Fudzilla: New AMD Zen APU boasts up to 16 cores (plus Greenland GPU with HBM)

[Fjodor2001]

The likelihood of a brand new uarch matching Intel's sixth generation Core processor (SKL) is extremely low. From what we know now, there are no new technologies being used that Intel hasn't already implemented and had the time to perfect. Once one adds in the inevitable advantage Intel will have in electrostatics, that likelihood becomes even smaller.

The cause for optimism is precisely because so little is known about Zen. This opens a window of opportunity for grandiose speculation.

So, if you care too, indulge now and enjoy. I really don't want to ruin your party and, I really hope Zen is much more competitive that the BD line ever was. I just felt a need to drop in a little reality check.

Good luck.

8 core Zen will not beat 4 core Intel in MT loads. Really?

 

[Ajay]

Have you hard numbers that suggest that intel process is better than samsung s and that as a consequence AMD Zen will have inferior perf/Watt.?.

Otherwise all those assumptions are based on some "intel is forcibly better in everythings", certainly that it can have an entertaining value but technicaly speaking it s just filling the thread with some void arguments.

Pay attention to what IDontCare has said about these two process nodes. He has knowledge and contacts that we can't even dream about having.

8 core Zen will not beat 4 core Intel in MT loads. Really?

Oh, it should, and it better. I thought we were comparing quad to quad.

 

[Abwx]

Pay attention to what IDontCare has said about these two process nodes. He has knowledge and contacts that we can't even dream about having.

So far intel s 14nm is not much better than their 22nm, we re talking of 10-15% better perf/Watt in a favourable segment of the frequency/voltage curve, it wont be difficult for Samsung/GF to match this process and eventualy get a few parameters that measure better, be it leakage or capacitance or whatever else that this process could be good at.

 

[JDG1980]

The likelihood of a brand new uarch matching Intel's sixth generation Core processor (SKL) is extremely low. From what we know now, there are no new technologies being used that Intel hasn't already implemented and had the time to perfect. Once one adds in the inevitable advantage Intel will have in electrostatics, that likelihood becomes even smaller.

The cause for optimism is precisely because so little is known about Zen. This opens a window of opportunity for grandiose speculation.

So, if you care too, indulge now and enjoy. I really don't want to ruin your party and, I really hope Zen is much more competitive that the BD line ever was. I just felt a need to drop in a little reality check.

Good luck.

First of all, it's worth pointing out that the speculation goes both ways. Before the recent slides describing Summit Ridge and Greenland, some of the most vocal AMD detractors on this forum were insisting that Zen would be a successor to cat cores and not even attempt to compete in HPC, HEDT, and gaming.

Now, it's probably true that AMD can't leapfrog Intel's newest (and as yet unreleased) architecture. But a lot of posters here are trying to claim that AMD can't even beat Intel's 2011 Sandy Bridge in terms of IPC. To me, that's just absurd. We're talking about a design half a decade old by the time Zen hits the market, one that was already at or near release when design work on Zen first began. And people think AMD is going to fall short of even that? To me, Sandy Bridge IPC is the minimum needed for Zen to be competitive, while somewhere around Haswell IPC is the maximum we could plausibly expect. It's important to note that this has likely been the primary focus of AMD's CPU R&D for some time. Although nothing was publicly confirmed, it appears that sometime around late 2012 - early 2013, Steamroller was reevaluated and severely cut back. We know, of course, that plans for FX Steamroller and Excavator were cancelled. This makes it likely that Bulldozer and its progeny were already recognized internally as flops, and the premier designers were moved away from polishing that turd. Instead, the best people in AMD would have been moved onto the Zen project. This isn't some last-minute thing that they're slapping together. This is the future of the company; Jim Keller and Lisa Su are going to want to make damn sure it's done right.

 

[Fjodor2001]

Oh, it should, and it better. I thought we were comparing quad to quad.

AMD quad only has to come close to Intel quad to sell, assuming price is lower. AMD iGPU will be better, that is a given.

AMD 8 core will for sure beat Intel 4 core in MT, and can thus sell at a higher price. It'll be the top SKU to get for consumers.

 

[CHADBOGA]

AMD quad only has to come close to Intel quad to sell, assuming price is lower. AMD iGPU will be better, that is a given.

AMD 8 core will for sure beat Intel 4 core in MT, and can thus sell at a higher price. It'll be the top SKU to get for consumers.

