New Zen microarchitecture details

[Dresdenboy]

In a series of recent Linux kernel driver patches AMD provided further details of the family 17h processors. All these findings let the 40% claim look more likely. What they don't tell you are achievable clock frequencies and power demands. So discussing absolute performance is moot at this time still.

Some highlights (or "lowlights" for some - as you wish ):

• GMI link confirmed (first appeared on Fudzilla slide)
• FMAC units are no longer tied to the FP3 pipe as said in the October patch - this could mean no bottleneck here - and maybe some other trick (the FP3 wasn't there by accident)
• uOp cache
• 4x parallel page table walker (Skylake has 2 per core)
• Checkpointing and atomic ops mentioned (could mean finally some HW transactional memory support after proposing it in 2009)
• 4 cycle latency 2R1W 32kB L1 D$

I also updated the microarchitecture diagram:

In depth analysis on my blog: http://dresdenboy.blogspot.com/2016/02/new-amd-zen-core-details-emerged.html

Edit: Fixed a typo.

 

[grimpr]

Very interesting, so, GMI is a new fabric interconnect or an upgraded version of hypertransport?

 

[Dresdenboy]

Very interesting, so, GMI is a new fabric interconnect or an upgraded version of hypertransport?

It's new, but I didn't see anything mentioned regarding its heritage. It's very likely, that it is based on HyperTransport (and PCIe) somehow.

 

[turtile]

It's new, but I didn't see anything mentioned regarding its heritage. It's very likely, that it is based on HyperTransport (and PCIe) somehow.

I would assume that they used something from Seamicro? Or did they just waste almost a half billion dollars on it?

 

[swilli89]

In a series of recent Linux kernel driver patches AMD provided further details of the family 17h processors. All these findings let the 40% claim look more likely. What they don't tell you are achievable clock frequencies and power demands. So discussing absolute performance is moot at this time still.

Some highlights (or "lowlights" for some - as you wish ):

• GMI link confirmed (first appeared on Fudzilla slide)
• FMAC units are on longer tied to the FP3 pipe as said in the October patch - this could mean no bottleneck here - and maybe some other trick (the FP3 wasn't there by accident)
• uOp cache
• 4x parallel page table walker (Skylake has 2 per core)
• Checkpointing and atomic ops mentioned (could mean finally some HW transactional memory support after proposing it in 2009)
• 4 cycle latency 2R1W 32kB L1 D$

I also updated the microarchitecture diagram:

In depth analysis on my blog: http://dresdenboy.blogspot.com/2016/02/new-amd-zen-core-details-emerged.html

Wow good stuff man! First CPU I've been truly excited about since Sandy Bridge!

 

[Atreidin]

The L1 latency could be a hint that they are targeting high clocks. That's nothing solid, of course.

 

[cytg111]

Nice job. uOp cache is new to AMD archs is it not?

 

[swilli89]

The L1 latency could be a hint that they are targeting high clocks. That's nothing solid, of course.

This is a good point.. If they came with a 4 core Zen @ 4ghz base and the claimed 40% IPC improvement it would be a game changer. For gamers, we have long been at the point of saturation, since Sandy Bridge. You would NOT feel a difference sitting at two beige boxes, one running a 2500K and one running a 6600K in pretty much any game.

However knowing AMD, hopefully they bring this performance to a 6-core @ $200 and shake up the market.

 

[myocardia]

I would assume that they used something from Seamicro? Or did they just waste almost a half billion dollars on it?

I had also hoped that. Do we have any reason(s) so far to think that they did not use Seamico's technology, or at least base their new fabric on Seamicro's tech?

 

[Dresdenboy]

I would assume that they used something from Seamicro? Or did they just waste almost a half billion dollars on it?

Right, of course, there will be SeaMicro IP in there. So this could be a mix with the best of all these worlds. One SeaMicro influence can be seen in some patents, where AMD describes a slot based server board with a torus like data fabric. And for many cores on a chip they're also going to use a torus topology.

 

[Dresdenboy]

The L1 latency could be a hint that they are targeting high clocks. That's nothing solid, of course.

I had the same thoughts. There are many more hints that they won't aim for some med/low clocks. FinFET also helps a lot here.

Nice job. uOp cache is new to AMD archs is it not?

Yep, it is.

 

[Arachnotronic]

This is a good point.. If they came with a 4 core Zen @ 4ghz base and the claimed 40% IPC improvement it would be a game changer. For gamers, we have long been at the point of saturation, since Sandy Bridge. You would NOT feel a difference sitting at two beige boxes, one running a 2500K and one running a 6600K in pretty much any game.

However knowing AMD, hopefully they bring this performance to a 6-core @ $200 and shake up the market.

Lol, this reminds me of what the AMD fans were saying on the eve of the Derpdozer launch.

 

[cytg111]

Lol, this reminds me of what the AMD fans were saying on the eve of the Derpdozer launch.

Value added : Zero.
Much like this post. Ill go report us both now. Later.

 

[NTMBK]

Thanks for this, Dresdenboy! Looking forward to seeing some Zen benchmarks at the end of the year.

 

[NTMBK]

Value added : Zero.
Much like this post. Ill go report us both now. Later.

