Discussion: AMD Ryzen

[juanrga]

Back to the 40% statement. I would imagine is under a regular floating point operation metric, probably using AVX1. If they are talking Integer performance, then AMD is gonna be screwed again in the consumer market.

XV core: 2ALU + 2AGU
Zen core: 4ALU + 2AGU

Thus Zen core can do up to 2x more integer instructions per cycle. However, this is a peak, because integer workloads are not unending sequences of integer instructions. Assuming on a typical integer x86 workload about one half of total instructions are loads and stores whereas remaining half are branches and integer/logic operations, we can do the next rough estimation:

XV core: 4 * 50 = 200
Zen core: 6 * 50 = 300

This implies Zen could be up to 50% faster clock-for-clock on integer workloads. But of course, this rough computation assumes that rest of microarchitecture was scaled up conveniently, including the cache subsystem. It could be that the microarchitecture was scaled-up for sustaining loads to 3ALUs and that the fourth ALU was added by reasons of symmetry. In that case the improvement over Excavator would be more in the 30% range.

 

[-Fran-]

XV core: 2ALU + 2AGU
Zen core: 4ALU + 2AGU

Thus Zen core can do up to 2x more integer instructions per cycle. However, integer workloads are not unending sequences of integer instructions. Assuming on a typical integer x86 workload about one half of total instructions are loads and stores whereas remaining 50% are branches and integer/logic operations, we can do the next rough estimation:

XV core: 4 * 50 = 200
Zen core: 6 * 50 = 320

This implies Zen could be up to 50% faster on integer workloads. But of course, this rough computation assumes that rest of microarchitecture was scaled up conveniently, including the cache subsystem. It could be that the microarchitecture was scaled-up for sustaining loads to 3ALUs and that the fourth ALU was added by reasons of symmetry. In that case the improvement over Excavator would be more in the 30% range.

I don't think an AGU scales 1:1 with an ALU for operands to be honest. And like you say, it will depend a lot on the stuff supporting the logical units. In regards to the cache, that is *supposedly* the big surprise with Zen. From what we read around the news and (IIRC) AMD itself, is that Mr. Keller did a fantastic job, but of course, no numbers, haha.

In any case, 30% over Excavator is still good. If they managed to nail the cache latency to decent levels, or at least prediction, they should be in a good place for most application types (as in, not depend if a program has A or B compile flag enabled).

Cheers!

 

[juanrga]

XV core: 2ALU + 2AGU
Zen core: 4ALU + 2AGU

Thus Zen core can do up to 2x more integer instructions per cycle. However, integer workloads are not unending sequences of integer instructions. Assuming on a typical integer x86 workload about one half of total instructions are loads and stores whereas remaining 50% are branches and integer/logic operations, we can do the next rough estimation:

XV core: 4 * 50 = 200
Zen core: 6 * 50 = 320

This implies Zen could be up to 50% faster on integer workloads. But of course, this rough computation assumes that rest of microarchitecture was scaled up conveniently, including the cache subsystem. It could be that the microarchitecture was scaled-up for sustaining loads to 3ALUs and that the fourth ALU was added by reasons of symmetry. In that case the improvement over Excavator would be more in the 30% range.

I don't think an AGU scales 1:1 with an ALU for operands to be honest.

I don't know what you mean by this.

 

[-Fran-]

XV core: 2ALU + 2AGU
Zen core: 4ALU + 2AGU

Thus Zen core can do up to 2x more integer instructions per cycle. However, integer workloads are not unending sequences of integer instructions. Assuming on a typical integer x86 workload about one half of total instructions are loads and stores whereas remaining 50% are branches and integer/logic operations, we can do the next rough estimation:

XV core: 4 * 50 = 200
Zen core: 6 * 50 = 320

This implies Zen could be up to 50% faster on integer workloads. But of course, this rough computation assumes that rest of microarchitecture was scaled up conveniently, including the cache subsystem. It could be that the microarchitecture was scaled-up for sustaining loads to 3ALUs and that the fourth ALU was added by reasons of symmetry. In that case the improvement over Excavator would be more in the 30% range.

