Discussion: AMD Ryzen

[os2wiz]

OK, what was Intel's highest performing single-core processor core back in 2012? At Intel's glacial pace of performance improvement, add 8-10% cumulatively for 4 years and I'm wondering if Zen will even have parity with Skylake?

That's leaving aside the monstrous gaffes such as the Bulldozer/Piledriver architecture missteps.

Then we need to look at TDP parity.

Then reconcile it with AMD's stated intent to not go after HEDT environments anymore.
Where did you pull the 8 to 10% per annum oerformance improvement from. That is fictional. The past 2 new iterations have been less than 6 % for Skylake and close to zero for Kabylake. You have to vack to Sandybridge to see an 8 to 10 % improvement over previous generation. They have averaged 5 to 6 % .
Then look at the unit cost of component manufacture. Will AMD chipset motherboards achieve price parity with Intel chipset motherboards?

So now that we have maybe achieved that, shall we look at the Intel processor line coming out after Kaby Lake, since I think Kaby Lake is due before Zen?

I'm sick of this Intel vs AMD thing. I'm pretty much getting reconciled to the fact that if I want to continue running Windows I have to bend over for Intel, Nvidia, and Microsoft. That I'm now a cash cow to be milked for every dollar I have.

We should look out, towards ARM, ChromeOS, or some other disruptor will introduce a platform that will become the next defacto environment where competition still drives innovation and price wars.

Intel has not had 8 to 10 % performance improvement since Sandybridge. It has been 5 to 6% per iteration over the past 5 years. Kabylake is near 0% and the Starlake was only 5%. Haswell was 5 to 6%. etc.

 

[sarwar_r87]

Since we are all dishing out numbers, ill have a go too:

from anandtech:
cinebench r15 single threaded vishera @ 4 Ghz is 102. excavator based FX would than be 112.2 (10% increment Singlethread) or 117.3(15% increment Singlethread).
not based on that zen would be:

if excavator was 10 % increment, 40%, 50% and 55% zen gain would equate to - 157.08, 168.3 and 174.4.
if excavator was 15 % increment, 40%, 50% and 55% zen gain would equate to - 164.22, 176.0 and 181.815.

and intel haswell (Intel (Haswell) Core i7 4790K) is 181 at 4Ghz with 8MB L3 and skylake (Intel (Skylake) Core i7 6700K ) is 183 and kaby (Intel (Kaby Lake) Core i7 7700K, normalized to 4 GHz) is also 183

so to conclude, with the most optimistic assumption zen will just about match haswell.

Assumption1: 15 % gain from vishra to a non-existent excavator (source http://www.bit-tech.net/hardware/2014/07/31/amd-a-10-7800-review/3)
Assumption2: 55 % gain from excavator to zen

but I think the 15% excavator increment is a bit too optimistic, its more like 10% which would make it 3.7% slower than haswell. now that difference is just peanuts.

for comparison, Intel (Ivy Bridge-E) Core i7 4930K and Intel (Sandy Bridge) Core i7 2600K are at 160 and 158.8 when normalized to 4 Ghz

 

[salgado18]

OK, what was Intel's highest performing single-core processor core back in 2012? At Intel's glacial pace of performance improvement, add 8-10% cumulatively for 4 years and I'm wondering if Zen will even have parity with Skylake?

That's leaving aside the monstrous gaffes such as the Bulldozer/Piledriver architecture missteps.

Then we need to look at TDP parity.

Then reconcile it with AMD's stated intent to not go after HEDT environments anymore.
Where did you pull the 8 to 10% per annum oerformance improvement from. That is fictional. The past 2 new iterations have been less than 6 % for Skylake and close to zero for Kabylake. You have to vack to Sandybridge to see an 8 to 10 % improvement over previous generation. They have averaged 5 to 6 % .
Then look at the unit cost of component manufacture. Will AMD chipset motherboards achieve price parity with Intel chipset motherboards?

So now that we have maybe achieved that, shall we look at the Intel processor line coming out after Kaby Lake, since I think Kaby Lake is due before Zen?

I'm sick of this Intel vs AMD thing. I'm pretty much getting reconciled to the fact that if I want to continue running Windows I have to bend over for Intel, Nvidia, and Microsoft. That I'm now a cash cow to be milked for every dollar I have.

