AMD Zen 2 Offers a 13% IPC Gain over Zen+, 16% over Zen 1

Nelly · 17 Oct 2018 at 16:58

TechPowerUP | Posted: 17 October 2018 said:
AMD "Zen" CPU architecture brought the company back to competitive relevance in the processor market. It got an incremental update in the form of "Zen+" which saw the implementation of an improved 12 nm process, and improved multi-core boosting algorithm, along with improvements to the cache subsystem. AMD is banking on Zen 2 to not only add IPC (instructions per clock) improvements; but also a new round of core-count increases. Bits n Chips has information that Zen 2 is making significant IPC gains.

According to the Italian tech publication, we could expect Zen 2 IPC gains of 13 percent over Zen+, which in turn posted 2-5% IPC gains over the original Zen. Bits n Chips notes that these IPC gains were tested in scientific tasks, and not in gaming. There is no gaming performance data at the moment. AMD is expected to debut Zen 2 with its 2nd generation EPYC enterprise processors by the end of the year, built on the 7 nm silicon fabrication process. This roughly 16 percent IPC gain versus the original Zen, coupled with higher clocks, and possibly more cores, could complete the value proposition of 2nd gen EPYC. Zen 2-based client-segment products can be expected only in 2019.

Source: Bits and Chips / TechPowerUp

humbug · 17 Oct 2018 at 17:29

Interesting... add that to this and Intel's performance lead is finished, AMD will take it.

https://forums.overclockers.co.uk/threads/why-amds-zen2-at-4-5ghz-will-win.18834329/

humbug said:
This is worth its own thread, it explains how Dr Agner Fog details a lot more IPC potential locked up in the Zen architecture, it actually has much higher IPC than Intel's Skylake architecture but a weak front-end is holding Ryzen back, improvements to this front end over time can alleviate the bottleneck put AMD's IPC way ahead of Intel.

Zen is actually running at 32 Bytes per Clock, this vs 16 Bytes per clock of Skylake.
SkyLake 4 Micro ops per clock vs 6 on Zen.
Decode: Skylake 1536 uOps vs Zen 2048 uOps
and so on.........

Rroff · 17 Oct 2018 at 17:34

13% seems really low gains for a move from "12nm" to a 7nm node.

hominid · 17 Oct 2018 at 17:40

Rroff said:
13% seems really low gains for a move from "12nm" to a 7nm node.

IPC not clock speed!

Distracted · 17 Oct 2018 at 17:43

That's only instructions per clock for 'scientific tasks'.

If the task is different it is reasonable to assume that the performance change will be too. Likewise we don't know if anything else is improved on top of ipc. The speeds of the chips could be upped, memory support and imc could change.

Of course however unlikely it seems, things could also get worse.

We don't know.

Just be glad AMD are competing and lets hope for the sake of competition that they take the top spot in every workload.

Bring on the cpu price wars! :cool:

Rroff · 17 Oct 2018 at 17:51

hominid said:
IPC not clock speed!

I know - 7nm gives a huge amount of potential for improvements over "12nm".

DragonQ · 17 Oct 2018 at 17:52

Yeah this is very early days, no need to place any credence on the 13% figure yet. They could easily achieve that in a specific workload with a new instruction set, for example.

Also, if daddy Intel hasn't been able to improve IPC beyond 5% since 2014, how can AMD do it?

Rroff said:
I know - 7nm gives a huge amount of potential for improvements over "12nm".

Node shrinks typically mean compromising over improving performance, power, or yields. If AMD literally just shrank Zen+ to 7nm they might get a large improvement in yields and a power usage improvement, but not much in terms of performance. If they make the die size the same by adding tonnes of transistors they'd bump up performance but probably not get much improvement in power or yields. Who knows which direction they're going.

Also remember that Intel node shrinks typically came with 0% IPC improvements, since that was the tick-tock methodology they used. So even 10% would be quite nice.

humbug · 17 Oct 2018 at 18:03

Rroff said:
I know - 7nm gives a huge amount of potential for improvements over "12nm".

What? think back to everytime Intel moved their architecture from one die shrink to the next, 3% IPC increments and those were architectural tweaks.

Die shrinks don't make any difference to IPC, its still the same architecture, you have to make changes to the architecture to gain IPC, that's how much performance you get at a given speed (Instructions Per Clock)

AMD are making changes to the architecture from Zen 1 to Zen 2, that's where the 13% extra performance 'at the same clock speed' comes from, you don't get that from die shrinks Rroff.

DragonQ · 17 Oct 2018 at 18:05

humbug said:
What? think back to everytime Intel moved their architecture from one die shrink to the next, 3% IPC increments and those were architectural tweaks.