I don't think an AGU scales 1:1 with an ALU for operands to be honest.

I don't know what you mean by this.

While calculating memory pointers is associated to integer operations, they are *not* integer operations per sé, so grouping them seems a bit off, even if they both do kind of the same.

Cheers!

 

[juanrga]

It is irrelevant how we call them, what is relevant is what unit will execute them, how many units of each there exists in the design and which is the distribution of instructions in the code.

 

[010TheMaster010]

Juangra, you make me feel like waiting for Zen is not worth it D:

I suppose it depends of personals expectations. Pessimists were expecting a new Bulldozer fiasco and optimists were expecting a new K8. I was somewhat in the middle.

Let us do some math.

We know that the X4 845 (Excavator with 3.8 GHz max turbo speed) does about 95 points in single-thread Cinebench R15. The Sandy Bridge i7-2600K also have a 3.8 GHz max turbo and does 135 points. AMD claims (see slide in the OP) that Zen core has 40% higher IPC than Excavator core. Thus 95 x 1.4 = 133, so that would bring Zen IPC up to Sandy Bridge levels.

On the other hand AMD has recently shared this slide

The right hand side is not labeled. but many of us believe AMD is saying that Zen CPU for desktop is about 2x faster than Orochi on Cinebench 15 multithreaded. The million dollar question is what AMD means by Orochi?

Orochi is the codename for the server-die used in the BD-based FX and Opteron CPUs. Thus some of us assume that AMD is saying that 8-core Zen is ~2x faster than the FX-8150. The FX-8150 scores about 552 points in CBR15 multithread. Multiply by 2x and we obtain about 1104 points. The 6-core Ivy-Bridge i7 4960X does about 1097.

Others claim that Orochi means FX-8350 because Vishera chips are based in the Orochi rev. C die. Let us then take FX-8350 as baseline. The FX-8350 does about 640 points, so double that would be 1280 points. That puts us right around the E5-2687W v2. This is 8-core Ivy Bridge CPU that does about 1297 points.

Aren't you ignoring the performance boost from L3 cache in Zen that is not present on the Athlon?

Is not the 40% quoted by AMD already considering the L3?

What I mean is that, isn't your calculation not considering the lack of L3 on the 845? And if you were to compare the 845 to the 2600k, wouldn't you have to account for that? Or is the official 40% IPC boost over an L3-less Excavator? The way I've been interpreting it is that it's over excavator with L3. Not 40% over something like an athlon which has no L3.

 

[juanrga]

What I mean is that, isn't your calculation not considering the lack of L3 on the 845? And if you were to compare the 845 to the 2600k, wouldn't you have to account for that? Or is the official 40% IPC boost over an L3-less Excavator? The way I've been interpreting it is that it's over excavator with L3. Not 40% over something like an athlon which has no L3.

I am assuming that the 40% given by AMD is against Excavator without L3, because no other implementation of Excavator does exist.

 

[-Fran-]

It is irrelevant how we call them, what is relevant is what unit will execute them, how many units of each there exists in the design and which is the distribution of instructions in the code.

Weird for you to say that, haha.

It is important to know the difference between an AGU and an ALU, since they do *different* things to boost performance in integer type of operations. How a CPU arranges them is *very* important.

But yeah, I won't dig any deeper on that front. I'm not 100% sure how to discriminate performance given by an ALU or an AGU in any given design anyway.

Cheers!

 

[juanrga]

It is irrelevant how we call them, what is relevant is what unit will execute them, how many units of each there exists in the design and which is the distribution of instructions in the code.

Weird for you to say that, haha.

It is important to know the difference between an AGU and an ALU, since they do *different* things to boost performance in integer type of operations. How a CPU arranges them is *very* important.

