AMD Raven Ridge 'Zen APU' Thread

[Valantar]

~220 vs 126 mm².

I expect the CPU performance to be around 10% lower that Kaby Lake, but the GPU performance to be 3x better (Kaby Lake has 0.4 TFLOPS on GT2, I expect RR to be clocked around 850 Mhz - 11*64*2*0.850=1.2 TFLOPS - similar to RX 550).

This doesn't sound very good for the mobile version. Not that many mobile CPUs are used for gaming. The margins will be a lot lower when going against Intel in this sector.

For the desktop version (even though I don't remember AMD saying anything about a desktop APU for Ryzen) I expect the CPU 5% slower than with Kaby Lake and 4x better at 1.6 TFLOPS (~1.1 Ghz).

Raven Ridge seems a great buy for e-sports laptops.

1. Although KBL GT2 is far more common, GT4 is a more relevant comparison in terms of die area. See my post above.
2. I think you're way off in terms of clocks. You really expect the Vega (optimized for higher clocks, remember?) iGPU in RR to be clocked the same as or lower than the i GPUs on 28nm Kaveri? That sounds very unlikely. I would expect it at at least 1200MHz, but more wouldn't surprise me. The mobile versions might be around your estimate, but I think that sounds low too.
3. 3x better performance than KBL GT2? Wow, that sounds crazy low. Sure, DDR4 is a bottleneck. But that big?

If AMD is making an 11CU RR, it's meant for light gaming/esports and/or compute, not as a "slightly better than Intel's iGPU" solution. That would be bonkers. What incentive would anyone have to buy it?

 

[T1beriu]

As shown in the pic with many coins, there are diff size.

Found this pic, the one in the upper right corner has same letters and seems to be 26.5 or 28mm

My analysis is correct. I'm going to upload this image and give you this link. I gave you all the clues you need.

 

[coercitiv]

For mobile, I expect the CPU performance to be around 10% lower that Kaby Lake 7700HQ

For the desktop version I expect the CPU to be 5% slower than with Kaby Lake 7400

So in mobile it's 10% lower than 4c/8t @ 3.4Ghz in MT loads, and in desktop it's 5% lower than 4c/4t @ 3.3Ghz in MT loads. Care to share why you expect such a drastic drop in performance from mobile to desktop TDP?

Even if we only take ST performance into account, going from 10% under 3.8Ghz KBL in mobile to 5% under 3.5Ghz KBL in desktop makes little sense considering KBL drops clocks by 8% while RR surely gets higher clocks in desktop.

 

[T1beriu]

Vs Kaby Lake, it will wipe the floor with it , that's certain because Kaby Lake is focused on dual core while RR is quad all the way or almost all the way.

How can RR wipe the floor with KL since R5 1400 is 3% behind i5 7400 according to Computerbase (app + games) and both run at similar clock speeds, BUT R5 is 4c/8t while i5 is 4c/4t? And 7700HQ is a better comparison because it's 4c/8t and clocks higher (3.6-3.8 in 1/2 cores)? We're ignoring 1500X because it has 16MB L3 cache.

As for Vega, hard to imagine that it's clocked bellow 1GHz if it was designed for high clocks

I think you're way off in terms of clocks. I would expect it at at least 1200MHz

You're probably right. I've thought of that myself. I'm thinking at 1.2 is turbo & sub 900 for base clocks for mobile & 1.4/1.1 for desktop. If there's a GPU ONLY workload, sure, I can imagine going over 1 to 1.2 Ghz, but when the CPU starts cranking, I see it dropping to sub 0.9Ghz, a real use case scenario.

So in mobile it's 10% lower than 4c/8t @ 3.4Ghz in MT loads, and in desktop it's 5% lower than 4c/4t @ 3.3Ghz in MT loads. Care to share why you expect such a drastic drop in performance from mobile to desktop TDP?

Because the number of threads doubles.