Die shrinks don't make any difference to IPC, its still the same architecture, you have to make changes to the architecture to gain IPC, that's how much performance you get at a given speed (Instructions Per Clock)

AMD are making changes to the architecture from Zen 1 to Zen 2, that's where the 13% extra performance at the same clock speed comes from, you don't get that from die shrinks Rroff.

Mostly true but node shrinks give you more transistor space, which is a big help to improving the architecture in the first place.

Rroff · 17 Oct 2018 at 18:09

DragonQ said:
Mostly true but node shrinks give you more transistor space, which is a big help to improving the architecture in the first place.

Difference in nodes can help with things like latency, etc. which can all enable higher IPC if you design around it. Assuming these CPUs are going from a fairly lousy GF12nm to the improved TSMC 7nm (maybe I have that wrong?) it isn't really that great - it is a bigger jump even than Intel's 22nm to 14nm and Intel really didn't make any effort.

humbug · 17 Oct 2018 at 18:09

DragonQ said:
Mostly true but node shrinks give you more transistor space, which is a big help to improving the architecture in the first place.

yes, yet it just means you can cram more into the same space, or use less space, an architectural change is needed to gain IPC, and even then you have to know what you are doing, skill / knowledge.... Bulldozer was bigger than Sandy Bridge on the same size 32nm node, the IPC was only 60% that of Sandy.

sidthesexist · 17 Oct 2018 at 18:42

Rroff said:
13% seems really low gains for a move from "12nm" to a 7nm node.

Given that it''s the first iteration of 7nm is it really that big a surprise? Even more given how long Intel have spent on their new node and are still struggling?

Rroff · 17 Oct 2018 at 18:47

sidthesexist said:
Given that it''s the first iteration of 7nm is it really that big a surprise? Even more given how long Intel have spent on their new node and are still struggling?

If TSMC are taken at their word their 7nm capability is coming along nicely, with EUV ramping now, and seeing nothing like the issues Intel is dealing with.

Alexrose1uk · 19 Oct 2018 at 08:49

Things are only going to get worse for Intel if the current pricing situation doesnt resolve.
Things are bad right now:
https://www.youtube.com/watch?v=T6KcnRIcG8U

Imagine how bad they'll be with Zen 2 out. Don't get me wrong, OEMs like Intel, but the cost rises vs AMD chips cannot be ignored in the numbers they are buying...

beany_bot · 19 Oct 2018 at 09:10

Does 1 tweet really require a dedicated new thread? Sorry to gripe but can't / shouldn't this go into the Zen2 thread??

Hagar · 19 Oct 2018 at 09:23

https://forums.overclockers.co.uk/threads/9700k-vs-9900k-gaming.18834251/

Did this really need a new thread also, sorry to gripe, but should this not go into core 9000 thread,

pot calling kettle etc.

beany_bot · 19 Oct 2018 at 09:25

nkata said:
https://forums.overclockers.co.uk/threads/9700k-vs-9900k-gaming.18834251/

Did this really need a new thread also, sorry to gripe, but should this not go into core 9000 thread, pot calling kettle etc.

That is a specific discussion on the i7 vs i9 in relation to gaming. Not a 10 word tweet... But yeah in hindsight it could have.

humbug · 19 Oct 2018 at 09:38

beany_bot said:
That is a specific discussion on the i7 vs i9 in relation to gaming. Not a 10 word tweet... But yeah in hindsight it could have.

This is also a discussion about performance, there is no difference between your 10 word thread and this one, except the brand of CPU.

SiDeards73 · 19 Oct 2018 at 10:14

Alexrose1uk said:
Things are only going to get worse for Intel if the current pricing situation doesnt resolve.
Things are bad right now:
https://www.youtube.com/watch?v=T6KcnRIcG8U

Imagine how bad they'll be with Zen 2 out. Don't get me wrong, OEMs like Intel, but the cost rises vs AMD chips cannot be ignored in the numbers they are buying...

Colleagues of mine were at a Dell EMC event last week at Shepperton Studios hosted by SNS and there was a lot of pushing EPYC products from DELL.. ive not yet noticed on the Dell Premier Portal much in the way of Ryzen in Business Clients yet though... been keeping my eyes peeled... However my Boss has said a few days ago, our Central IT (I work for a worldwide FTSE 500 Manufacturing company) has started pushing why arent we looking at AMD / Zen / EPYC etc.. so i can imagine before long we will see a shift.. Money men make the decisions after all.

pandem0nium · 19 Oct 2018 at 11:21

An IPC improvement in "Scientific tasks" could just be down to instruction set changes. Such as introducing AVX-512 or making the AVX-256 work more like Intel's implementation.