But yeah, I won't dig any deeper on that front. I'm not 100% sure how to discriminate performance given by an ALU or an AGU in any given design anyway.

Cheers!

The difference between AGU and ALU was already considered in the computation. Precisely they were taken as two different kind of execution units each one executing different instructions. As also stated before it was an rough computation. It doesn't have the same accuracy and validity than a cycle-accurate simulator of the whole microarchitecture. But well, there are limits to what I can do.

 

[Ecky]

http://www.digitimes.com/news/a20160607PD202.html

Zen Delayed Until 2017

Market watchers originally expected Intel and AMD to launch their new CPU platforms in the third and fourth quarter, respectively, to help boost PC demand in the second half, but both CPU makers have recently postponed their launch schedules to January 2017 at CES 2017 since the upstream supply chain still has serious inventory issues, and market demand has been slow.

Intel's Kaby Lake platform was originally scheduled to launch at the end of the third quarter and AMD's Zen architecture in the fourth. Related brand vendors have been aggressively clearing their inventory of older-generation models and preparing for the new products since May, but will need to delay the plans as both CPU makers have postponed their launches.

Since most vendors still have Skylake and Haswell-based inventory waiting to be digested, while the Kaby Lake architecture does not feature any major improvements from the previous-generation architectures, Intel is not in a hurry to release Kaby Lake-based CPUs until early 2017.

AMD's Zen architecture is produced by Globalfoundries' 14nm process and chipsets are designed by Taiwan-based players. However, the platform is also expected to be delayed until January 2017.

With the delays, the PC supply chain will not be able to begin mass production for the next-generation products until November or December and PC demand is also unlikely to pick up until the first quarter of 2017.

 

[juanrga]

I am not surprised. Since 2015 I have stated multiple times my firm belief that Zen was a 2017 product,

http://semiaccurate.com/forums/showpost.php?p=247828&postcount=890
https://semiaccurate.com/forums/showpost.php?p=252444&postcount=1761

 

[Cazalan]

http://www.digitimes.com/news/a20160607PD202.html

Zen Delayed Until 2017

Market watchers originally expected Intel and AMD to launch their new CPU platforms in the third and fourth quarter, respectively, to help boost PC demand in the second half, but both CPU makers have recently postponed their launch schedules to January 2017 at CES 2017 since the upstream supply chain still has serious inventory issues, and market demand has been slow.

Intel's Kaby Lake platform was originally scheduled to launch at the end of the third quarter and AMD's Zen architecture in the fourth. Related brand vendors have been aggressively clearing their inventory of older-generation models and preparing for the new products since May, but will need to delay the plans as both CPU makers have postponed their launches.

Since most vendors still have Skylake and Haswell-based inventory waiting to be digested, while the Kaby Lake architecture does not feature any major improvements from the previous-generation architectures, Intel is not in a hurry to release Kaby Lake-based CPUs until early 2017.

AMD's Zen architecture is produced by Globalfoundries' 14nm process and chipsets are designed by Taiwan-based players. However, the platform is also expected to be delayed until January 2017.

With the delays, the PC supply chain will not be able to begin mass production for the next-generation products until November or December and PC demand is also unlikely to pick up until the first quarter of 2017.

They're just guessing like every body else. Note they also qualify it with "expected to be delayed" so they have no certainty.

 

[010TheMaster010]

What I mean is that, isn't your calculation not considering the lack of L3 on the 845? And if you were to compare the 845 to the 2600k, wouldn't you have to account for that? Or is the official 40% IPC boost over an L3-less Excavator? The way I've been interpreting it is that it's over excavator with L3. Not 40% over something like an athlon which has no L3.

I am assuming that the 40% given by AMD is against Excavator without L3, because no other implementation of Excavator does exist.

