Personal computing discussed

Moderators: renee, Flying Fox, morphine

 
ultima_trev
Gerbil XP
Topic Author
Posts: 363
Joined: Sat Mar 27, 2010 11:14 am
Contact:

Quick and dirty estimation of Zen's performance and hardship in the market

Tue Jan 17, 2017 2:10 am

1) This assumes the Zen leak by Canard PC was legitimate and bases scores off that (see this if you have not already: https://www.reddit.com/r/Amd/comments/5 ... _magazine/)

2) Zen per CES demo will be clocked at 3.6 base and 3.9 turbo, gains of 14.3/11.4% versus the engineering sample (3.15/3.5) however performance doesn't necessarily scale 1:1 with clock speeds so I thought up a worst case to best case range.

The Good:  Price.  Rumors of two Ryzen SR7 SKUs launching at 350 and 500.  I doubt it will be this low since it should give performance similar to the Haswell-E i7 5960X (a 1,000 dollar CPU). Most likely they will be 600 and 700 dollars. 

The Bad:  Performance and IPC.  Ryzen is closer to Haswell-E than Broadwell-E.  It is still a wonderful improvement over Bulldozer and its derivatives however not enough to match Broadwell-E.  I'm not comparing to Skylake or Kaby Lake as I'm thinking Zen will be more focused on HEDT and server loads than consumer.

Pro Work Loads...
i7 6900K - 193.4
i7 6800K - 152.5
CPC Zen ES - 168.7
Estimated Zen at CES speeds - 180.9 to 192.3

Est Zen CES version compared to i7 6900K:  93.5% to 99.4%
Compared to i7 6800K:  118.6 to 126.1%


Gaming...
i7 6900K - 107.4
i7 6800K - 105.8
CPC Zen ES - 97.3
Est Zen at CES speeds - 102.9 to 108.4

Est Zen CES version compared to i7 6900K:  95.8 to 100.9%
Compared to i7 6800K:  97.3 to 102.5%


The ugly: Skylake-E/EP/EX

I have no doubt Intel has a good stockpile of production Skylake E/EP/EX chips at the ready, only waiting for AMD's move.  These chips will surely offer a modest IPC and power efficiency gain over Broadwell-E however Intel will price these more aggressively.  If the SR7 flagship is $700, Intel will price the Skylake-E 8 core at $750 and will be seen as a value compared to the Zen because of its noticeable gains in power efficiency, IPC and performance.

I also have no doubt Skylake EP/EX will crush Naples in highly parallel workloads as well as AMD has had little time to work on their implementation of SMT while Intel had well over a decade to do so.  Supposedly Naples will have octo-channel memory however I doubt this will be enough to overcome Intel's higher IPC, performance-per-watt and mature SMT methodology.
Ryzen 7 1800X - Corsair H60i - GA AB350 Gaming - 32GB DDR4 2933 at 16,16,16,36 - GTX 1080 at 1924 / 5264 (undervolted) - 250GB WD Blue SSD - 2TB Toshiba 7200rpm HDD
 
NTMBK
Gerbil XP
Posts: 371
Joined: Sat Dec 21, 2013 11:21 am

Re: Quick and dirty estimation of Zen's performance and hardship in the market

Tue Jan 17, 2017 4:27 am

Zen vs. Skylake-EP is going to be a very interesting fight. Intel has put a lot of effort into adding AVX-512 to Skylake-EP, meaning much bigger and more complex vector units, along with datapaths to feed it, and upgraded caches. (It's meant to have a much bigger L2 cache than consumer Skylake.) Meanwhile AMD has focused on mostly getting good performance on 128-bit vectors (i.e. SSE code, or AVX working on 128-bit vectors). 

I suspect that AMD might be able to beat Intel on core density (i.e. number of cores per socket). That would make them very competitive in plenty of traditional server workloads which rely heavily on integer code and scalar FP, but will get completely curb-stomped in any workload which can be rewritten to use AVX-512. But given that there is almost no code out there using AVX-512 right now outside of HPC (where they already have it in Xeon Phi), I'm not sure how much difference that will make in the real world. 
 
ptsant
Gerbil XP
Posts: 397
Joined: Mon Oct 05, 2009 12:45 pm

Re: Quick and dirty estimation of Zen's performance and hardship in the market

Tue Jan 17, 2017 4:57 am

I really think (and hope) that Zen will be priced at $350 and $500 for the base and "special" 8c16t model.