We should look out, towards ARM, ChromeOS, or some other disruptor will introduce a platform that will become the next defacto environment where competition still drives innovation and price wars.

Intel has not had 8 to 10 % performance improvement since Sandybridge. It has been 5 to 6% per iteration over the past 5 years. Kabylake is near 0% and the Starlake was only 5%. Haswell was 5 to 6%. etc.

Replying to Karsten, improvements have been slow for two reasons: it's very hard to reach improvements higher than 10% in one year without changing the uarch entirely, and lack of competition from AMD because they tried something different and failed.

But don't lose hope just yet, Ryzen is almost here, and will change the landscape of CPUs and motherboards. If not for being a great processor, but for promoting competition, forcing Intel to push the pace. Pentium has hyperthreading now, could you see that without AMD having a strong CPU? I'd even say the i3 could become 4c/4t, i5 to 4c/8t and i7 to 6c onwards.

But I also think that "disruption" will not come from ARM or ChromeOS or Linux, because all of them still relly on the same silicon tech that's reaching its limits. We will have new tech for CPUs in the next years, and they could bring a completely new instruction set and development environment. So let's just sit back, relax, and enjoy these last years of silicon-based CPUs

 

[cdrkf]

I think despite Juan's best pseudomath efforts to downplay / ridicule / discount AMD, that RyZen will give many a pleasant surprise at the launch.

Its not like Intel have done anything special since sandy Bridge ... just suckered people into upgrading with dodgy synthetic benchmarks ... while they reaped the benefits on product costs from die shrinks.

AMD brought you x86-64 ... AMD brought you cheap CPU's ... AMD brought you HT/DCA ... AMD provided innovation ... Intel did not.

Intel's current i-series CPU's are just an extension of the P3 (after the p4 failed miserably to scale, and had dogy pipes) with an onboard memory controller ... the core series all harks back to that.

Without competition Intel would give you nothing for another 5 years ... except hype.

Well that is my 50 cents worth.

I think Juan has mainly just been trying to point out to people not to expect *too* much from Zen. If the hype train is left at full tilt instead of being happy that AMD are pushing broadwell performance people would be claiming that Zen is 20% faster clock for clock than Kaby and will 'mop the floor' with Intel- and then there would be a huge backlash when the reality of the product sinks in. Thankfully people have been somewhat more realistic in their expectations of AMD this time so the fact they look set to have a product stack that will span the majority of Intel is getting the positive reaction it deserves.

To give you an example of unchecked hype- I've just watched a video of a tech reviewer who is *already* really disappointed with Vega- because the leaks and info AMD have shared point to it being 'only' 10% faster than the 1080 (and as a result a tad lower than the GP102 based Titan XP). The thing is though, that level of performance puts it comfortably 30 - 40% faster than the Fury X, AMD's previous top gpu- and more impressively with the same core count (so that isn't just another- throw more shaders at it improvement). At the end of the day, at that level of performance it will provide a product in a market segment AMD currently doesn't occupy with anything from this gen. The other complain was that it will 'only' have 8gb of vram.... 4gb of Vram on the Fury was an issue due to it being billed as a 4k card. 8gb is plenty for 4k, I really can't see that being a bottleneck within the useful lifespan of the card. The whole issue of that boils down to the hype train though convincing everyone that Vega will be equipped with 16 - 32gb of vram and will be the fastest card in the world- none of which came from AMD.

 

[sarinaide]

I worked out the MP ratio scaling in CB and it reveals a lot.

*all scores include turbo*

6700K - 5.05
4790K - 5
4770k - 5.07
3770k - 4.95
2600k - 5

6800k - 7.5
5960X - 9.5

FX8370 - 6.3
1090T - 5.1

Lets assume the Chinese leak of 1188 @ 3.1 is correct that will equate to 141 with a MP ratio of about 9 that is rather if you scale to match the 5960X @3.4 = 1337/9.5 = 144 that is very close single thread to multi thread scaling and performance. if both are at 3ghz the ratio is 1170/1150 that is ~ 1% IPC differential.

I still think the top end Ryzen will be equal to or faster than a i7 5960X with less IPC vs IPC + SMT scaling.