Well if that assumption is true, that makes 40% a little more lackluster :/

 

[8350rocks]

What I mean is that, isn't your calculation not considering the lack of L3 on the 845? And if you were to compare the 845 to the 2600k, wouldn't you have to account for that? Or is the official 40% IPC boost over an L3-less Excavator? The way I've been interpreting it is that it's over excavator with L3. Not 40% over something like an athlon which has no L3.

I am assuming that the 40% given by AMD is against Excavator without L3, because no other implementation of Excavator does exist.

Well if that assumption is true, that makes 40% a little more lackluster :/

L3 may account for part of that...but, if you are constantly digging into L3, those are some pretty big mispredicted code branches, too.

From what I understand, branch prediction is much improved, cache is much improved, and many other things that all sum up into the IPC category are improved.

So, if you figure solid gains in multiple areas, 40% is still impressive. Especially considering that Intel often gains their single digit increases generation over generation by adding extraneous things that 90% of software does not use. So, for all intents and purposes, most people cannot see a difference between haswell/broadwell and skylake, outside of clockspeed differences.

Now, if they are running a CPU render farm or something, sure...but your average Joe Q Gamer is not doing that.

 

[Ecky]

Even with 40% over desktop excavator, we're looking at somewhere around Ivy Bridge IPC, give or take. Although there's nothing wrong with Ivy Bridge performance today, most people will take a 6700K over a 5820K because the extra cores are pretty useless for most desktop workloads, so 8c/16t Zen will be a pretty tough sell against Kaby Lake and Skylake-E.

 

[010TheMaster010]

What I mean is that, isn't your calculation not considering the lack of L3 on the 845? And if you were to compare the 845 to the 2600k, wouldn't you have to account for that? Or is the official 40% IPC boost over an L3-less Excavator? The way I've been interpreting it is that it's over excavator with L3. Not 40% over something like an athlon which has no L3.

I am assuming that the 40% given by AMD is against Excavator without L3, because no other implementation of Excavator does exist.

Well if that assumption is true, that makes 40% a little more lackluster :/

L3 may account for part of that...but, if you are constantly digging into L3, those are some pretty big mispredicted code branches, too.

From what I understand, branch prediction is much improved, cache is much improved, and many other things that all sum up into the IPC category are improved.

So, if you figure solid gains in multiple areas, 40% is still impressive. Especially considering that Intel often gains their single digit increases generation over generation by adding extraneous things that 90% of software does not use. So, for all intents and purposes, most people cannot see a difference between haswell/broadwell and skylake, outside of clockspeed differences.

Now, if they are running a CPU render farm or something, sure...but your average Joe Q Gamer is not doing that.

I definitely agree it's still impressive, but it's unfortunate to see that I've been (likely) misinterpreting that this whole time. I doubt I would see a huge difference between my FX-6300 and something like an i3 4xxx (comparable cost here) in daily computing, but once people catch on that X is falling behind Y by Z much(no matter what unusual program(probably built for select user cases) that metric was taken from), it'll make X look like the poorer choice. And at first people see that it's only in certain scenarios, but soon after, it becomes distorted into "Don't buy X because it's Z% worse than Y." and that is not a good position to be in.

 

[juanrga]

http://www.digitimes.com/news/a20160607PD202.html

Zen Delayed Until 2017

Market watchers originally expected Intel and AMD to launch their new CPU platforms in the third and fourth quarter, respectively, to help boost PC demand in the second half, but both CPU makers have recently postponed their launch schedules to January 2017 at CES 2017 since the upstream supply chain still has serious inventory issues, and market demand has been slow.

Intel's Kaby Lake platform was originally scheduled to launch at the end of the third quarter and AMD's Zen architecture in the fourth. Related brand vendors have been aggressively clearing their inventory of older-generation models and preparing for the new products since May, but will need to delay the plans as both CPU makers have postponed their launches.

Since most vendors still have Skylake and Haswell-based inventory waiting to be digested, while the Kaby Lake architecture does not feature any major improvements from the previous-generation architectures, Intel is not in a hurry to release Kaby Lake-based CPUs until early 2017.

