AMD Zen CPUs?

[rabidz7]

Will the AMD Zen CPUs that are expected to come out in 2016 be able to compete against 5770Ks and 5670Ks, or will they be true high-end processors, taking on Xeons and 6960Xs? Will they support dual-CPU configuration?

 

[maxius]

#1 2016 and it is ment for replacement for the cmt construction cores and aimed at the top end of the market so yes.
#2 Zen is a new smt architecture from the masterminds that gave us x86-64, the k8, the imc and ddr1 (if it wasnt for amd we would be on ia64 cpu's fed by rambus memory)
#3 amd is playing this close to the vest gl finding info before they want it out in the public
#4 very valid questions no answers yet turn in after excavator is out the door

 

[OFaceSIG]

#1 2016 and it is ment for replacement for the cmt construction cores and aimed at the top end of the market so yes.
#2 Zen is a new smt architecture from the masterminds that gave us x86-64, the k8, the imc and ddr1 (if it wasnt for amd we would be on ia64 cpu's fed by rambus memory)
#3 amd is playing this close to the vest gl finding info before they want it out in the public
#4 very valid questions no answers yet turn in after excavator is out the door

I really, really, really hope that this is true. AMD needs to get back in the game and at least this would be in the parking lot of the ballpark.

 

[maxius]

I really, really, really hope that this is true. AMD needs to get back in the game and at least this would be in the parking lot of the ballpark.

from what ive seen a few years back when they brought back the talent from other companies i am expecting them to be on the upswing with a full range of new designs in 2016 across embedded, gpu, cpu and apu

 

[Simmonz]

from what ive seen a few years back when they brought back the talent from other companies i am expecting them to be on the upswing with a full range of new designs in 2016 across embedded, gpu, cpu and apu

I am really optimistic about some of their recent hires. I think having a strong core of minds is what they need. They will never have the money Intel does but if they have smart people making the decisions and their products then they can still be strong.

Here is a quick overview of the Zen CPU architecture.

 

[KazeoHin]

There aren't any real details on the actual architecture of Zen, simply that it is not based on the BD module scheme, and AMD thinks it will be competitive at the high end. Personally I don't believe them, I want to believe them, but I don't.

They may move from their current multi threading (CMT) to something like Intel's Hyperthreading, but that would require vast changes to how they advertise their core counts. I'm hoping to see a real 8-core CPU with 'AMD threading' (shows as 16 cores in system) on high end.

 

[Mackintire]

Personally mostly because of the process that Intel will be on and their ability to scale on a given tech. I expect Intel to maintain the performance lead...


That said I would not be surprised if Jim Keller and his team were able to close the gap to within 5% per clk cycle worth of performance AND hit 4Ghz while doing so.

Considering we're at a 30-35% per clk IPC disadvantage presently, that's a massive jump in performance.

IMO the only way AMD can beat intel on the desktop is if INTEL literally screws up for 2 development cycles....which has never occurred.

 

[NaroonGTX]

Nothing is publicly known about Zen other than it exists and that it is their new scalable x86 architecture, and that it is scheduled to launch in 2016 alongside K12, the ARMv8-based part as well. The "it will be SMT-based" stuff was just rumors, no one at AMD ever said that. It could very well have elements of SMT in it, but we don't know for sure.

From what I've heard from insiders, the x86 performance is really good and it'll supposedly be more than enough to compete with Skylake in 2016. Though I don't know if this info is true since I don't work for AMD and I haven't seen any ES or prototypes or anything, so please don't quote me here as if it's a fact lol. Officially, we have no clue how it'll perform when it actually releases. Documents I've seen related to Zen and K12 label some of the IP blocks as "High IPC Cores" so I guess they weren't miffing around when they said K12 and Zen will be high-performance based.

My two cents is that you don't just go out and get some of the best architects in their respective fields like Jim Keller and Raja Koduri if you're not actually serious about building competitive, high-performance products. I think what we'll see in 2016 and onwards is, for example, Zen being very scalable to where it can go from 90W (a high-end consumer desktop/enthusiast CPU/APU) all the way down to 10~15W. You'll see them utilize these IP blocks across various different process nodes, packages, and market areas, like dense server, embedded, consumer, mobile, etc. Same for the ARM part.

So while I personally don't know how Zen will perform (it's too early to say), I think I can safely say it will be better than anything 15h has produced and will be good enough that most of the negative perception generated by Bulldozer's initial release will finally fade.

