New Zen microarchitecture details

[swilli89]

I assume we leave die area out of the equation.

Considering AMD does come up with a competitive product from a performance perspective, how do you reckon prices will evolve? Do you really think Intel will let AMD have a hefty piece of the pie just because they want margins to stay high? My gut instinct says they will go nuclear rather than concede precious market share. If anything, it will be Intel bringing the prices to the ground because they can.

On the other hand, if AMDs product fails to deliver as many predict it will... how much is a 8 threaded BD CPU today?!

Great that you bring up a potential move by Intel. Let's turn to history for some context.

In 2003 Intel was riding high on its Northwood based Pentium 4's. They beat the Athlon XP's by a fair margin, but not quite as bad as current Skylakes are beating up AMD's lineup.

Q4 2003 AMD launches its Athlon64. Its totally trumps what Intel is offering. At 2.0Ghz it beats the Pentium 4 @ 3.2Ghz. It is an utterly better product in every sense of the word (empowered by an nVidia chipset, no less :sneaky 50% better IPC and lower power consumption.

Oct 2003 Pricing
Athlon64 3200+: $414
Pentium 4C 3.2: $590

Now this is with Intel getting demolished by AMD's offering. If AMD merely comes out with a competitive product, based on historical data, do you really want to say that "it will be Intel bringing the prices to the ground because they can."

Intel will protect their margins just like they always have. Sure, they will adjust, but AMD will hopefully once again be relevant and not have to price their own offering too low either.

AMD doesn't need to get a majority of the market. They just need some more OEM wins for both desktop and laptop, some data center design wins (efficiency key here), and a bigger share in the home DIY builder.

 

[frozentundra123456]

I assume we leave die area out of the equation.

Considering AMD does come up with a competitive product from a performance perspective, how do you reckon prices will evolve? Do you really think Intel will let AMD have a hefty piece of the pie just because they want margins to stay high? My gut instinct says they will go nuclear rather than concede precious market share. If anything, it will be Intel bringing the prices to the ground because they can.

On the other hand, if AMDs product fails to deliver as many predict it will... how much is a 8 threaded BD CPU today?!

Or on OTOH, do you think if AMD finally brings out a competitive product, that they will not try to maximize profits as well, (ahem, original price of 9590, 7850K, and Fury X)?


The unbelievable part of this thread is that expectations for performance keep increasing and expectations for pricing get lower and lower. Now if Zen has SB/IB ipc and clocks somewhat lower than X99, 8 core zen will compete more closely with 6 core Intel. If that is the case, yes, it could sell in the 350.00 range, because 5820 is very close in price to 6700K. This actually seems like the most likely scenario to me, and it still would be a win for AMD and the consumer. But to think AMD will sell a chip of comparable performance to 5960x at 1/3 the price, well I guess some insist on believing that, but I certainly dont.

 

[swilli89]

Or on OTOH, do you think if AMD finally brings out a competitive product, that they will not try to maximize profits as well, (ahem, original price of 9590, 7850K, and Fury X)?


The unbelievable part of this thread is that expectations for performance keep increasing and expectations for pricing get lower and lower. Now if Zen has SB/IB ipc and clocks somewhat lower than X99, 8 core zen will compete more closely with 6 core Intel. If that is the case, yes, it could sell in the 350.00 range, because 5820 is very close in price to 6700K. This actually seems like the most likely scenario to me, and it still would be a win for AMD and the consumer. But to think AMD will sell a chip of comparable performance to 5960x at 1/3 the price, well I guess some insist on believing that, but I certainly dont.

Refer to my quote above. AMD has historically presented a much better value even when annihilating intel . If they come out with a product that is slightly behind in IPC but offers more cores there's not historical data to suggest that they won't try to undercut Intel's pricing to gain back design wins and market share.

 

[Shehriazad]

I legitimately wonder...are 8 cores and 16 threads going to be beneficial to DX12 and Vulkan gaming or are we still gonna durk-a-durk around with 2-4 threads even in most AAA titles? (And now don't tell me about that one in 10 titles that utilizes more...Im well aware)


Because from a gaming standpoint...these cores are only interesting if they are utilized...and DX12 and Vulkan seem to be the only way that this will ever happen...if at all.