Adding a healthy pinch of salt is adding value in my eyes Let's not get too hyped before we actually see some benches.

 

[Arachnotronic]

Value added : Zero.
Much like this post. Ill go report us both now. Later.

It is a fact -- before BD launched, these forums were buzzing with claims that AMD was going to kick SNB into the stone age with 8 cores, more IPC than Phenom II, etc. Your join date is 2008, so I am sure you saw all of the internet buzz around Bulldozer, too.

That hype led to a lot of disappointment/angst, particularly among those who held off on buying SNB because they were told to "wait for Bulldozer."

Zen might be different, but to borrow the words of swilli89, "knowing AMD" it'll under-perform comparable Intel parts. Past performance is the best indicator of future performance and AMD has a lot to prove in CPUs.

 

[Sweepr]

Value added : Zero.
Much like this post. Ill go report us both now. Later.

There's nothing wrong with his post, stop being overly sensitive about any criticism towards AMD. Whether you like it or not, claiming they will sell a 6C/12T CPU with competitive IPC/clocks at $200 sounds way too optimistic.

 

[cytg111]

Adding a healthy pinch of salt is adding value in my eyes Let's not get too hyped before we actually see some benches.

A pinch, sure, but truckload? At every turn? No thanks.
This common behavioral group dynamics of ganging up on one individual or minority exposes, imo, some pretty substandard characteristic properties of some of our fellow members. I believe this board have had enough of it so I report away.. If enough besides me do it, maybe things will change.

(see, crowding up as I type)

 

[Arachnotronic]

A pinch, sure, but truckload? At every turn? No thanks.
This common behavioral group dynamics of ganging up on one individual or minority exposes, imo, some pretty substandard characteristic properties of some of our fellow members. I believe this board have had enough of it so I report away.. If enough besides me do it, maybe things will change.

In other words, even when a poster does not violate the ToS, you report them anyway because you don't like their opinions? Okay.

 

[Arachnotronic]

There's nothing wrong with his post, stop being overly sensitive about any criticism towards AMD. Whether you like it or not, claiming they will sell a 6C/12T CPU with competitive IPC/clocks at $200 sounds way too optimistic.

If AMD has a 6C/12T winner like that, they should not charge just $200 for it. They could easily sell it for $400+ and there would be a lot of people who would gladly pay.

Anyway, it's interesting to see the hype around Zen. It is very easy to make an unreleased chip seem like anything you want it to be, especially when you are comparing it to chips that are in the here and now.

 

[frozentundra123456]

This is a good point.. If they came with a 4 core Zen @ 4ghz base and the claimed 40% IPC improvement it would be a game changer. For gamers, we have long been at the point of saturation, since Sandy Bridge. You would NOT feel a difference sitting at two beige boxes, one running a 2500K and one running a 6600K in pretty much any game.

However knowing AMD, hopefully they bring this performance to a 6-core @ $200 and shake up the market.

How would a quad with a even with a 40% IPC increase at 4ghz be a game changer? We already have that in Skylake, probably even Haswell. If you cant tell any difference between SB and skylake, you could not tell any difference with this mythical quad with a 40% IPC increase either. And if you believe AMD will bring a competitive hex core at 200.00, well, I have some land you might wish to look at. Even AMD's marketing is not that inept.

 

[Arachnotronic]

How would a quad with a even with a 40% IPC increase at 4ghz be a game changer? We already have that in Skylake, probably even Haswell. If you cant tell any difference between SB and skylake, you could not tell any difference with this mythical quad with a 40% IPC increase either. And if you believe AMD will bring a competitive hex core at 200.00, well, I have some land you might wish to look at.

40% IPC increase from XV is about Haswell level. This kind of performance has been "buyable" since 2014 with the launch of HSW-E 6C/12T for under $400 as long as you're willing to overclock.

I love how when Intel provides a certain level of performance/functionality, they are accused of screwing customers, but if AMD is able to provide it, it's a "game changer" and "most exciting architecture in the history of ever."

 

[MrTeal]

40% IPC increase from XV is about Haswell level. This kind of performance has been "buyable" since 2014 with the launch of HSW-E 6C/12T for under $400 as long as you're willing to overclock.

I love how when Intel provides a certain level of performance/functionality, they are accused of screwing customers, but if AMD is able to provide it, it's a "game changer" and "most exciting architecture in the history of ever."

It's not about whether or not you could have bought HSW-E performance for the last 18 months anyway. It's that AMD is going to be selling 8 core Haswell-E levels of performance in the same $100-$200 range as the current octo-core FX processors. I don't recall the exact quote, but I think Lisa Su herself said something along the lines of "AMD cherishes its history as a value leader, and seeks to continue being a budget brand going forward." It was something like that, at least.

 

[ShintaiDK]

I thought she said she was done being the discount brand.

 

[cytg111]

In other words, even when a poster does not violate the ToS, you report them anyway because you don't like their opinions? Okay.

For disturbing the peace. Yea sure.
What, you see it as your sacred duty to keep it real so you sprinkle some anti-amd-fan rhetoric around when a poster steps over the line (guesstimating Zen performance and pricepoint)? Thank god we have you to smack the "keeping-it-real" hammer down.

insulting other members is not allowed
Markfw900