You forgot to add in your calculations the fact that 6900 is a 140W chip, will Zen is (probably) below 120W for the high-end model. This is very important for the server variant, because it means that the 32c64t Naples chip will be able to clock higher at the same thermal headroom as the comparable intel chip.

Anyway, I don't care much about who makes the absolute best chip. A "halo" product is just that for most buyers. What I want is more competition below the $400 zone. If the base 8c variant is indeed at $350, I will be sorely tempted. AMD doesn't need to win anything, they just need a solid product that is competitive and makes profit.
Image
 
just brew it!
Administrator
Posts: 54500
Joined: Tue Aug 20, 2002 10:51 pm
Location: Somewhere, having a beer

Re: Quick and dirty estimation of Zen's performance and hardship in the market

Tue Jan 17, 2017 5:26 am

Even if the 8C/16T end up being a bit pricey at launch, there could be affordable (and effective) 4C/8T and/or 6C/12T versions. Fingers crossed.

Image

:wink:
Nostalgia isn't what it used to be.
 
LASR
Gerbil First Class
Posts: 147
Joined: Fri Jan 10, 2014 9:35 pm

Re: Quick and dirty estimation of Zen's performance and hardship in the market

Tue Jan 17, 2017 6:24 am

I just want a 8C chip at $400. The jump from dual core -> quad core was way faster. 
 
just brew it!
Administrator
Posts: 54500
Joined: Tue Aug 20, 2002 10:51 pm
Location: Somewhere, having a beer

Re: Quick and dirty estimation of Zen's performance and hardship in the market

Tue Jan 17, 2017 6:49 am

LASR wrote:
I just want a 8C chip at $400. The jump from dual core -> quad core was way faster. 

Careful what you wish for. You can have an 8C chip today for less than $150 (and could have, for a few years already) if you're willing to settle for a Piledriver core on 32nm process. :wink:
Nostalgia isn't what it used to be.
 
Topinio
Gerbil Jedi
Posts: 1839
Joined: Mon Jan 12, 2015 9:28 am
Location: London

Re: Quick and dirty estimation of Zen's performance and hardship in the market

Tue Jan 17, 2017 8:31 am

just brew it! wrote:
LASR wrote:
I just want a 8C chip at $400. The jump from dual core -> quad core was way faster. 

Careful what you wish for. You can have an 8C chip today for less than $150 (and could have, for a few years already) if you're willing to settle for a Piledriver core on 32nm process. :wink:

This, and the fact that when IPC is identical the physics dictates that a higher core count CPU is for throughput and a lower core count one is for latency (due to its on-average higher frequency).*

[*] Absent cache effects. Given sufficient cores/threads to achieve escape from the OS's gravity. &c.
Desktop: 750W Snow Silent, X11SAT-F, E3-1270 v5, 32GB ECC, RX 5700 XT, 500GB P1 + 250GB BX100 + 250GB BX100 + 4TB 7E8, XL2730Z + L22e-20
HTPC: X-650, DH67GD, i5-2500K, 4GB, GT 1030, 250GB MX500 + 1.5TB ST1500DL003, KD-43XH9196 + KA220HQ
Laptop: MBP15,2
 
just brew it!
Administrator
Posts: 54500
Joined: Tue Aug 20, 2002 10:51 pm
Location: Somewhere, having a beer

Re: Quick and dirty estimation of Zen's performance and hardship in the market

Tue Jan 17, 2017 8:50 am

Topinio wrote:
just brew it! wrote:
LASR wrote:
I just want a 8C chip at $400. The jump from dual core -> quad core was way faster. 

Careful what you wish for. You can have an 8C chip today for less than $150 (and could have, for a few years already) if you're willing to settle for a Piledriver core on 32nm process. :wink:

This, and the fact that when IPC is identical the physics dictates that a higher core count CPU is for throughput and a lower core count one is for latency (due to its on-average higher frequency).*

[*] Absent cache effects. Given sufficient cores/threads to achieve escape from the OS's gravity. &c.

I hear where you're coming from but that is a bit of an over-simplification. It will only be true if performance/watt is also identical and we're entirely limited by thermals.
Nostalgia isn't what it used to be.
 