AMD's Zen architecture is produced by Globalfoundries' 14nm process and chipsets are designed by Taiwan-based players. However, the platform is also expected to be delayed until January 2017.

With the delays, the PC supply chain will not be able to begin mass production for the next-generation products until November or December and PC demand is also unlikely to pick up until the first quarter of 2017.

They're just guessing like every body else. Note they also qualify it with "expected to be delayed" so they have no certainty.

From the source:

both CPU makers have recently postponed their launch schedules to January 2017 at CES 2017

 

[juanrga]

What I mean is that, isn't your calculation not considering the lack of L3 on the 845? And if you were to compare the 845 to the 2600k, wouldn't you have to account for that? Or is the official 40% IPC boost over an L3-less Excavator? The way I've been interpreting it is that it's over excavator with L3. Not 40% over something like an athlon which has no L3.

I am assuming that the 40% given by AMD is against Excavator without L3, because no other implementation of Excavator does exist.

Well if that assumption is true, that makes 40% a little more lackluster :/

L3 may account for part of that...but, if you are constantly digging into L3, those are some pretty big mispredicted code branches, too.

From what I understand, branch prediction is much improved, cache is much improved, and many other things that all sum up into the IPC category are improved.

So, if you figure solid gains in multiple areas, 40% is still impressive. Especially considering that Intel often gains their single digit increases generation over generation by adding extraneous things that 90% of software does not use. So, for all intents and purposes, most people cannot see a difference between haswell/broadwell and skylake, outside of clockspeed differences.

Now, if they are running a CPU render farm or something, sure...but your average Joe Q Gamer is not doing that.

"Most people" will not be purchasing an 8C/16T CPU. Enthusiast will do. The "average Joe Q Gamer" will not be purchasing an 8C/16T CPU. Enthusiast will do.

A lot of enthusiast care about having the fastest that money can purchase. They care about having 15% more IPC, or 2 extra cores (4 extra threads) or about higher clocks, or about higher overclocking capabilities, or about extra instructions, or about extra memory channels...

 

[-Fran-]

"Most people" will not be purchasing an 8C/16T CPU. Enthusiast will do. The "average Joe Q Gamer" will not be purchasing an 8C/16T CPU. Enthusiast will do.

A lot of enthusiast care about having the fastest that money can purchase. They care about having 15% more IPC, or 2 extra cores (4 extra threads) or about higher clocks, or about higher overclocking capabilities, or about extra instructions, or about extra memory channels...

I side with Juan on this one wholeheartedly. The steam survey is great to back up this sort of statements:

Cores: http://store.steampowered.com/hwsurvey/cpus/
Speed: http://store.steampowered.com/hwsurvey/processormfg/

Combining the information from those 2, you will know the sweet spot for *most* gamers is around the $300 mark. Then you *drop* down to $100. Anything above is certainly out of a gamers scope and jumps into enthusiast class (deep pockets? XD).

AMD needs something to attack that market for gamers if they want to be relevant for them. If they go after enthusiasts, then they won't get many sales at all. I find it hard to justify spending ~$500+ on a single component. Even if the performance is great. If they manage to price a decent performing Zen alternative under $400 and close to $350 (head on for the i7 non-E), they will get market back and most definitely sell CPUs; even if the 8C/16T part is $500+. Whatever they can get in the $300 that can sit next to Intel's offerings is a win.

Cheers!

 

[Ecky]

The problem is, how to compete with Intel if your product is 3-4 generations behind in IPC, in a software environment that is not going to see much benefit past 4 threads? AMD tried that with APUs, but found they really couldn't provide a lot of value due to the memory bandwidth constraints of shared system memory, effectively capping iGPU performance at "barely adequate" levels.

Additionally, I don't think anyone else has built a part on the node(s) that Zen will be using that is much past 2ghz. 16t Zen is looking to me mostly like a server part that will also be sold to anyone who might want a workstation (or, more of an afterthought, a desktop).

