And how much does it matter? TSX is great thing no doubt there. But the adoption? What can you name of real software what uses and get significant benefit of it?
I blame Intel stupid marketing for cutting TSX from too many versions and killing the adoption.
One of the most important features of TSX are checkpoints. Zen supports checkpoints in its execution pipeline. Otherwise, I've not seen anything that said Zen did or did not support TSX, not that the tech is widely used at this time.
From there, you just need tagging and a few other features to add support. It's something that could be included in Zen+ if Zen does not have it.
It looks like Zen was developed to accelerate the vast majority of software, and rely on core count for everything else. It might explain the lack of focus on AVX.
If cache stats were any indication of performance, it would appear that Zen was destined to compete with Broadwell, but not quite match the Lake CPUs; Zen+ would perhaps close the gap albeit a bit late. Bulldozer was hamstrung by half-speed writes and horrific L3 latency - would it be remiss to assume that they've at least fixed those two issues?
I'm not sure anybody can truly predict performance however, even with a Blender demonstration, and certainly not to work out prospective Cinebench or SuperPi performance. You could have a monster of an architecture, but if the software isn't optimised for it, it's not going to be representative of its true performance.
I'd still want the TSX instructions before even thinking about the server market. I guess they surrendered that before the overall architecture was finished. Although considering how badly it has worked for Intel (essentially turned off after errata was noted in the first generation), maybe it wasn't worth risk.
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.
Facebook
Twitter
RSS
106 Comments
View All Comments
eldakka - Wednesday, August 24, 2016 - link
The first page link, AMD Server CPUs and Motherboard Analysis, is wrong, it actually links to the ARM v8-A article.atlantico - Friday, August 26, 2016 - link
Yes, it's also wrong here: http://www.anandtech.com/show/10585/unpacking-amds...Sigh.
TristanSDX - Wednesday, August 24, 2016 - link
Zen do not support transactional memory, big disadvantage comparing to IntelSenti - Wednesday, August 24, 2016 - link
And how much does it matter? TSX is great thing no doubt there. But the adoption? What can you name of real software what uses and get significant benefit of it?I blame Intel stupid marketing for cutting TSX from too many versions and killing the adoption.
coder111 - Wednesday, August 24, 2016 - link
As far as I know, Azul JVMs do support transactional memory. So if you have a Java app, you can use it.Other than that, yes, I haven't seen TSX used much...
68k - Wednesday, August 24, 2016 - link
Isn't the version of glibc in recent Linux-distributions using the lock elision feature of TSX?https://lwn.net/Articles/534758/
https://01.org/blogs/tlcounts/2014/lock-elision-gl...
If so, then essentially every single Linux program does make use of TSX when present.
looncraz - Wednesday, August 24, 2016 - link
One of the most important features of TSX are checkpoints. Zen supports checkpoints in its execution pipeline. Otherwise, I've not seen anything that said Zen did or did not support TSX, not that the tech is widely used at this time.From there, you just need tagging and a few other features to add support. It's something that could be included in Zen+ if Zen does not have it.
silverblue - Wednesday, August 24, 2016 - link
It looks like Zen was developed to accelerate the vast majority of software, and rely on core count for everything else. It might explain the lack of focus on AVX.If cache stats were any indication of performance, it would appear that Zen was destined to compete with Broadwell, but not quite match the Lake CPUs; Zen+ would perhaps close the gap albeit a bit late. Bulldozer was hamstrung by half-speed writes and horrific L3 latency - would it be remiss to assume that they've at least fixed those two issues?
I'm not sure anybody can truly predict performance however, even with a Blender demonstration, and certainly not to work out prospective Cinebench or SuperPi performance. You could have a monster of an architecture, but if the software isn't optimised for it, it's not going to be representative of its true performance.
wumpus - Wednesday, August 24, 2016 - link
I'd still want the TSX instructions before even thinking about the server market. I guess they surrendered that before the overall architecture was finished. Although considering how badly it has worked for Intel (essentially turned off after errata was noted in the first generation), maybe it wasn't worth risk.Alexvrb - Sunday, August 28, 2016 - link
Yeah they need to take their time. A faulty implementation would do more harm than good at this point.