AMD Zen

[Prysin]

I'm sure it'll perform at Haswell levels. All the improvements they've made just add up to that and if they can't get that level performance and instead get Ivybridge...I don't even want to think how that would happen.

hope is worthless.

hope wont give you more performance.

hope is for the naive.

 

Either it will or it wont. Probability and well, AMDs history of predicting their own end product is highly against ZEN being any better then haswell AT ALL

[TechGod]

hope is worthless.

hope wont give you more performance.

hope is for the naive.

 

Either it will or it wont. Probability and well, AMDs history of predicting their own end product is highly against ZEN being any better then haswell AT ALL

I don't think it will be BETTER than Haswell but perform around it. 

[JoshB2084]

Nope, I am not hype about Zen. Just a CPU, right? :rolleyes:

[Nena Trinity]

hope is worthless.

hope wont give you more performance.

hope is for the naive.

 

Either it will or it wont. Probability and well, AMDs history of predicting their own end product is highly against ZEN being any better then haswell AT ALL

True but in "theory" and note "theory" it can be up to 100% better per-core but AMD claims it will 100% sure be a 40% increase in IPC so we know that much!  (then again if they say 40% that could mean 20%)

[looncraz]

They haven't given us ANY informations except the name... 

 

We know the following, from AMD, about Zen (I will probably miss something).

 

4x full decoders (FastPath and simple)

4x ALU

2x AGU

4x 64-bit FPU

32KB L1 cache (unclear if this is data/instruction, both, either)

512KB L2

MMX, SSE, SSE2, SSE3, SSE4A, SSSE3, SSE4.1, SSE4.2, AES, AVX, AVX2, SHA, etc...

FinFet

DDR4 (possibly even DDR3 support, but AMD may be making a clean break from DDR3)

40% IPC over Excavator

Should be to market in 2016, likely late Q3 or Q4, first full year is 2017.

 

We know little else for certain (and some of the gcc patch details may be wrong, such as the FPU pipeline assignments and specific instruction latencies).

[TechGod]

We know the following, from AMD, about Zen (I will probably miss something).

 

4x full decoders (FastPath and simple)
4x ALU
2x AGU
4x 64-bit FPU
32KB L1 cache (unclear if this is data/instruction, both, either)
512KB L2
MMX, SSE, SSE2, SSE3, SSE4A, SSSE3, SSE4.1, SSE4.2, AES, AVX, AVX2, SHA, etc...
FinFet
DDR4 (possibly even DDR3 support, but AMD may be making a clean break from DDR3)
40% IPC over Excavator
Should be to market in 2016, likely late Q3 or Q4, first full year is 2017.

 

We know little else for certain (and some of the gcc patch details may be wrong, such as the FPU pipeline assignments and specific instruction latencies).

Would all these changes indicate Haswell level performance easily?

True but in "theory" and note "theory" it can be up to 100% better per-core but AMD claims it will 100% sure be a 40% increase in IPC so we know that much! (then again if they say 40% that could mean 20%)

Eh. It's unlikely it'll only be 20% over excavator.

[looncraz]

Would all these changes indicate Haswell level performance easily?

 

Easily?  Not so much.  It looks like it will be a very close call, though.  Zen should be above Ivy Bridge, though, and much closer to Haswell.

 

Zen should absolutely match Haswell for integer workloads (the two CPUs are quite similar in their basic execution unit counts), the FPU, however, is a very big question.

 

Zen has dedicated FPU pipelines (four of them), whereas Haswell shares pipelines (called "ports") with the integer units (ALUs).  This should allow Zen to encounter less of a reduction in performance when executing both types of code at once, but the effect will be small in optimized code (which will avoid using mixed code).  This could be a good way, though, for Zen to have better SMT (what Intel's hyper-threading actually is) and may well give it an extra boost in many older games.

 

After all that (and much more) come into play, we have to consider the cache latencies.  If Zen's caches are as fast as Haswell's (unlikely), then Zen should have the advantage as it has a larger L2 cache. Most likely, though, the larger L2 cache is specifically in place to bring parity to Haswell's cache hit rate.  Let's not kid ourselves and pretend AMD developed Zen without an idea as to what they were competing against :rolleyes:

 

And, finally, the thing that matters most, in the end, is clock speeds.  If Zen can't hit >4Ghz, it doesn't matter if they have Haswell IPC they can deliver more cores at a appreciable clock rate.

 

If Zen hits Sandy/Ivy Bridge IPC, AMD will have to compensate with clock-rate (they will stay well behind unless they can achieve 4.5GHz STOCK - highly unlikely).  The other option is Sandy/Ivy Bridge IPC and higher core counts... but they will still need decent clocks.