 

[sarinaide]

I think despite Juan's best pseudomath efforts to downplay / ridicule / discount AMD, that RyZen will give many a pleasant surprise at the launch.

Its not like Intel have done anything special since sandy Bridge ... just suckered people into upgrading with dodgy synthetic benchmarks ... while they reaped the benefits on product costs from die shrinks.

AMD brought you x86-64 ... AMD brought you cheap CPU's ... AMD brought you HT/DCA ... AMD provided innovation ... Intel did not.

Intel's current i-series CPU's are just an extension of the P3 (after the p4 failed miserably to scale, and had dogy pipes) with an onboard memory controller ... the core series all harks back to that.

Without competition Intel would give you nothing for another 5 years ... except hype.

Well that is my 50 cents worth.

I think Juan has mainly just been trying to point out to people not to expect *too* much from Zen. If the hype train is left at full tilt instead of being happy that AMD are pushing broadwell performance people would be claiming that Zen is 20% faster clock for clock than Kaby and will 'mop the floor' with Intel- and then there would be a huge backlash when the reality of the product sinks in. Thankfully people have been somewhat more realistic in their expectations of AMD this time so the fact they look set to have a product stack that will span the majority of Intel is getting the positive reaction it deserves.

To give you an example of unchecked hype- I've just watched a video of a tech reviewer who is *already* really disappointed with Vega- because the leaks and info AMD have shared point to it being 'only' 10% faster than the 1080 (and as a result a tad lower than the GP102 based Titan XP). The thing is though, that level of performance puts it comfortably 30 - 40% faster than the Fury X, AMD's previous top gpu- and more impressively with the same core count (so that isn't just another- throw more shaders at it improvement). At the end of the day, at that level of performance it will provide a product in a market segment AMD currently doesn't occupy with anything from this gen. The other complain was that it will 'only' have 8gb of vram.... 4gb of Vram on the Fury was an issue due to it being billed as a 4k card. 8gb is plenty for 4k, I really can't see that being a bottleneck within the useful lifespan of the card. The whole issue of that boils down to the hype train though convincing everyone that Vega will be equipped with 16 - 32gb of vram and will be the fastest card in the world- none of which came from AMD.

Nobody has claimed broadwell level, maybe somewhere between haswell and broadwell is a very realistic estimation given the numbers that float around.

 

[sarwar_r87]

Lets assume the Chinese leak of 1188 @ 3.1 is correct that will equate to 141 with a MP ratio of about 9

1. you have no reason to make this assumption.


2. 1188/9 *1.097= 144. 1188*1.097/8.8 = 148. that means the sensitivity of your MP is 1.249%. that is to say if your MP is off by 1% (MP = +/- 0.09), your value will change by 2.249%. compared to haswell.

that too error prone.

EDIT: also what juan is saying is that is IPC is lower, it will leave more room for SMT, meaning your MP would be more than 9.5 for it to be more efficient than intel. also to assume an MP of 9 would also mean that amd's SMT is less efficient than intel's. I think there are CES video showing how ryzen is better at multithreading than intel when it comes to streaming games.

 

[sarinaide]

Lets assume the Chinese leak of 1188 @ 3.1 is correct that will equate to 141 with a MP ratio of about 9

1. you have no reason to make this assumption.

2. 1188/9 *1.097= 144. 1188*1.097/8.8 = 148. that means the sensitivity of your MP is 1.249%. that is to say if your MP is off by 1%, your value will change by 1.249%. or about 1% IPC difference.

that too error prone.

1% IPC is inside margin of error, you need to affect the result by around 5% to get a noticeable difference.

What we don't know is AMD SMT scaling but we assume that it is similar to Intel's as the blender result is quite similar and handbrake, we will also give Intel benefit on SMT as they ahve used it since P4 days so they have done it a lot longer so I doubt AMD's SMT implementation would actually give a greater ratio than Intel given that a 8 core FX could only push 6.3 which is way less than th 6800k's 7.5

EDIT:

i7 5960 @3.1 1220/9.5 = 128
Zen@3.1 1190/9.5 = 125

5960X @3.4 1337/9.5 = 141
Zen @3.4 1307/9.5 = 137

That is if you take Ryzen having the same level of SMT as Intel.