 

[-Fran-]

The rationale behind people flocking over i5s and some to i7s is that they are "enough for what I want and more than plenty for what I need". You have to also add they sit at an accessible price point, which is important. The most important aspect for AMD here, is to nail down the quote with whatever product they put in the $300 range. It doesn't matter if it's *inferior* to Intel offerings, as long as that gap is justified by price AND performs within strike range.

When the 8350 came out, it actually was competitive with Sandy i5s, but Sandy destroyed the 8350 after OC and in games they weren't that close to justify spending *more* in one. Now, Intel can't OC that well as the Sandy era, so AMD has a one-time opportunity to bring parity and tick a lot of boxes for the average gamer.

Cheers!

 

[TehPenguin]

I think/hope Zen will not only be for enthusiast but also for people like me, whole love to game but also need a workstation capable of rendering "heavy duty" tasks fairly quickly.

 

[juanrga]

The rationale behind people flocking over i5s and some to i7s is that they are "enough for what I want and more than plenty for what I need". You have to also add they sit at an accessible price point, which is important. The most important aspect for AMD here, is to nail down the quote with whatever product they put in the $300 range. It doesn't matter if it's *inferior* to Intel offerings, as long as that gap is justified by price AND performs within strike range.

When the 8350 came out, it actually was competitive with Sandy i5s, but Sandy destroyed the 8350 after OC and in games they weren't that close to justify spending *more* in one. Now, Intel can't OC that well as the Sandy era, so AMD has a one-time opportunity to bring parity and tick a lot of boxes for the average gamer.

Cheers!

Even the most optimistic expect Zen to have lower clocks and not much overclocking capabilities. I have to agree with Ecky, if your IPC is behind, your clocks are behind and your OC is behind, that doesn't look good for your product.

 

[TechyInAZ]

Just wanted to let everybody know that I've redone the original post to now include the latest updates on Zen. So if anybody has any new news on Zen that I missed, please feel free to post it here and I will do my best to post it up above.

 

[8350rocks]

What I mean is that, isn't your calculation not considering the lack of L3 on the 845? And if you were to compare the 845 to the 2600k, wouldn't you have to account for that? Or is the official 40% IPC boost over an L3-less Excavator? The way I've been interpreting it is that it's over excavator with L3. Not 40% over something like an athlon which has no L3.

I am assuming that the 40% given by AMD is against Excavator without L3, because no other implementation of Excavator does exist.

Well if that assumption is true, that makes 40% a little more lackluster :/

L3 may account for part of that...but, if you are constantly digging into L3, those are some pretty big mispredicted code branches, too.

From what I understand, branch prediction is much improved, cache is much improved, and many other things that all sum up into the IPC category are improved.

So, if you figure solid gains in multiple areas, 40% is still impressive. Especially considering that Intel often gains their single digit increases generation over generation by adding extraneous things that 90% of software does not use. So, for all intents and purposes, most people cannot see a difference between haswell/broadwell and skylake, outside of clockspeed differences.

Now, if they are running a CPU render farm or something, sure...but your average Joe Q Gamer is not doing that.

I definitely agree it's still impressive, but it's unfortunate to see that I've been (likely) misinterpreting that this whole time. I doubt I would see a huge difference between my FX-6300 and something like an i3 4xxx (comparable cost here) in daily computing, but once people catch on that X is falling behind Y by Z much(no matter what unusual program(probably built for select user cases) that metric was taken from), it'll make X look like the poorer choice. And at first people see that it's only in certain scenarios, but soon after, it becomes distorted into "Don't buy X because it's Z% worse than Y." and that is not a good position to be in.

I expect that a 6 core part on a binned wafer will end up 200-400 MHz over the flagship 8C Zen, with a lower price point. That, much like your 6300, will end up being the "market winning" CPU. It would probably fall into high end i5/low end i7 money, and be quite competitive at that price point. That would put one well within reach of your average gamer.