That will surely depend on the single threaded performance.

 

[Fjodor2001]

That will surely depend on the single threaded performance.

Depends on the workload it's intended for.

 

[RussianSensation]

AMD quad only has to come close to Intel quad to sell, assuming price is lower. AMD iGPU will be better, that is a given.

AMD 8 core will for sure beat Intel 4 core in MT, and can thus sell at a higher price. It'll be the top SKU to get for consumers.

What if AMD positions 8-core parts against i5 and i7 and uses 4-core against the i3? Imagine 25-30% higher IPC from BD and much lower power usage courtesy of 14nm node. All of a sudden take 9590's performance and you are dangerously close to the i7 4790K. Add DX12 gaming and it's possible a true 8-core Zen will beat 4770K in games. Since a lot of people on our forum care for gaming and rendering/encoding performance, an 8-core Zen with 25-30% IPC increase over 9590 would be very good compared to where AMD is sitting today.

Think about it -- right now hardly anyone would take an FX9590 @ 4.8Ghz over even the 4.5 year old i5 2500K @ 4.8Ghz. If AMD's Zen can compete with i7 4770K/4790K and be just 1 generation behind Intel's Skylake, that would be a huge comeback! As the total package, FM3 mobo + 8-core Zen would cost less than Skylake i7 anyway which means it'll be a pretty good trade-off for a lot of consumers. I don't think Zen needs to beat Skylake to be good, but imo AMD needs to get to i7 4770K level to stay in the game and keep building on Zen in a similar fashion to how they keep improving GCN.

I mean right now FX9590 isn't even as good as the i7 920 @ 4.0Ghz. So if AMD can catch up to be just 1 generation behind Intel, that would be massive progress and a step in the right direction. Right now, FX9000 is like 4-5 generations behind.

We have to be realistic though. If you have a true 8-core CPU with 95W TDP and Intel's i7 6700K is a 4-core with HT but has 95W TDP, chances are AMD is once again counting that software becomes more multi-threaded with DX12, etc. and that each of those cores isn't as fast as each of Intel's. However, given how much behind AMD is today in perf/watt and perf/core, I think even 25-30% better performance on top of 9590 in a 95W power envelope would be an incredible catch up on their part.

 

[turtile]

I'm actually surprised that they have an 8 core model. How well could it really sell? Even Intel doesn't even try to put a product out in this area. My assumption is that the 8 core model will be the only one using high performance libraries derived from the server product.

If the new GPU arch is really efficient, I think this will definitely compete with Intel. Intel keeps making the GPU bigger and bigger...

 

[RussianSensation]

If the new GPU arch is really efficient, I think this will definitely compete with Intel. Intel keeps making the GPU bigger and bigger...

That is a great point. I totally forgot about that. Oh snap! :biggrin:

"Breaking things down to the GPU portion of Haswell, based in turn on these measurements I came up with an 87mm^2 adder for the extra hardware in Haswell GT3 vs. GT2. Doubling that 87mm^2 we get a rough idea of how big the full 40 EU Haswell GPU might be: 174mm^2. If my math is right, this means that in a quad-core Haswell GT3 die, around 65% of the die area is GPU. This is contrary to the ~33% in a quad-core Haswell GT2. I suspect a dual-core + GT3 design is at least half GPU. Meanwhile Crystalwell, the 128MB eDRAM, adds another 84mm2 die (by our measurements) to the entire package." ~AT

That means AMD can use all that extra space for the CPU cores if it makes a pure 8-core CPU, and doesn't include the GPU. That would be much more impressive imo. I think AMD should do that because Intel's GPU is useless for most of this forum anyway and just wastes useful die space. Right now I would rather take an 8-core i7 4790K at 130W TDP or 4 much larger cores than the existing version. 😛

 

[imported_ats]

If things where that easy, why didn't everyone buy Apple stock in 2001?

Looked like Apple was going bankrupt back then. Why did they let a goldmine pass...?

Actually, I've happened to know/met the investor/analyst that made the buy call on Apple at the bottom, perfectly timed for a multiplicative return. She made quite a killing as did her firm off that bet.