 

[sarwar_r87]

What we don't know is AMD SMT scaling but we assume that it is similar to Intel's as the blender result is quite similar and handbrake, we will also give Intel benefit on SMT as they ahve used it since P4 days so they have done it a lot longer so I doubt AMD's SMT implementation would actually give a greater ratio than Intel given that a 8 core FX could only push 6.3 which is way less than th 6800k's 7.5

now you are seeing the light. with how SMT performance effects this form of backward calculation.

what juan is saying is that if IPC is lower, it will leave more room for SMT, meaning your MP would be more than 9.5 for it to be more efficient than intel. also to assume an MP of 9 would also mean that amd's SMT is 11% less efficient than intel's (that's multicore + SMT-- the difference is too large, its more likely to be 9.4-9.6 if you assume intel to be at 9.5). I think there are CES video showing how ryzen is better at multithreading than intel when it comes to streaming games.

i.e. your assumption of MT = 9 is just a number that has no justification.


intel has 6 full core on 6800k, while amd 4 modules with 4 FPs. you cannot draw any conclusion from that

 

[juanrga]

AMD has never qualified 40% better at same clocks, SKU vs SKU or any number that really quantifies that, they may have inferred 40% IPC from a 9590 @5.2Ghz.

Hot Chips slides from the presentation given by one of Zen engineers

In case the footnote is too small to be read, it says Zen core compared to Excavator core.

 

[juanrga]

Therefore they measured total throughput per clock

IPC+SMT ~ Haswell

Now it remains to measure IPC. There are three possibilities:

IPC < Haswell
IPC ~ Haswell
IPC > Haswell

My belief is IPC < Haswell

I did that the other day. I took the "up to 55%" that supposedly was leaked in a Swedish forum and I got "up to Ivy/Haswell" using CB.

https://t.co/Lsa12Y6Uwu

This is a moot point if you are basing your math on kaveri.

I am not.

 

[sarwar_r87]

i7 5960 @3.1 1220/9.5 = 128
Zen@3.1 1190/9.5 = 125

5960X @3.4 1337/9.5 = 141
Zen @3.4 1307/9.5 = 137

That is if you take Ryzen having the same level of SMT as Intel.

that's a difference of 2.8%, roughly the same as what I calculated. this is more reasonable.

but I think the 15% excavator increment is a bit too optimistic, its more like 10% which would make it 3.7% slower than haswell. now that difference is just peanuts.

But a MP value beyond 9.4-9.6 is probably not correct.

I am not.

I know you are intel biased, but you have to relax your assumptions.

from anandtech:
cinebench r15 single threaded vishera @ 4 Ghz is 102. excavator based FX would than be 112.2 (10% increment Singlethread) or 117.3(15% increment Singlethread).
not based on that zen would be:

if excavator was 10 % increment, 40%, 50% and 55% zen gain would equate to - 157.08, 168.3 and 174.4.
if excavator was 15 % increment, 40%, 50% and 55% zen gain would equate to - 164.22, 176.0 and 181.815.

and intel haswell (Intel (Haswell) Core i7 4790K) is 181 at 4Ghz with 8MB L3 and skylake (Intel (Skylake) Core i7 6700K ) is 183 and kaby (Intel (Kaby Lake) Core i7 7700K, normalized to 4 GHz) is also 183

so to conclude, with the most optimistic assumption zen will just about match haswell.

Assumption1: 15 % gain from vishra to a non-existent excavator (source http://www.bit-tech.net/hardware/2014/07/31/amd-a-10-78...)
Assumption2: 55 % gain from excavator to zen

but I think the 15% excavator increment is a bit too optimistic, its more like 10% which would make it 3.7% slower than haswell. now that difference is just peanuts.

for comparison, Intel (Ivy Bridge-E) Core i7 4930K and Intel (Sandy Bridge) Core i7 2600K are at 160 and 158.8 when normalized to 4 Ghz

 

[juanrga]

I think despite Juan's best pseudomath efforts to downplay / ridicule / discount AMD, that RyZen will give many a pleasant surprise at the launch.