 

[Yuriman]

I legitimately wonder...are 8 cores and 16 threads going to be beneficial to DX12 and Vulkan gaming or are we still gonna durk-a-durk around with 2-4 threads even in most AAA titles? (And now don't tell me about that one in 10 titles that utilizes more...Im well aware)


Because from a gaming standpoint...these cores are only interesting if they are utilized...and DX12 and Vulkan seem to be the only way that this will ever happen...if at all.

My best guess is that we're going to continue seeing a slow progression toward more cores being used, especially in higher budget titles. Compared with 2011, utilization of more than 2-4 cores is a lot better but still nowhere near universal.

 

[jpiniero]

I legitimately wonder...are 8 cores and 16 threads going to be beneficial to DX12 and Vulkan gaming or are we still gonna durk-a-durk around with 2-4 threads even in most AAA titles?

Probably not. Games are starting to use 6 threads pretty decently but that's about it. How much CPU power do you need anyway, when all the games are designed for the consoles?

 

[SK10H]

Soooo.....basically if AMD will sell you a 5960X equivalent for $300-400 you are ready to buy?? Good luck with that, you and about a million others better be hitting F5 all day long.

I am not looking for the top binned pre-oc golden chip reserved for $1000, there will definitely have one like the x2 days. Basically, this may be the 3rd chip from above, $1000, $600, $350, maybe with some cache cut but not core count as this is the main attraction. They can start cutting to 6core/12 thread at i5 price level.

Second, $400/cpu is very fat margin for AMD given enough buying interest while GF/Samsung can supply the volume. The 5960x is a Titan without a Hawaii.

 

[FlanK3r]

still Im thinking information about Zen and October relase+IPC near Broadwell are fakes. What I know, the Zen is planed at the end of this year, seems December time. And nospecific information about comparison performance to Intel.

 

[coercitiv]

Intel will protect their margins just like they always have. Sure, they will adjust, but AMD will hopefully once again be relevant and not have to price their own offering too low either.

AMD doesn't need to get a majority of the market. They just need some more OEM wins for both desktop and laptop, some data center design wins (efficiency key here), and a bigger share in the home DIY builder.

Flash forward 2014-2015: Intel's margins on their Atom class CPUs are high, yet their mobile division is bleeding billions. Their premium mobile products, with nice fat margins, end up in some of the cheapest phones and tablets on the market. How can that be?! (we know, but let's act surprised)

Last time Intel had to fight in the high performance CPU market, they did it behind closed doors. Fines bigger than 1 billion dollars showed us the scale of the operation. Next time, if there ever is a next time, they won't have the luxury of maintaining appearances: either they fight through pricing, or they lose market share.

[PS] I just noticed FlanK3r's signature and thought about a secret agent's closet full of guns, only his is filled with CPUs

 

[itsmydamnation]

still Im thinking information about Zen and October relase+IPC near Broadwell are fakes. What I know, the Zen is planed at the end of this year, seems December time. And nospecific information about comparison performance to Intel.

Jim Kellar when he still worked at AMD said Zen would put AMD back on performance parity with Intel. Make of that what you will.

Here is a better question, why wont Zen perform like broadwell in "high IPC"?

its wider int then broadwell,STARS,CON cores
its wider fp (in terms of units) with lower latency then boradwell,stars,con cores
it has solid prefetchers and predictors ( this was never CON's problem, see agner)
it has same cache L:S (2:1 just 1/2 as wide as broadwell)
pipeline length will be shorter then CON but will have to wait and see how short
it has a op cache (assume mop cache) , STARS/CON didn't have this, CON's really long pipeline length hurt on failed branches
CON has solid L/S pipeline (memory disambiguation) no reason Zen wont
it will have same L1 size, latency, association yet to be determined
L2 is more unknown, seen 512k thrown about, but latency should be back around STARS cores ( no write coalescing cache)
expect a large PRF like Broadwell, CON

its easy to see why bulldozer failed at "IPC" long pipeline, poor cache latency, narrow execution resources. So everyone making "IPC" guesses list your actual technical reasons why, not just cuz AMD.

i have no idea what IPC will be like, its why in all my posts here i have avioded making the guess, i only look at the data that has been found/published and so far none of that has said that Zen cant get Broadwell like performance.