 

[KazeoHin]

The thing I have a problem with is that AMD COULD hav a 'good enough' i5k competitor if it were to make 8-core steamroller/excavator parts, but it chooses not to. What makes them feel so different with Zen?

 

[NaroonGTX]

AMD canceled plans for the SR octocores ages ago because there wasn't a suitable process node for it to be built upon. They never spoke publicly about an XV successor but most likely if they had a suitable process node to jump to, they would've done one of those as well.

Zen has nothing to do with that, though, seeing as how GloFo has stopped sucking so hard and their 16/14nm FinFET process is apparently coming along well enough.

 

[Michaelius]

This will be interesting as Intel will probably still be on 14nm then so process advantage will be closest it has been in last few years.

 

[KazeoHin]

AMD canceled plans for the SR octocores ages ago because there wasn't a suitable process node for it to be built upon. They never spoke publicly about an XV successor but most likely if they had a suitable process node to jump to, they would've done one of those as well.

Zen has nothing to do with that, though, seeing as how GloFo has stopped sucking so hard and their 16/14nm FinFET process is apparently coming along well enough.

The problem I have with that is AMD released unlocked 4 core APUs with a ton of the die devoted to a (most likely) hot running GPU, so replacing that GPU with 4 more cores and some cache is definitely possible, and could have a greater computational strength than the current FX series. Like I said, AMD chooses not to compete in the high-end arena. It's probably the best choice for them right now, but it's a choice.

 

[ebduncan]

There isn't enough information about Zen to say where they will stack up in the market.

They will likely have 2 designs. One to take the place of a High end desktop. The other to fill the mainstream market (APU).

Some reason I see them releasing a 16 core cpu and marketing it as the worlds first 16 core desktop cpu. Then having two versions of it, one with ARM cores and one with X86 cores.

 

[drescherjm]

I see them going the other way. I mean making the cores more powerful not less. I would not be surprised if the top Zen had 6 total cores.

 

[KazeoHin]

I see them going the other way. I mean making the cores more powerful not less. I would not be surprised if the top Zen had 6 total cores.

In my opinion, AMD have yet to make a TRUE successor to the phenom II X6. The 8150 was a quad-module chip, and it's performance was a step-up from a phenom X4, but a step down from X6. The 8350 is the successor to the 8150, even though it's slightly faster than an Phenom II X6, it's not an evolution of that product target. A true successor would be a 12 core (6 module) chip, but that silicon doesn't even exist even at the server level.

 

[OFaceSIG]

In my opinion, AMD have yet to make a TRUE successor to the phenom II X6. The 8150 was a quad-module chip, and it's performance was a step-up from a phenom X4, but a step down from X6. The 8350 is the successor to the 8150, even though it's slightly faster than an Phenom II X6, it's not an evolution of that product target. A true successor would be a 12 core (6 module) chip, but that silicon doesn't even exist even at the server level.

My old 1100T is now running my VMware ESX box and it's doing a fanstastic job with 16GB of DDR3 at it.

 

[Michaelius]

In my opinion, AMD have yet to make a TRUE successor to the phenom II X6. The 8150 was a quad-module chip, and it's performance was a step-up from a phenom X4, but a step down from X6. The 8350 is the successor to the 8150, even though it's slightly faster than an Phenom II X6, it's not an evolution of that product target. A true successor would be a 12 core (6 module) chip, but that silicon doesn't even exist even at the server level.

You know I was wondering why AMD doesn't try to sell some 16 core opterons as FX cpus in consumer space.

It's not like they have any sales to canibalize in server market But 16 overclockable cores could interest some enthusiasts.

 

[drescherjm]

But 16 overclockable cores could interest some enthusiasts.

I say that 2.8 to 3.0 GHz stock frequency @ 200W would make it a hard sell as a desktop processor. AMD would lose millions of dollars on on such a product targeted at such a small market.

 

[ccityinstaller]

In my opinion, AMD have yet to make a TRUE successor to the phenom II X6. The 8150 was a quad-module chip, and it's performance was a step-up from a phenom X4, but a step down from X6. The 8350 is the successor to the 8150, even though it's slightly faster than an Phenom II X6, it's not an evolution of that product target. A true successor would be a 12 core (6 module) chip, but that silicon doesn't even exist even at the server level.

In my opinion, AMD have yet to make a TRUE successor to the phenom II X6. The 8150 was a quad-module chip, and it's performance was a step-up from a phenom X4, but a step down from X6. The 8350 is the successor to the 8150, even though it's slightly faster than an Phenom II X6, it's not an evolution of that product target. A true successor would be a 12 core (6 module) chip, but that silicon doesn't even exist even at the server level.