If it wasn't by my efforts to maintain expectations realistic, people would still be believing on a magical RyZen with IPC well above Skylake, that comes in 12-cores for desktop, uses quad-channel AM4 mobos, has 256bit FMAC units and support for AVX512, with a base clock of 4.3GHz for the 8C, and OC to 5GHz on air with easiness, 60% more efficiency than BDWE,...

And that is the short list of overhype and nonsense posted on forums. Some tech sites have contributed to the overhype and nonsense. Consider, for instance the clever guys at WCCFTECH, which have been during month writing plain nonsense like this:

One particularly interesting part of the spec is the cache. Naples features an astronomically large L3 cache, half a gigabyte large in fact. Which is unprecedented in the industry.

Naples has 64MB of L3, not 512MB. They only have been wrong by factor of 8x.

 

[juanrga]

I think Juan has mainly just been trying to point out to people not to expect *too* much from Zen. If the hype train is left at full tilt instead of being happy that AMD are pushing broadwell performance people would be claiming that Zen is 20% faster clock for clock than Kaby and will 'mop the floor' with Intel- and then there would be a huge backlash when the reality of the product sinks in. Thankfully people have been somewhat more realistic in their expectations of AMD this time so the fact they look set to have a product stack that will span the majority of Intel is getting the positive reaction it deserves.

Indeed, once a clever guy said me on a disqus thread that AMD fanboys and Intel/Nvidia fanboys overhype future AMD products by different reasons, with Intel/Nvidia fanboys waiting people to be disappointed at launch.

I prefer to be accurate on my predictions/guess, but if I have to fail, I prefer to fail by underestimating than by over-hyping.

 

[jaymc]

Latest gossip... Ryzen to be released on the 28th Feb.

http://wccftech.com/amd-launching-ryzen-cpu-gdc-2017/

 

[cdrkf]

I think despite Juan's best pseudomath efforts to downplay / ridicule / discount AMD, that RyZen will give many a pleasant surprise at the launch.

Its not like Intel have done anything special since sandy Bridge ... just suckered people into upgrading with dodgy synthetic benchmarks ... while they reaped the benefits on product costs from die shrinks.

AMD brought you x86-64 ... AMD brought you cheap CPU's ... AMD brought you HT/DCA ... AMD provided innovation ... Intel did not.

Intel's current i-series CPU's are just an extension of the P3 (after the p4 failed miserably to scale, and had dogy pipes) with an onboard memory controller ... the core series all harks back to that.

Without competition Intel would give you nothing for another 5 years ... except hype.

Well that is my 50 cents worth.

I think Juan has mainly just been trying to point out to people not to expect *too* much from Zen. If the hype train is left at full tilt instead of being happy that AMD are pushing broadwell performance people would be claiming that Zen is 20% faster clock for clock than Kaby and will 'mop the floor' with Intel- and then there would be a huge backlash when the reality of the product sinks in. Thankfully people have been somewhat more realistic in their expectations of AMD this time so the fact they look set to have a product stack that will span the majority of Intel is getting the positive reaction it deserves.

To give you an example of unchecked hype- I've just watched a video of a tech reviewer who is *already* really disappointed with Vega- because the leaks and info AMD have shared point to it being 'only' 10% faster than the 1080 (and as a result a tad lower than the GP102 based Titan XP). The thing is though, that level of performance puts it comfortably 30 - 40% faster than the Fury X, AMD's previous top gpu- and more impressively with the same core count (so that isn't just another- throw more shaders at it improvement). At the end of the day, at that level of performance it will provide a product in a market segment AMD currently doesn't occupy with anything from this gen. The other complain was that it will 'only' have 8gb of vram.... 4gb of Vram on the Fury was an issue due to it being billed as a 4k card. 8gb is plenty for 4k, I really can't see that being a bottleneck within the useful lifespan of the card. The whole issue of that boils down to the hype train though convincing everyone that Vega will be equipped with 16 - 32gb of vram and will be the fastest card in the world- none of which came from AMD.

Nobody has claimed broadwell level, maybe somewhere between haswell and broadwell is a very realistic estimation given the numbers that float around.

My comment was 'pushing Broadwell'- what I meant by that was 'approaching' or 'close to'- which if they have Haswell IPC it would be (as we've proved the gap between Haswell and Broadwell is very small, depending on specific scenario).

