and can service multiple transcode streams. Consumer Nvidia GPUs are limited to 2. I don't think Plex supports hardware transcoding on AMD Ryzen or discrete GPUs at all right now. I haven't really looked into why yet, and could likely just be a software support issue in Plex.

This, unfortunately, would be a deal breaker for my home-server use case, although I'm very happy AMD is finally competitive again, overall.

The Zen2 chips being released and discussed here don't have a GPU. The Ryzen 3000 series APUs being released soon are using last year's Zen+ CPU architecture; and are only 4 cores, so not super comparable to a $2000 Intel CPU :)

The rr debugger had problems with Ryzen. AMD doesn't have AVX512 but that's more of a performance difference; code shouldn't be written to require it.

+1. I've heard some optimizations in scientific computing (e.g. Gaussian 16, from someone who does chemistry) is not available without AVX, but I'm not a practitioner of scientific computing, and know nothing about the current situation. Could anyone share your experience? Also, what is the performance figure of multimedia/encoding performance?

I have too frequently had problems with hardware support in OS outside Linux and Windows for non Intel chipsets.

One that sticks out is virtualization support in freebsd and others. I don't know if it's still true but the time risk isn't worth the cost benefit for me any more.

I'm considering Intel for my next workstation because I want to be able to run macOS at least in VirtualBox (or just run Hackintosh, did not decide yet). For some reason macOS does not like AMD CPUs.

There was a recent video on Linus Tech Tips where they set up a Threadripper workstation running macOS via KVM and it out-performed a top of the line iMac Pro.

Does it work with VirtualBox? I want to run Windows.

You can run Windows in KVM. I haven't used VirtualBox for years so my knowledge is out of date.

> For some reason macOS does not like AMD CPUs.

Reddit:"Mojave, now on Ryzen!" https://www.reddit.com/r/hackintosh/comments/a95xlo/mojave_n...

Installing custom built kernel on a proprietary OS sounds like a recipe for disaster. It's likely to break on the first update. More like a proof of concept rather than stable solution.

With a little elbow grease this hasn't been true for Hackintosh builds nor virtualization for ages: https://amd-osx.com

You can run MacOS through KVM easy on Ryzen. Even GPU passthrough works fine.

iMessage and other features generally break when running amd macintoshes, intel ones are indistinguishable.

Why would they break if macOS itself runs fine on AMD? These softwares run on macOS and don't make any system calls directly by bypassing the OS, right?

They typically check for a network interface from a real Mac. Using a Hackintosh you can't spoof some of the NIC details that are detected and cause iCloud/iMessage to not work.

Oh, wasn't aware of that. So does iCloud / iMessage work when macOS is run virtualised?

Oh man, my suggestion buy a used dual socket Xeon V2 (or V3 if you can afford it) server with room for a GPU. The price for performance is pretty unbelievable since you get a whole system, which you can run ESXi or whatever flavor of hypervisor you want.

V2/V3 Xeons are dirt cheap on ebay and if you don't mind a little risk the ES versions are mighty cheap for what you can get.

I just got Dual 2667 V2, 96GB of ram, dual 1200W PSUs, and room for 20 SAS2 HDs for less than $1,000 (replaced the 2637 V2s that came with the system). I'll note power consumption will be quite a bit higher, but power is pretty cheap where I live.

.... I'll be really stoked in a few years when the Rome chips are available used.

I have a Dell Precision T7600. The only thing I hate is the almost full minute to boot with a perc controller and other servery type checks

It won't work in a virtual machine? Any reason why?

Also interested in this