 

[Enigmoid]

Now, it's probably true that AMD can't leapfrog Intel's newest (and as yet unreleased) architecture. But a lot of posters here are trying to claim that AMD can't even beat Intel's 2011 Sandy Bridge in terms of IPC. To me, that's just absurd. We're talking about a design half a decade old by the time Zen hits the market, one that was already at or near release when design work on Zen first began. And people think AMD is going to fall short of even that? To me, Sandy Bridge IPC is the minimum needed for Zen to be competitive, while somewhere around Haswell IPC is the maximum we could plausibly expect. It's important to note that this has likely been the primary focus of AMD's CPU R&D for some time. Although nothing was publicly confirmed, it appears that sometime around late 2012 - early 2013, Steamroller was reevaluated and severely cut back. We know, of course, that plans for FX Steamroller and Excavator were cancelled. This makes it likely that Bulldozer and its progeny were already recognized internally as flops, and the premier designers were moved away from polishing that turd. Instead, the best people in AMD would have been moved onto the Zen project. This isn't some last-minute thing that they're slapping together. This is the future of the company; Jim Keller and Lisa Su are going to want to make damn sure it's done right.

Conroe launched July 2006
Thuban launched April 2010
Sandy Bridge launched January 2011
Stars (llano) launched August 2011
Skylake will launch ~ Aug 2015
Zen will launch ~ mid 2016

Conroe is just barely faster than stars (which is around steamroller level though steamroller has many new instructions). Give conroe an integrated memory controller and it would significantly widen the gap (you can only integrate the memory controller once though I suppose you can add HBM). And that was when AMD had money and the gap was significantly smaller than it is today.

Intel added another 5 years of progress since 2011. AMD is starting from where Intel was 10 years ago.

Steamroller's design would have been nearly completely finished 2 years before launch. If it got cut back it got cut back in other areas (like the GDDR5 rumours).

Building chips from scratch takes time. Go back and look at the intel demos. Intel demoed Haswell at IDF 2011, more than 18 months before launch. If Zen is launching next year AMD has to be finished or finishing up the last major touches on Zen. Its then a long haul to finish validation and final tweaks (generally not major changes or huge performance tweaks), adapting the chip to the process (TSMC/Samsung don't know the final finished electrical parameters of high volume 14 nm when Zen started development) with the last 6 months before launch using basically a completely locked in design - making masks and manufacturing/building up inventory/distribution.

 

[CHADBOGA]

Depends on the workload it's intended for.

Just as was the case for Bulldozer/Piledriver.

 

[cbn]

My assumption is that the 8 core model will be the only one using high performance libraries derived from the server product.

Yes, what kind of libraries AMD used on various parts is something to consider.

P.S. I have wondered how Carrizo on FM2+ would have really fared? If the CPU didn't downclock and the iGPU clocked high enough, A DDR4 version might have actually worked better in gaming than Kaveri (although non iGPU tasks would have probably been worse due to the high density library on the CPU).

 

[cbn]

One academic comparison I do hope we see in the future:

1. Quad core Carrizo with any CPU downclocking disabled (if necessary) and iGPU clockspeed tuned to desktop levels. (If can even be the DDR3 model)

vs

1. One of the quad Kaveri APUs @ 45 watt and 65 watt cTDP.

^^^^ Which of the above APUs have better iGPU gaming performance per @ 45 watt and 65 watt.

 

[cbn]

Even a 95 watt Carrizo vs. Kaveri gaming comparison would have been interesting.

Apparently Carrizo's two modules don't scale well past 50 watts (see graph on the right in the slide below), but 45 watts or more of iGPU could have been added to make a 95 watt EX APU (based on Carrizo). Examples: 50 watt EX based CPU + 45 watt iGPU = 95 watts or 40 watt EX based CPU + 55 watt iGPU = 95 watts.

EX= Excavator

 

[JoJoman88]

I look forward to a day i can comtemplate using an AMD processor again. I have skipped the whole construction equipment line much as i skipped the whole Intel P4 line. Looks like AMD has learned and like Intel did on P4 with a change in direction. I think Zen will not match Skylake but it will be a big step in it's direction. So lets hope we are welcoming AMD back to the game because a competive AMD only pushes Intel harder and we all win in the end.

 

[itsmydamnation]

Intel added another 5 years of progress since 2011. AMD is starting from where Intel was 10 years ago.

thats not true, lets look at the things AMD have added since llano that would likely get used in a new core.

1. loop cache
2. load forwarding
3. store forwarding
4. clock mesh
5. the whole CON core FPU
6. adaptive clocking + AVFS
7. L2 stream predictors
8. dynamic cache resizing
9. use of a PRF not a ROB
10. hardware DIV

Plus all the other various general improvements to components over the generations.