 

The obvious thing, from what we know, though, is that Zen will not overtake Intel, and the most we should expect is matching Haswell on average (beating it in integer, losing to it in floating point).

[TechGod]

Easily?  Not so much.  It looks like it will be a very close call, though.  Zen should be above Ivy Bridge, though, and much closer to Haswell.

 

Zen should absolutely match Haswell for integer workloads (the two CPUs are quite similar in their basic execution unit counts), the FPU, however, is a very big question.

 

Zen has dedicated FPU pipelines (four of them), whereas Haswell shares pipelines (called "ports") with the integer units (ALUs).  This should allow Zen to encounter less of a reduction in performance when executing both types of code at once, but the effect will be small in optimized code (which will avoid using mixed code).  This could be a good way, though, for Zen to have better SMT (what Intel's hyper-threading actually is) and may well give it an extra boost in many older games.

 

After all that (and much more) come into play, we have to consider the cache latencies.  If Zen's caches are as fast as Haswell's (unlikely), then Zen should have the advantage as it has a larger L2 cache. Most likely, though, the larger L2 cache is specifically in place to bring parity to Haswell's cache hit rate.  Let's not kid ourselves and pretend AMD developed Zen without an idea as to what they were competing against :rolleyes:

 

And, finally, the thing that matters most, in the end, is clock speeds.  If Zen can't hit >4Ghz, it doesn't matter if they have Haswell IPC they can deliver more cores at a appreciable clock rate.

 

If Zen hits Sandy/Ivy Bridge IPC, AMD will have to compensate with clock-rate (they will stay well behind unless they can achieve 4.5GHz STOCK - highly unlikely).  The other option is Sandy/Ivy Bridge IPC and higher core counts... but they will still need decent clocks.

 

The obvious thing, from what we know, though, is that Zen will not overtake Intel, and the most we should expect is matching Haswell on average (beating it in integer, losing to it in floating point).

That's literally all that matters. Red doesn't need lay waste to Blue, they just need to show them their fancy new weapon. 

Why do you think it is unlikely for Zen to not have cache latency as good as Intel's though? Oh and I read about GF's 14nm process won't leave AMD with the option to clock really high, is that true?

[looncraz]

True but in "theory" and note "theory" it can be up to 100% better per-core but AMD claims it will 100% sure be a 40% increase in IPC so we know that much!  (then again if they say 40% that could mean 20%)

 

It is entirely possible (and actually likely) that some areas will see a 70%+ improvement, and other areas will see none, very little, or even less performance.

 

Given that Lisa Su is an engineer, I'd expect her definition of IPC to be related entirely to the instruction retirement rate than benchmark results (actual Instructions Per Cycle, not the commonly used IPC =  Performance Per Cycle).

 

Retiring 40% more instructions per cycle on a core being our metric to examine, things become very cloudy.  There are times when 40% more instructions will double your performance, and times when 40% more instructions is worth 15% more performance... or even less (when something else becomes a bottleneck).  Then we don't know how well a single thread can really push the core, that 40% may only be true with SMT in action to fully saturate the core - though I kinda doubt that - Zen's a very wide design and should do quite well.

[TechGod]

It is entirely possible (and actually likely) that some areas will see a 70%+ improvement, and other areas will see none, very little, or even less performance.

 

Given that Lisa Su is an engineer, I'd expect her definition of IPC to be related entirely to the instruction retirement rate than benchmark results (actual Instructions Per Cycle, not the commonly used IPC =  Performance Per Cycle).

 

Retiring 40% more instructions per cycle on a core being our metric to examine, things become very cloudy.  There are times when 40% more instructions will double your performance, and times when 40% more instructions is worth 15% more performance... or even less (when something else becomes a bottleneck).  Then we don't know how well a single thread can really push the core, that 40% may only be true with SMT in action to fully saturate the core - though I kinda doubt that - Zen's a very wide design and should do quite well.

Did I ever mention that I love your insights?

[looncraz]

That's literally all that matters. Red doesn't need lay waste to Blue, they just need to show them their fancy new weapon. 

Why do you think it is unlikely for Zen to not have cache latency as good as Intel's though? Oh and I read about GF's 14nm process won't leave AMD with the option to clock really high, is that true?

 

Very true, if AMD can even just reach Sandy Bridge levels, they can easily compete with platform, price, and core count and they can become a viable upgrade path.  Considering their poor market positioning, Zen will help them almost no matter how it does if the net performance of a quad core Zen w/ SMT is at least equal to a stock i7 2600k in most/all areas. Below that, and the argument for using them dies as so many better options become available (though price can always change this equation).

 

As for 14nm LPP limiting clock rates, it actually shouldn't hurt them much, if at all.  In theory, it can actually be better than 32nm SOI, where they have 5GHz chips.  People hear "Low Power" in the name and automatically assume that means it can't hit high clocks, but that's not the case - it just makes the design much more important.