 

[PPB]

My only concern is Cache performance. Until this day, AMD has the cache performance of almost decade-old products. And just recently they went smart about it, instead of trying to rework it completely (lots of $$$) on ongoing uarchs, at least they decided to save die area and get some latency shaved off by reducing it's size when it didnt give big performance penalties. It is totally important that they overcome the pandemic low performance of this area in their designs with Zen.

 

[Yuriman]

My only concern is Cache performance. Until this day, AMD has the cache performance of almost decade-old products. And just recently they went smart about it, instead of trying to rework it completely (lots of $$$) on ongoing uarchs, at least they decided to save die area and get some latency shaved off by reducing it's size when it didnt give big performance penalties. It is totally important that they overcome the pandemic low performance of this area in their designs with Zen.

^ And memory controller performance.

 

[CHADBOGA]

Jim Kellar when he still worked at AMD said Zen would put AMD back on performance parity with Intel. Make of that what you will.

Do you have a link for that?

 

[Boze]

Jim Kellar when he still worked at AMD said Zen would put AMD back on performance parity with Intel. Make of that what you will.

 

[itsmydamnation]

Do you have a link for that?

you guys are gonna go make me find it D: .......

 

[swilli89]

Flash forward 2014-2015: Intel's margins on their Atom class CPUs are high, yet their mobile division is bleeding billions. Their premium mobile products, with nice fat margins, end up in some of the cheapest phones and tablets on the market. How can that be?! (we know, but let's act surprised)

Last time Intel had to fight in the high performance CPU market, they did it behind closed doors. Fines bigger than 1 billion dollars showed us the scale of the operation. Next time, if there ever is a next time, they won't have the luxury of maintaining appearances: either they fight through pricing, or they lose market share.

[PS] I just noticed FlanK3r's signature and thought about a secret agent's closet full of guns, only his is filled with CPUs

Very very true. I forget Intel poured hundreds of millions to illegally coerce companies not to buy AMD's products. So dirty, I remember being semi-ashmed at buying the dominant 2500k half a decade ago. We the people need AMD to bring the fight this time and hopefully claw itself some decent revenue and profits to keep the cycle going.

Jim Kellar when he still worked at AMD said Zen would put AMD back on performance parity with Intel. Make of that what you will.

Here is a better question, why wont Zen perform like broadwell in "high IPC"?

its wider int then broadwell,STARS,CON cores
its wider fp (in terms of units) with lower latency then boradwell,stars,con cores
it has solid prefetchers and predictors ( this was never CON's problem, see agner)
it has same cache L:S (2:1 just 1/2 as wide as broadwell)
pipeline length will be shorter then CON but will have to wait and see how short
it has a op cache (assume mop cache) , STARS/CON didn't have this, CON's really long pipeline length hurt on failed branches
CON has solid L/S pipeline (memory disambiguation) no reason Zen wont
it will have same L1 size, latency, association yet to be determined
L2 is more unknown, seen 512k thrown about, but latency should be back around STARS cores ( no write coalescing cache)
expect a large PRF like Broadwell, CON

its easy to see why bulldozer failed at "IPC" long pipeline, poor cache latency, narrow execution resources. So everyone making "IPC" guesses list your actual technical reasons why, not just cuz AMD.

i have no idea what IPC will be like, its why in all my posts here i have avioded making the guess, i only look at the data that has been found/published and so far none of that has said that Zen cant get Broadwell like performance.

Great post, based on probable assumptions and what we know, its posts like these that keep me coming back to this forum.

 

[DrMrLordX]

My only concern is Cache performance. Until this day, AMD has the cache performance of almost decade-old products. And just recently they went smart about it, instead of trying to rework it completely (lots of $$$) on ongoing uarchs, at least they decided to save die area and get some latency shaved off by reducing it's size when it didnt give big performance penalties. It is totally important that they overcome the pandemic low performance of this area in their designs with Zen.