To pretend that AMD is throwing out the baby with the bath water because its moving away from CON cores is disingenuous.

 

[inf64]

Why I think Zen 8C/16T will roughly equal 2x8350 in threaded workloads (simplified version):

My assumption is that AMD aimed at around 30% higher (integer) ST IPC vs Excavator core, which is IMO realistic given the additional pipelines, necessary changes to load/store unit and cache changes. Because AMD states that XV core has around 5% higher IPC than SR core and we know from benchmarks that SR is round ~10% faster than PD , we could potentially have : 1.3x1.05x1.1= 1.5 or 50% higher IPC vs PD core inside 8350.

Now in MT workloads situation changes drastically. While in ST case we have a PD core at 1 and Zen core at ~1.5 mark (given equal frequency), in MT case PD module would achieve 80% of what CMT would do. CMD PD core would naturally do close to 2pts, so 80% gives you 1.6pts for a PD module's throughput in MT workload. This varies of course and scaling is sometimes worse and sometimes better, that is just an average.
Zen core on the other hand would gain additional ~25-30% due to SMT, similarly to post-SB cores. So Zen now could potentially score 1.5x1.25=1.87pts in MT workload.

I assumed that 8C/16T Zen cannot reach the clock of 8350, but would clock at roughly maximum 90% of 8350's frequency. This would reduce the MT throughput number to 1.87x0.9=1.68pts. Roughly the same as what 8350 achieves in same workload scenario.

So to sum up: if AMD managed 30% higher ST IPC vs Excavator and 90% of 8350's clock with Zen (resulting in ~3.6-3.8Ghz clock range), with 8C/16T Zen they could potentially achieve 2x8350's performance in MT workloads and around ~35% higher aggregated ST performance, stock vs stock.

 

[ShintaiDK]

If we look at Cinebench 11.5. 4790K and FX8350.

4790K scales at 4.71x. Or around 59% per core including SMT.
FX8350 scales at 6.46x. Or around 81% per core.

Just as compare, a 4690K scales at 3.95x. Or around 98.7% per core.

I say you are way overly optimistic. Not to mention Zen is still limited to dualchannel.

The TDP is 95W for the Zen octocore. That itself imposes limitations for 8 cores, SMT and 16MB L3.

And scaling is only going to get worse the more bottlenecks you have. While a 5960X can keep up the scaling like a 4790K with roughly 60% per thread due to quadchannel, 20MB cache and DDR4. Something like an X5697v3 with 28 threads drops to 50%.

 

[NTMBK]

The TDP is 95W for the Zen octocore. That itself imposes limitations for 8 cores, SMT and 16MB L3.

Remember that at 14nm, Intel can squeeze 8 2-2.6GHz Broadwell cores with SNT, 12MB L3, and a pair of 10Gb/s ethernet controllers into 45W. Even allowing for the gap between Samsung/GloFo and Intel's 14nm processes (which I won't deny), it gives you an idea of what is possible on a modern process node. And hopefully AMD have finally got a good Turbo implementation- they have been trying hard to improve it recently- so single core turbo performance could still be pretty impressive even if multi-core clocks are limited to ~3GHz.

 

[Fjodor2001]

The TDP is 95W for the Zen octocore.

Is the TDP of the 8 core Zen known? I haven't seen anything about that in the leaked slides so far...?

 

[ShintaiDK]

Remember that at 14nm, Intel can squeeze 8 2-2.6GHz Broadwell cores with SNT, 12MB L3, and a pair of 10Gb/s ethernet controllers into 45W. Even allowing for the gap between Samsung/GloFo and Intel's 14nm processes (which I won't deny), it gives you an idea of what is possible on a modern process node. And hopefully AMD have finally got a good Turbo implementation- they have been trying hard to improve it recently- so single core turbo performance could still be pretty impressive even if multi-core clocks are limited to ~3GHz.

Problem is frequency. We also know it takes Intel 140W to put 8 cores in a 5960X with its limited frequency. And Intels 14nm is substaintially better than Samsungs 14nm.

 

[ShintaiDK]

Is the TDP of the 8 core Zen known? I haven't seen anything about that in the leaked slides so far...?

Its from the insider that posted the slides.

 

[Fjodor2001]

Its from the insider that posted the slides.

Funny, I must have missed that. Got any link?