AMD had Operton 12 core SKUs that were 2 X6 desktop dies married into one package (12 full cores in 1 cpu socket) before they went full on with the BullDozer SMT design..We had quite a few guys on our F@H team running 2P and 4P setups with them.

I am confident that Jim Keller has one more A64 type product in him..He will develop the solid performance while the rest of the team will work with him to make sure it can scale down to various power levels..

The other thing AMD may very well do is offer a marriage of high power X64 cores with some low power ARM cores that could use a translator to run native X64 code that isn't heavily demanding..When they announced the low power chipset/MB that would allow for either an X64 or ARM cpu with the same cpu socket.

 

[Michaelius]

I say that 2.8 to 3.0 GHz stock frequency @ 200W would make it a hard sell as a desktop processor. AMD would lose millions of dollars on on such a product targeted at such a small market.

Well they are selling 8 core 220W cpus anyway

 

[NaroonGTX]

You know I was wondering why AMD doesn't try to sell some 16 core opterons as FX cpus in consumer space.

It's not like they have any sales to canibalize in server market But 16 overclockable cores could interest some enthusiasts.

They don't try it because it would be a waste of time, money, and resources to do so. Those products that have more than 8 cores are simply die-harvested MCM (multi-chip module) packages which are just two Bulldozer chips on a single package. It literally couldn't physically fit into Socket AM3, so they would need to release a new separate platform and then have all their board partners make more boards for a product line which would be such an incredibly small niche that it wouldn't even be remotely profitable.

Yes, Intel offers their enthusiast platform where they sell gimped low-end Xeon's to people for outrageous prices but that's because server is where they make most of their money from anyway, so it's smarter to simply utilize die harvesting to re-badge those chips that failed verification for their intended market segment but were still in good enough working condition to meet another market (budding enthusiasts) who are willing to buy them. AMD doesn't have that luxury since their server marketshare is anemic.

That said, even if they did do it, the power consumption and TDP once OC'd would instantly turn any potential buyers away besides only the most ardent and passionate of red kool-aid drinkers.

I am confident that Jim Keller has one more A64 type product in him..He will develop the solid performance while the rest of the team will work with him to make sure it can scale down to various power levels..

Jim Keller is already a master of low-power designs. Look at his work at Apple. He and his team will definitely create a great series of products, but how great exactly it remains to be seen.

Well they are selling 8 core 220W cpus anyway

True, but the key difference here is that those parts are literally the same silicon as the rest of the FX chips on socket AM3, all they did was make sure the stock clocks were high enough to where 99% of OC'ers would have next to no headroom for further overclocking. With those chips you're basically just getting guaranteed high-clocks with it.

 

[R3MF]

The problem I have with that is AMD released unlocked 4 core APUs with a ton of the die devoted to a (most likely) hot running GPU, so replacing that GPU with 4 more cores and some cache is definitely possible, and could have a greater computational strength than the current FX series. Like I said, AMD chooses not to compete in the high-end arena. It's probably the best choice for them right now, but it's a choice.

very much agreed.
what logic reverses that choice in in eighteen months time...?

 

[NaroonGTX]

They didn't do that for FM1/FM2 because they saw no point in having such a chip for mobile devices (what most of these APU's are targeted at as they all use the same silicon from the same production line). Why bother producing such a chip on another production line if only a handful of people (let's be real here) would buy them? AMD needs to make as much cash as they can, and squandering those funds on pointless products wouldn't be a good business strategy for them.

A hypothetical chip where they "take out the GPU and replace it with more CPU cores" would also suffer due to their unfixed L3$ they had with 15h. AMD went on record saying that they hadn't fixed L3$ and it wasn't a high priority for them like a year before Kaveri came out, which pretty much hinted heavily that they had no plans to make any successors to Abu Dhabi besides that Warsaw refresh. AMD did indeed choose to not compete in high-end until they could bring forth a new design that would be more suited for such a segment...and we will see that a little over a year from now.

 

[ManofGod]

Either way, I am looking forward to it. It will definitely be an upgrade from my FX8350 and 8320. With DDR4, it will be a good time to upgrade.

 

[R3MF]

They didn't do that for FM1/FM2 because they saw no point in having such a chip for mobile devices (what most of these APU's are targeted at as they all use the same silicon from the same production line). Why bother producing such a chip on another production line if only a handful of people (let's be real here) would buy them?

that would seem to be circular logic:

because we don't produce any APU's with more cores no one will buy them, but why produce APU's with more core when no-one seems willing to buy them...

i guess you'd argue that the L3 problems led AMD to bow out until 2016, i am not yet convinced.