NTMBK
Gerbil XP
Posts: 371
Joined: Sat Dec 21, 2013 11:21 am

Re: Quick and dirty estimation of Zen's performance and hardship in the market

Tue Jan 17, 2017 9:12 am

Topinio wrote:
This, and the fact that when IPC is identical the physics dictates that a higher core count CPU is for throughput and a lower core count one is for latency (due to its on-average higher frequency).*

[*] Absent cache effects. Given sufficient cores/threads to achieve escape from the OS's gravity. &c.


[*]
Depends on how good the Turbo implementation is. If in low-thread-count situations it can boost up to the same clock speed as a theoretical 4 core, you shouldn't see much of a difference.
 
ultima_trev
Gerbil XP
Topic Author
Posts: 363
Joined: Sat Mar 27, 2010 11:14 am
Contact:

Re: Quick and dirty estimation of Zen's performance and hardship in the market

Wed Jan 18, 2017 12:38 am

Higher core count also seems to create an overhead of sorts with memory bandwidth in respect to certain workloads.  Obviously something like this on a quad/octo channel DDR4 setup will be negligible for most work loads however with desktop Zen being only dual channel I definitely foresee Broadwell-E having a distinct advantage on some rendering, encoding, compression/encryption and/or DB query benchmarks, despite Zen's "native" DDR4-3200 support (via AMP profiles, AMD's version of XMP).  
Ryzen 7 1800X - Corsair H60i - GA AB350 Gaming - 32GB DDR4 2933 at 16,16,16,36 - GTX 1080 at 1924 / 5264 (undervolted) - 250GB WD Blue SSD - 2TB Toshiba 7200rpm HDD
 
Topinio
Gerbil Jedi
Posts: 1839
Joined: Mon Jan 12, 2015 9:28 am
Location: London

Re: Quick and dirty estimation of Zen's performance and hardship in the market

Wed Jan 18, 2017 5:11 pm

NTMBK wrote:
[*]
Depends on how good the Turbo implementation is. If in low-thread-count situations it can boost up to the same clock speed as a theoretical 4 core, you shouldn't see much of a difference.

The turbo is typically not going to give as high clocks at 4T workloads on an e.g. 8C16T chip as it would on the same 4-way job on a 4C8T chip.

Part of this is the spec, part is the effect of all the background software/system threads, which on a 6C or 8C chip is to tend to keep more cores active than the 4T job needs -- so the 8C16T CPU might be running at its 6C turbo bin so this would need to be at the level that the 4C8T chip's 4C bin is.
Desktop: 750W Snow Silent, X11SAT-F, E3-1270 v5, 32GB ECC, RX 5700 XT, 500GB P1 + 250GB BX100 + 250GB BX100 + 4TB 7E8, XL2730Z + L22e-20
HTPC: X-650, DH67GD, i5-2500K, 4GB, GT 1030, 250GB MX500 + 1.5TB ST1500DL003, KD-43XH9196 + KA220HQ
Laptop: MBP15,2
 
Waco
Maximum Gerbil
Posts: 4850
Joined: Tue Jan 20, 2009 4:14 pm
Location: Los Alamos, NM

Re: Quick and dirty estimation of Zen's performance and hardship in the market

Wed Jan 18, 2017 5:26 pm

Topinio wrote:
NTMBK wrote:
[*]
Depends on how good the Turbo implementation is. If in low-thread-count situations it can boost up to the same clock speed as a theoretical 4 core, you shouldn't see much of a difference.

The turbo is typically not going to give as high clocks at 4T workloads on an e.g. 8C16T chip as it would on the same 4-way job on a 4C8T chip.

Part of this is the spec, part is the effect of all the background software/system threads, which on a 6C or 8C chip is to tend to keep more cores active than the 4T job needs -- so the 8C16T CPU might be running at its 6C turbo bin so this would need to be at the level that the 4C8T chip's 4C bin is.

Doesn't Windows have a few patches courtesy of Bulldozer to combat this? That, and unless AMD's implementation is different, there's a variable level of turbo per-core. 4 threads could be running the 4T turbo max while others are at a middling state to keep up with background work.

In short, it's complicated. :P

I surely hope Zen is every bit as competitive as we have been led to believe!
Victory requires no explanation. Defeat allows none.

Who is online

Users browsing this forum: No registered users and 1 guest
GZIP: On