 

[sarinaide]

I have no issue with Juan having passive expectations,. My issue is he takes data like i posted above which was his posting comparing a ryzen 4/8 cpu at 4ghz vs a 4790k then he says he corrected data for same clocks but posts Anandtechs results which i have proven many times is turbo core scores.

I7 5960X 8/16 3-3.5ghz

Anandtech: 140/1337 MP ratio 9.5

@3ghz
140.73x3/3.5 = 120.62
120.62x9.5= 1146 or 1337*3/3.5= 1146

@3.3GHZ

1146*3.3/3=1260.6
1260.6/9.5=132.69 or 132,9*9.5=1260.6

@3.5ghz

120.62x3.5/3=140.72
140.72*9.5= 1336.87

Or

1146*3.5/3= 1337
1337/9.5= 140.73

So Anandtechs scores = turbo core.

I7 6900k 153/1547 MP ratio 10.1@3.7

@3.1ghz

153*3.1/3.7=128.19 or 1547*3.2/3.7= 1296
128.19*10.1=1296 or 1296/10.1=128

Ryzen at 3.1~estimate

1188/9.75 (split a 5960 and 6900 SMT)
121.84

@3.4

121.84*3.4/3.1= 133.54*9.75=1303

Putting all the SKU together @3.4

5960x 1297
Ryzen 1303
6900K 1421

So maybe 8.6% difference between a 5960X/Ryzen and the 6900K

I think that estimate is rather feasible

 

[sarinaide]

I know Juan is going to say 133 is sandybridge so lets just use math to prove a point.


Per Anandtech

2600K - 135/672 mp ratio 4.98

Step 1 calculate target frequency ie 3.4ghz to match ryzen

135*3.4/3.8= 120.79
120.79x4.98=601

Or

672*3.4/3.8=601
601/4.98=120.79

Correct back to turbo

601x3.8/3.4 or 120.79x4.98=672

3770

143*3.4/3.9 or 708*3.4/3.9
126 or 617
Correct back to 3.9

126*4.95 or 617*3.9/3.4=708

@3.4

Sandybrige = 120.79
Ivybridge = 126.12
Haswell = 133
Ryzen = 133

 

[juanrga]

Latest gossip... Ryzen to be released on the 28th Feb.

http://wccftech.com/amd-launching-ryzen-cpu-gdc-2017/

The description of the GDC session has been updated to

http://schedule.gdconf.com/session/optimizing-for-amd-ryzen-cpu-presented-by-amd

It could mean a new delay or not.

 

[manleysteele]

If Ryzen is a sufficing solution for the vast majority of users at a much better price point, it should be a success for AMD. That begs the question, "What would constitute a sufficing solution for the majority of users?" Looking at the final 8-10% of users, that would mean running a GTX1080 at it's maximum framerate in the games these users run. Even if they can't afford a 1080 or won't spend the money, that's the target they all see. The leaked gaming benches have demonstrated that ability. Whether it will do that across the board remains to be seen, but I believe it do it much more often than it doesn't. I don't have a price point in mind for Ryzen. I see the arguments about price on this and other, less polite threads, but I'm willing to wait and decide after the release to consider price as a metric for an evaluation. We will have the answer to these and other questions at the end of February.

Edit: Sigh. Ok, we'll see the results whenever. Waiting.

 

[sarinaide]

Im guessing March end is most likely

 

[sarwar_r87]

http://schedule.gdconf.com/session/optimizing-for-amd-ryzen-cpu-presented-by-amd

It could mean a new delay or not.

clearly those are the only two direction things can go.

@ sarinaide: they explicitly said Q1, which is till March. and also explicitly said not at the end of q1, i.e. not end of march. its more likely that they edited the entry to avoid confusion in the event that they cannot meet their target, whatever it may be

 

[sarinaide]

I was thinking 16 month year, thankfully not. But end of march is end of Q1

 

[wah007]

About a month left for release then. Should we expect any independent reviews a week or two before it releases?

 

[jdwii]

No major purchase for me just spend 330$ today on a new gaming device

However February launch sounds responsible as Amd was telling some big youtubers that Ryzen will be launching sooner in the 1st quarter then later. Plus it almost seems like they are ready for launch.