 

[ManofGod]

that would seem to be circular logic:

because we don't produce any APU's with more cores no one will buy them, but why produce APU's with more core when no-one seems willing to buy them...

i guess you'd argue that the L3 problems led AMD to bow out until 2016, i am not yet convinced.

Except that nothing is quite cut and dry as that. The CMT architecture, as much as I liked it, was insufficient for high end competition and did not sell as well as AMD needed them too. There was no realistic reason to produce another 8 Core CPU on the old platform they knew was coming to an end anyways.

Also, by making a new CPU today, they would have only cut into sales of the 8350 and 8320. Most folks who bought Intel this time around would not be coming back. I did want a new one but, I will just have to wait for the Zen CPU's to come out. Oh, and the post you quoted did not say no one, just very few.

 

[NaroonGTX]

i guess you'd argue that the L3 problems led AMD to bow out until 2016, i am not yet convinced.

The L3$ was just one problem from a larger picture. The bigger reason is because they didn't have a suitable process node to produce the Steamroller-based successor to Abu Dhabi (it was meant to be shrunken down to a 28nm node) so they opted to scrap plans for those chips altogether. That's sort of the same reason Kaveri initially got delayed a year. If GloFo weren't sucking so much back then, everyone would've gotten the chips they wanted by now. Alas, it didn't turn out that way.

 

[Pieter3dnow]

I would say that it is way to early to tell where the cpu will end up. Supposedly were getting more information about the design in 2015.

Even in 2016 I would have some reservations on how well this CPU will do, in general AMD had struggled with releasing new architecture and were not even going to touch on other problems related to that.

 

[defaultluser]

What L3 cache issue are you going on about?

The only issue I know of with AMD's L3 cache for Piledriver is that it's the highest latency of any previous processor design, but that's all I've heard.

 

[Quix]

Intel has been stagnate for a while so I hope AMD manages to shake it up.

 

[drescherjm]

Intel has been stagnate for a while so I hope AMD manages to shake it up.

I believe both AMD and Intel have hit a wall at the fabrication side. Hopefully I am wrong but I do not expect big gains from either until a breakthrough.

 

[ManofGod]

I believe both AMD and Intel have hit a wall at the fabrication side. Hopefully I am wrong but I do not expect big gains from either until a breakthrough.

Well, considering this will be a whole new architecture, I am expecting big gains from AMD. How much? Who knows, but, they will not be going backwards, of that, I am certain. Also, I would think that it will at least beat all the Intel Quad Cores that are out now.

 

[drescherjm]

I expect this new architecture to have a large (30% or so) improvement in IPC at the same time a reduction in cores (down to 6 cores max) and also reduction in processor frequency (there will not be a 5GHz stock+turbo chip).

 

[Digital Viper-X-]

I expect this new architecture to have a large (30% or so) improvement in IPC at the same time a reduction in cores (down to 6 cores max) and also reduction in processor frequency (there will not be a 5GHz stock+turbo chip).

I hope it's more than 30%!

 

[Geforcepat]

I expect this new architecture to have a large (30% or so) improvement in IPC at the same time a reduction in cores (down to 6 cores max) and also reduction in processor frequency (there will not be a 5GHz stock+turbo chip).

Well technically that'd be up two cores.but im hoping for an 8 core this time

 

[Trimlock]

I hope they move away from the "more smaller cores" approach.

I imagine I will have quite a few new builds by the time the Zen CPU drops so I'm hoping for some greatness from them.

 

[NaroonGTX]

What L3 cache issue are you going on about?

The only issue I know of with AMD's L3 cache for Piledriver is that it's the highest latency of any previous processor design, but that's all I've heard.

15h's L3$ was slow and woefully inefficient, especially compared to its predecessor.

I hope they move away from the "more smaller cores" approach.

Well AMD has said both Zen and K12 are "performance oriented", but that could mean anything. I don't think they will try a repeat of the Bulldozer ideology, though.

 

[ebduncan]

What ever they do with the Zen cpu's they have to be highly scale-able. Meaning one design must adapt to the low end market, high end market, and server market.

What ever they release they are going to want to match the 5960x at the very least on the high end side. So you will likely see the core counts exceed 8. Intel will have different higher performing cpus out at that time, so anything less would be considered a failure to get back into the high end market.

 

[defaultluser]

15h's L3$ was slow and woefully inefficient, especially compared to its predecessor.

