I recall ryzen having a particular weakness when it comes to SSD performance in low queue depths at 4k and below, sometimes 50% or more compared to intel platforms.
In the context of an all flash data server, I wonder if amd have addressed that weaknes?
intel vs ryzen write/read % percent intel advantage
960 evo 1tb pcie 512b - 78/80 vs 43/47 % 81/70 1k - 156/163 vs 86/94 % 81/73 2k - 307/324 vs 176/193 % 74/68 4k - 608/632 vs 357/391 % 70/59
750 evo 500gb sata 512b - 64/71 vs 28/31 % 128/129 1k - 126/137 vs 63/58 % 100/136 2k - 212/237 vs 123/128 % 72/85 4k - 324/365 vs 213/243 % 52/50
The intel platform also enjoys around 50% and better access times, which leads me to suspect that ryzen suffers such sigificant losses because its ssd access is not "fine grained" enough.
so buy a 16 lane nvme controller, like u have to with niggardly io intel mobos, for big such jobs.
the exciting uses for this stuff involves 32k+ blocks, which performs fine.
i notice a popular app consolidates small blocks into 32k blocks
amdS fancy new memory management system, should ensure the various resources that comprise the managed memory pool, are used appropriately to their strengths and weaknesses.
maybe hardware remedies like extra dram on transactional app ssdS?
Could be Skylakes Speedshift in action. No CPU reaches its working frequencies as quickly. Could be very relevant under low queue depths and may be "fixed" by switching to a high performance profile.
This wouldn't concern servers, though, as any right-sized storage server shouldn't constantly run at low queue depths.
Very exciting stuff. Love the Epyc coverage and can't wait for an in house review from AT. I'd like to hear more about those OCuLink connectors.
Question - aren't those two PCIe slots too physically close to actually use both? It just stands out as really odd looking compared to consumer board spacing.
Saw this on Thinkmate earlier: http://www.thinkmate.com/amd-epyc Seems like Tyan has a 1U model with support for Epyc AND 4 GPUS, though that specific model doesn't seem to be listed. I've been using Supermicro and it's been good, anyone have experience with Tyan boxes?
The ancient S2882 and S38-something or other (updated version of the S2882) motherboards for Opteron CPUs were excellent workhorses ... so long as you didn't need IPMI support. They required the use of a separate daughter-card for IPMI and we could never get the damned things to work. :( But the boards themselves ran beautifully. We still have a few running in the server room hosting KVM and Xen VMs. And a small handful of them out in the elementary schools running as Linux file/proxy servers for diskless Linux desktops.
Haven't used any newer Tyan boards. The above were from the single-core Opteron days.
We switched to SuperMicro boards shortly after that and haven't looked back. The builtin IPMI support is excellent (although the Java-based web client sucks, ipmitool + SoL works great), the boards run stable, and the board itself tends to outlast CPUs, RAM, and other hardware.
Wow, based off that close up photo of the motherboard that CPU socket looks like the size of a 2.5" laptop hard drive. Is this size of CPU the direction we are moving towards in the future?
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
36 Comments
Back to Article
K_Space - Thursday, June 22, 2017 - link
So who's getting the Epyc review units: Ryan, Ian, Ganesh? Fight fight fight :PArchie2085 - Thursday, June 22, 2017 - link
May End up with Anton ??SunnyNW - Thursday, June 22, 2017 - link
JohanRyan Smith - Thursday, June 22, 2017 - link
It's a secret. We don't want you guys getting any ideas...Alexvrb - Thursday, June 22, 2017 - link
Fine! But upload the results of the AnandTech Reviewer Battle Royale at a future date, please. Good luck, Ryan, I'll be pulling for you.Lord of the Bored - Saturday, June 24, 2017 - link
I volunteer! For the good of the community, of course.ddriver - Thursday, June 22, 2017 - link
I recall ryzen having a particular weakness when it comes to SSD performance in low queue depths at 4k and below, sometimes 50% or more compared to intel platforms.In the context of an all flash data server, I wonder if amd have addressed that weaknes?
eSyr - Thursday, June 22, 2017 - link
Well, by pushing a lot of dirrectly attached disks: http://cdn.wccftech.com/wp-content/uploads/2017/06... http://images.anandtech.com/doci/11562/14979897942... (for some reason these are not in the slide deck from the epyc presentation)Alexvrb - Thursday, June 22, 2017 - link
SATA or PCIe?ddriver - Friday, June 23, 2017 - link
The problem is quite evident in both cases:http://www.tweaktown.com/articles/8073/amd-ryzen-s...
http://www.tweaktown.com/articles/8073/amd-ryzen-s...
intel vs ryzen write/read % percent intel advantage
960 evo 1tb pcie
512b - 78/80 vs 43/47 % 81/70
1k - 156/163 vs 86/94 % 81/73
2k - 307/324 vs 176/193 % 74/68
4k - 608/632 vs 357/391 % 70/59
750 evo 500gb sata
512b - 64/71 vs 28/31 % 128/129
1k - 126/137 vs 63/58 % 100/136
2k - 212/237 vs 123/128 % 72/85
4k - 324/365 vs 213/243 % 52/50
The intel platform also enjoys around 50% and better access times, which leads me to suspect that ryzen suffers such sigificant losses because its ssd access is not "fine grained" enough.