Zen's cache design is radically different from Construction cores, Stars, K8, and K7. It's inclusive instead of exclusive, for one thing. That ought to shake things up a bit.

 

[beginner99]

I expect sandy bridge IPC at best. AND still think that would be really good for AMD, if efficiency is there SB IPC is very good already.

At SB IPC it's basically a still born for consumer market. Any one with a nehalem or newer intel chip would have no reason to upgrade to it. They either don't buy anything or buy intel.

 

[Eric1987]

I know I am derailing but what exactly is IPC? I cannot garner the knowledge via wikipedia alone. if AMD comes with a comparable CPU soon I will definitely pick it up. I like to root for the under dog.

 

[itsmydamnation]

I know I am derailing but what exactly is IPC? I cannot garner the knowledge via wikipedia alone.

its really nothing (instructions per clock),

what people mean is high performance for a workload that has low ILP (instruction level parallelism).

 

[cbn]

I know I am derailing but what exactly is IPC? I cannot garner the knowledge via wikipedia alone.

In Layman's terms it is basically the amount of work done per clock.

So high IPC and high clockspeed would be what a person wants if seeking a processor with a powerful single thread performance.

In contrast, low IPC and low clockspeed would be the opposite (ie, weak processor single thread).

With that mentioned, the term "IPC" really means "instructions per clock" (and some instructions do more work than others)....but in this forum we do not use the term in the literal sense.

P.S. Some CPUs have more cores than others, so even if IPC and clockspeed are low the total multi-thread can still be quite good if enough cores are present.

 

[deasd]

Well, I don't like use 'IPC' to describe performance, I'm afraid it's not any better than 'Ghz'.
Intel side has improved FPU performance step by step since SandyBridge but the integer execution performance which is below 64bit roughly stay the same since Nehalem. That's why you can't feel any difference for daily usage in the previous few generations.
Reason why AMD looks worse is they don't want to follow what Intel's did such as SMT, FPU(AVX), so many factors have been differentiate, but it still being benched the same way as Intel side, when the Bulldozer born, it initial failed to clock above 4Ghz and cache problem become a disaster.
If Zen diagram is true then it imply AMD has absolutely turn around its strategy against Intel, follow what it did these years, but it would have no differentiate if it did. Only time will tell whether AMD's decision is correct or not.

 

[Exophase]

Intel side has improved FPU performance step by step since SandyBridge but the integer execution performance which is below 64bit roughly stay the same since Nehalem. That's why you can't feel any difference for daily usage in the previous few generations.

I don't know what you mean by "below 64-bit" here, AVX2 in Haswell expanded integer SIMD to 256-bit the same as it was for floating point. The SIMD part is often calling the floating point part but that is a misnomer.

There isn't a big difference because most software is simply not using AVX2 heavily, SIMD is not really easily applicable terribly broadly.

 

[cbn]

At SB IPC it's basically a still born for consumer market. Any one with a nehalem or newer intel chip would have no reason to upgrade to it. They either don't buy anything or buy intel.

I don't think Zen will have SB IPC, but remember Pollack's rule:

So if the Zen core being weaker actually resulted in substantially better performance per watt, it might work out better for mobile or ultra high core count server (eg, 32C/64T)

Then there are also people who would like to move into a desktop as console alternative (or perhaps combination NAS/Console alternative). For such people a 4C/8T or better should work out well enough especially if the price is right.

 

[JDG1980]

My only concern is Cache performance. Until this day, AMD has the cache performance of almost decade-old products. And just recently they went smart about it, instead of trying to rework it completely (lots of $$$) on ongoing uarchs, at least they decided to save die area and get some latency shaved off by reducing it's size when it didnt give big performance penalties. It is totally important that they overcome the pandemic low performance of this area in their designs with Zen.

Well, the 2015 Financial Analyst Day slides specifically mention a "new high-bandwidth, low-latency cache system" for Zen. So AMD clearly knows that this is an area that needs improvement. Of course, how they execute is the big question...