Yup, bulldozer L3 was woefully inefficient at shipping speeds, but I posit that it was not improperly designed. If you will entertain my crazy thought experiment...

It's been a known thing that the latency is massive, but it was most likely designed that way for the much higher clock speeds they were planning to achieve. I mean, you don't just cut the number of ALUs and decoders per-core by 1/3 and the number of 128-bit FPUS per-core by half and expect to somehow outperform Thuban if you add "just two more cores." It's not mathematically possible without a massive clock speed increase!

See the analysis here, showing the increased cache latency over Thuban:

http://www.extremetech.com/computing/100583-analyzing-bulldozers-scaling-single-thread-performance/3

If you extrapolate, the Bulldozer shows about the same cache latency as Thuban when it hits ~5-5.5 GHz, indicating this was the target product clock speed (it would have to top 4.5 GHz to be a significant performance improvement over Thuban). We know that Piledriver can hit 5GHz in production (and several GHz higher if you use cryo cooling), so the rest of the processor seems to be built for speed. That massive 2MB L2 cache per-module is also a clear sign L3 cache was expected to be nearly as slow as main memory (see Thuban with 512K per-core, and Intel with 256K per-core, both expecting fast L3 cache).

At 5.5 GHz it's much quicker than the pathetic 3.6 GHz Bulldozer launched at. At 5.5 GHz, the CPU is still starving for bandwidth just as much *percentage-wise* as it was at 3.6 GHz, but the performance should increase nearly linearly with that massive clock speed difference, letting it run circles around Thuban/i7 2600k even with that slow cache architecture.

Only, you can't reach 5+Ghz without top-end cooling, and 6GHz is out-of-the-question. They dreamed of 10GHz just like Intel did with the P4 (which had similarly slow L2 cache), and hit the same process/power wall. When you consider how far back this design was begun (2005), it's not hard to see why they thought the sky was the limit on frequency, and process shrinks would continue to give them massive power reductions for decades to come. But I still wonder why they didn't see the failure of the P4 as something they should try not to replicate!

 

[durquavian]

Yup, bulldozer L3 was woefully inefficient at shipping speeds, but I posit that it was not improperly designed. If you will entertain my crazy thought experiment...

It's been a known thing that the latency is massive, but it was most likely designed that way for the much higher clock speeds they were planning to achieve. I mean, you don't just cut the number of ALUs and decoders per-core by 1/3 and the number of 128-bit FPUS per-core by half and expect to somehow outperform Thuban if you add "just two more cores." It's not mathematically possible without a massive clock speed increase!

See the analysis here, showing the increased cache latency over Thuban:

http://www.extremetech.com/computing/100583-analyzing-bulldozers-scaling-single-thread-performance/3

If you extrapolate, the Bulldozer shows about the same cache latency as Thuban when it hits ~5-5.5 GHz, indicating this was the target product clock speed (it would have to top 4.5 GHz to be a significant performance improvement over Thuban). We know that Piledriver can hit 5GHz in production (and several GHz higher if you use cryo cooling), so the rest of the processor seems to be built for speed. That massive 2MB L2 cache per-module is also a clear sign L3 cache was expected to be nearly as slow as main memory (see Thuban with 512K per-core, and Intel with 256K per-core, both expecting fast L3 cache).

At 5.5 GHz it's much quicker than the pathetic 3.6 GHz Bulldozer launched at. At 5.5 GHz, the CPU is still starving for bandwidth just as much *percentage-wise* as it was at 3.6 GHz, but the performance should increase nearly linearly with that massive clock speed difference, letting it run circles around Thuban/i7 2600k even with that slow cache architecture.

Only, you can't reach 5+Ghz without top-end cooling, and 6GHz is out-of-the-question. They dreamed of 10GHz just like Intel did with the P4 (which had similarly slow L2 cache), and hit the same process/power wall. When you consider how far back this design was begun (2005), it's not hard to see why they thought the sky was the limit on frequency, and process shrinks would continue to give them massive power reductions for decades to come. But I still wonder why they didn't see the failure of the P4 as something they should try not to replicate!

Agreed. The max speeds seen by me have been 5.0Ghz mostly with 5.2 and 5.4 in a very slim minority. Of course this makes me wish I had more time to run some tests as far as over-all performance. Most of us users notice that after 4.7Ghz, more so with 4.8Ghz and over that the performance is night and day over 4.6Ghz and lower. I have never really tried to quantify it but it was easily seen in Skyrim and many others mention same in other games.