 

In the end, the only thing that determines how fast a processor can clock is the time it takes the longest running stage of the CPU to complete and have its (correct) results sitting on the output when the next clock cycle arrives.  A low power process usually means you have to trade transistor switching speed for power savings, which means lower clock rates (slower transistors means the longest stage now takes even longer).  14nm LPP has faster switching transistors than 32nm SOI, while also having lower leakage current (less power use, lower heat generation).

 

The limiting factor created by the process node, now, is the ability to feed power where you need it reliably.  Intel compensated by rising the CPU voltage so they could saturate the FinFets faster and have them switch more quickly.  This is an old trick - it's basic overclocking 101 ;-)  Raise the voltage, transistors switch faster, signals are cleaner, the longest running stage now takes less time to complete and you can increase the clock rate.  The limiting factors here are heat and the line current handling abilities - and both of those get worse with smaller lithography.

[TechGod]

Very true, if AMD can even just reach Sandy Bridge levels, they can easily compete with platform, price, and core count and they can become a viable upgrade path.  Considering their poor market positioning, Zen will help them almost no matter how it does if the net performance of a quad core Zen w/ SMT is at least equal to a stock i7 2600k in most/all areas. Below that, and the argument for using them dies as so many better options become available (though price can always change this equation).

 

As for 14nm LPP limiting clock rates, it actually shouldn't hurt them much, if at all.  In theory, it can actually be better than 32nm SOI, where they have 5GHz chips.  People hear "Low Power" in the name and automatically assume that means it can't hit high clocks, but that's not the case - it just makes the design much more important.

 

In the end, the only thing that determines how fast a processor can clock is the time it takes the longest running stage of the CPU to complete and have its (correct) results sitting on the output when the next clock cycle arrives.  A low power process usually means you have to trade transistor switching speed for power savings, which means lower clock rates (slower transistors means the longest stage now takes even longer).  14nm LPP has faster switching transistors than 32nm SOI, while also having lower leakage current (less power use, lower heat generation).

 

The limiting factor created by the process node, now, is the ability to feed power where you need it reliably.  Intel compensated by rising the CPU voltage so they could saturate the FinFets faster and have them switch more quickly.  This is an old trick - it's basic overclocking 101 ;-)  Raise the voltage, transistors switch faster, signals are cleaner, the longest running stage now takes less time to complete and you can increase the clock rate.  The limiting factors here are heat and the line current handling abilities - and both of those get worse with smaller lithography.

Do you think that AMD's (assuming Zen delivers on what they promise) Zen+ has a chance to catch up with Kaby/Icelake if Intel keeps going on their path of at most 5% performance bumps? 

[TechGod]

*Bunch of awesome new info for me* 

Oh and I would say the unified socket would be a massive advantage. Right now for intel you have the enthusiast grade motherboards (X99) the mainstream enthusiast (Z170) and then the H170 (I think) chipset. If AMD delivers Haswell level performance with all CPUs unlocked and a single platform for their chips, then they've easily got a customer right here. Might even stop Intel from fragmenting CPUs...

[rvborgh]

Oh and I would say the unified socket would be a massive advantage. Right now for intel you have the enthusiast grade motherboards (X99) the mainstream enthusiast (Z170) and then the H170 (I think) chipset. If AMD delivers Haswell level performance with all CPUs unlocked and a single platform for their chips, then they've easily got a customer right here. Might even stop Intel from fragmenting CPUs...

Rumors are that the first samples of Zen were supposedly at Haswell level performance... 

http://seekingalpha.com/user/19326301/comments

"What I heard from some AMD guys is that they already start testing "their new cpu" in the lab and seeing haswell level performance."

[TechGod]

Rumors are that the first samples of Zen were supposedly at Haswell level performance... 

http://seekingalpha.com/user/19326301/comments

"What I heard from some AMD guys is that they already start testing "their new cpu" in the lab and seeing haswell level performance."

Talking with a grain of salt but still very good news. 

[superjuddy]

They will be 14nm compared to the current gen being mostly 28nm so it should be quite a big boost of performance. plus general improvements. go on wccftech or some other site if you want to learn more about the actual process.

[BrightCandle]

Haswell level performance at the end of 2017 would be a bit of a disaster really, that would put them 3 CPUs behind Intel at that point. That isn't going to save the company, being faster and better is the only way to do that and the problem is they are so far behind and have such enormous problems with delivery and fullfilling the promises they make about their products that its unfortunately not really looking all that good. If we could buy one today it wouldn't be terrible but still uncompetitive and a budget option, next year it'll be bargain bin for that level of performance.