Lolimaster - Friday, June 23, 2017 - link
All those tests are obsolete after bios, windows updates.msroadkill612 - Friday, June 23, 2017 - link
If so, nobodys perfect.so buy a 16 lane nvme controller, like u have to with niggardly io intel mobos, for big such jobs.
the exciting uses for this stuff involves 32k+ blocks, which performs fine.
i notice a popular app consolidates small blocks into 32k blocks
amdS fancy new memory management system, should ensure the various resources that comprise the managed memory pool, are used appropriately to their strengths and weaknesses.
maybe hardware remedies like extra dram on transactional app ssdS?
FriendlyUser - Sunday, June 25, 2017 - link
You are talking about Ryzen, the article is about Epyc.ddriver - Monday, June 26, 2017 - link
Its the same microarchitecture LOL...MrSpadge - Monday, June 26, 2017 - link
Could be Skylakes Speedshift in action. No CPU reaches its working frequencies as quickly. Could be very relevant under low queue depths and may be "fixed" by switching to a high performance profile.This wouldn't concern servers, though, as any right-sized storage server shouldn't constantly run at low queue depths.
lakedude - Friday, June 23, 2017 - link
ddriver, Ryzen or Epyc? Epyc seems much better equipped for I/O than Ryzen.highlnder69 - Thursday, June 22, 2017 - link
Three of those servers would make for once heck of a nice ESXi vSan cluster.Rocket321 - Thursday, June 22, 2017 - link
Very exciting stuff. Love the Epyc coverage and can't wait for an in house review from AT. I'd like to hear more about those OCuLink connectors.Question - aren't those two PCIe slots too physically close to actually use both? It just stands out as really odd looking compared to consumer board spacing.
SunnyNW - Thursday, June 22, 2017 - link
For use with riser cardsSunnyNW - Thursday, June 22, 2017 - link
SHould have said I believe they're meant to be used with risersGrayswean - Thursday, June 22, 2017 - link
So that'd be a ryzen riser.Spunjji - Friday, June 23, 2017 - link
An Epyc Ryzen riser, to be precise.DanNeely - Thursday, June 22, 2017 - link
I only see 2 x16 slots though, unless they're putting a PLX on one of the risers that's only enough to support 4 x8 cards not 5.anoother - Sunday, June 25, 2017 - link
They are x24 slotsChicagoNerd77 - Thursday, June 22, 2017 - link
Saw this on Thinkmate earlier: http://www.thinkmate.com/amd-epycSeems like Tyan has a 1U model with support for Epyc AND 4 GPUS, though that specific model doesn't seem to be listed.
I've been using Supermicro and it's been good, anyone have experience with Tyan boxes?
phoenix_rizzen - Friday, June 23, 2017 - link
The ancient S2882 and S38-something or other (updated version of the S2882) motherboards for Opteron CPUs were excellent workhorses ... so long as you didn't need IPMI support. They required the use of a separate daughter-card for IPMI and we could never get the damned things to work. :( But the boards themselves ran beautifully. We still have a few running in the server room hosting KVM and Xen VMs. And a small handful of them out in the elementary schools running as Linux file/proxy servers for diskless Linux desktops.Haven't used any newer Tyan boards. The above were from the single-core Opteron days.
We switched to SuperMicro boards shortly after that and haven't looked back. The builtin IPMI support is excellent (although the Java-based web client sucks, ipmitool + SoL works great), the boards run stable, and the board itself tends to outlast CPUs, RAM, and other hardware.
TheOriginalTyan - Wednesday, July 5, 2017 - link
Welcome anyone who wants to take a look at them again.Glock24 - Thursday, June 22, 2017 - link
Haven't heard of Tyan in a looong time, not since the Athlon MP days.I had a consumer Tyan motherboard, the Trinity KT400. I'm feeling old.
edlee - Thursday, June 22, 2017 - link
Yes, where has tyan been last ten yearsDanNeely - Friday, June 23, 2017 - link
Making server mobos and bare bone rackmount systems that are of minimal interest to consumers.TheOriginalTyan - Wednesday, July 5, 2017 - link
Ouch.NanoTec - Thursday, June 22, 2017 - link
The back in tyan_epyc_678_678x452.jpg is not the TN70A in tyan_epyc_1.jpgAnton Shilov - Thursday, June 22, 2017 - link
Thank you so much for sharp eyes. Fixed.Razneu - Tuesday, June 27, 2017 - link
Wow, based off that close up photo of the motherboard that CPU socket looks like the size of a 2.5" laptop hard drive. Is this size of CPU the direction we are moving towards in the future?TheOriginalTyan - Friday, June 30, 2017 - link
Eventually the whole board ends up in the chip. (Don't quote us on that.)petteyg359 - Thursday, June 29, 2017 - link
Where does one get a 4U or 5U rack case with 140mm fan mounts?