[BuckGup]

Probably

[looncraz]

Do you think that AMD's (assuming Zen delivers on what they promise) Zen+ has a chance to catch up with Kaby/Icelake if Intel keeps going on their path of at most 5% performance bumps? 

 

 

It will catch up with Skylake, but it really depends on the performance profile and where the improvements are targeted.

 

 

Oh and I would say the unified socket would be a massive advantage. Right now for intel you have the enthusiast grade motherboards (X99) the mainstream enthusiast (Z170) and then the H170 (I think) chipset. If AMD delivers Haswell level performance with all CPUs unlocked and a single platform for their chips, then they've easily got a customer right here. Might even stop Intel from fragmenting CPUs...

 

A unified socket can be a huge advantage as AMD could, conceivably, charge exactly what Intel does for high end CPUs (at the same performance level) and rely entirely on the cheaper platform to drive sales.

 

Another thing that could be a decent advantage, though they may not do, would be to use their already existing DDR3/4 controller and allow the use of normal DDR3 (not DDR3L).  I'd love to keep my DDR3 when I upgrade, that will save me a good amount of money, most of which would go into the CPU.  They've done this before, but it sometimes limits the upgrade path.  I'm just hoping Zen+ will be on AM4 and not AM4+, so that motherboards will need to switched out to upgrade.

[looncraz]

Haswell level performance at the end of 2017 would be a bit of a disaster really, that would put them 3 CPUs behind Intel at that point. That isn't going to save the company, being faster and better is the only way to do that and the problem is they are so far behind and have such enormous problems with delivery and fullfilling the promises they make about their products that its unfortunately not really looking all that good. If we could buy one today it wouldn't be terrible but still uncompetitive and a budget option, next year it'll be bargain bin for that level of performance.

 

Not really, the difference between Skylake and Haswell is rather small.  AMD doesn't need to be faster, they need to be fast enough to be useful. Haswell CPUs still sell like hotcakes.  If AMD can match the i5 4690k with the same core count, they will have a small CPU they can sell for $220, versus the massive 8-cores they have which can't even command that price.  They can then enable SMT and add $50 and create an entire market niche which belongs entirely to them and which would be quite lucrative (assuming 15% SMT scaling).

 

I'd love a 6/8 core CPU with Haswell IPC with high enough clocks to match Sandy Bridge IPC @ 4.5GHz.

 

AMD is a good 40% behind Haswell on average, nearly 100% in some areas, so achieving parity would be a fantastic improvement.

[App4that]

Not really, the difference between Skylake and Haswell is rather small. 

 

Why oh why do people keep saying this? WHY?

 

[Citadelen]

People keep throwing around the 40% IPC improvement as though that's the overall performance increase, it's a purely architectural improvement, there will be a 40% IPC improvement no matter what AMD do. Throw in Hyper-Threading, a competent caching system, the 14nm node, newer instruction sets and the fact that the team is being allowed to do what they want with it, instead of being restricted like they were with Bulldozer and you've got something really promising. It was also being designed by Jim Keller, involved in K7, the lead designer of K8,  co-designer of x86-64 instruction set and HyperTransport interconnect.

People I don't see how it can't be amazing.

[valentas]

Why oh why do people keep saying this? WHY?

 

This one is really good review.

[rvborgh]

Haswell level performance at the end of 2017 would be a bit of a disaster really, that would put them 3 CPUs behind Intel at that point. That isn't going to save the company, being faster and better is the only way to do that and the problem is they are so far behind and have such enormous problems with delivery and fullfilling the promises they make about their products that its unfortunately not really looking all that good. If we could buy one today it wouldn't be terrible but still uncompetitive and a budget option, next year it'll be bargain bin for that level of performance.

 

i could be wrong (i'm just a layman) when it comes to these things but i really doubt that there is too much low hanging fruit with AMD64... there is only so much parallelism (ILP) that exists in the code to be exploited.  Perhaps someone more informed can correct, but i would think that at some point this means that Intel and AMD are converging to the same core integer performance.  Intel has just gotten there faster.  i do wonder if that is why despite throwing a lot more transistors at it, single thread performance seems to have tapered off these past few generations for Intel.

 

[TechGod]

People keep throwing around the 40% IPC improvement as though that's the overall performance increase, it's a purely architectural improvement, there will be a 40% IPC improvement no matter what AMD do. Throw in Hyper-Threading, a competent caching system, the 14nm node, newer instruction sets and the fact that the team is being allowed to do what they want with it, instead of being restricted like they were with Bulldozer. It was also being designed by Jim Keller, involved in K7, the lead designer of K8, co-designer of x86-64 instruction set and HyperTransport interconnect.

People I don't see how it can't be amazing.

Exactly! This so much...