 

[tamz_msc]

I am expecting a positive outcome in terms of power consumption, and therefore, battery life. Even though performance might be slightly slower. On the other hand, the 1700 in the Asus ROG laptop scoring 1410 in CB15 gives me hope.

 

[coercitiv]

Because the number of threads doubles.

You're gonna have to make more sense than that.

 

[T1beriu]

You're gonna have to make more sense than that.

I already explained it in the first paragraph of the previous post. When comparing a R5 4c/8t to a i5 4c/4t the difference is 5% but I expect it to increase when the comparison is made a against 4c/8t because mobile KL has extra 4 threads to play with while RR is already maxed out on threads.

It's kinda the same when comparing a 1500X with 7700K in Computerbase (apps + games). Both are 4c/8t, but R5 is cheating a bit because it has 16MB L3. The difference is 25% for 7700K. The extra 4 threads have an impact.

Now, on closer inspection, I think I was too conservative with the 5/10% performance comparison. I expect KL (2+2) to be 10/20% faster that RR in CPU workloads.

 

[dark zero]

Doesn't seem to be the same coin, the writing is different.
The one in the pic is like this

While the one on wiki is

Indeed, but look well at the bottom right. Seems that the coins are from different editions too.

 

[coercitiv]

I already explained it in the first paragraph of the previous post. When comparing a R5 4c/8t to a i5 4c/4t the difference is 5% but I expect it to increase when the comparison is made a against 4c/8t because mobile KL has extra 4 threads to play with while RR is already maxed out on threads.

The 7700HQ is only 10% slower than it's desktop 7700 counterpart (3.6Ghz vs 4Ghz). Look at the same Computerbase benchmark you used as example and notice the 7700 is around 28-35% faster than 7400, hence 7700HQ is bound to be around 20% faster than 7400. Based on your estimate the RR mobile SKU will be faster than the desktop SKU, which we both know is the other way around.

 

[piesquared]

Wasn't Bristol Ridge already very efficient on 28nm? With Raven Ridge having 100% better efficiency than that is crazy, it should be a clear winner in the mobile market.

 

[Arachnotronic]

Looks like no HBM2 for Raven Ridge, just as I predicted oh so long ago...

 

[majord]

I thought that was confirmed oh so long ago now?

 

[Valantar]

I thought that was confirmed oh so long ago now?

Not confirmed, but no real indication of it, so the realistic view has been that for a while, yes. Still, it's a shame that an iGPU that good has to contend with a 64-bit DDR4 bus shared with the CPU. I would love to see RR or a follow-up with 1-2GB of HBM2 (that should be plenty!) reserved for the GPU. That would make for a kick-ass system.

 

[imported_jjj]

How can RR wipe the floor with KL since R5 1400 is 3% behind i5 7400 according to Computerbase (app + games) and both run at similar clock speeds, BUT R5 is 4c/8t while i5 is 4c/4t? And 7700HQ is a better comparison because it's 4c/8t and clocks higher (3.6-3.8 in 1/2 cores)? We're ignoring 1500X because it has 16MB L3 cache.

You're probably right. I've thought of that myself. I'm thinking at 1.2 is turbo & sub 900 for base clocks for mobile & 1.4/1.1 for desktop. If there's a GPU ONLY workload, sure, I can imagine going over 1 to 1.2 Ghz, but when the CPU starts cranking, I see it dropping to sub 0.9Ghz, a real use case scenario.

.

Slow down a bit and think because you entirely missed my point.
Raven Ridge is mobile first and i am talking mobile, hope you are too.

In the mobile context we are not talking IPC , we are talking perf per target TDP, we are not talking ST we are talking MT.
- You don't compare clock for clock, you compare W for W.
- It's not about ST because a single core loaded doesn't max the TDP. You hit the TDP limits when you load all cores.

Vs Kaby Lake, KbL is dual core centric while almost all (if not all) RR SKUs will be quad cores. RR will have a 4 cores CCX so every die has 4 cores and maybe they have some lower end SKUs with cores disabled. but they gain more financially if almost all are quads.
This is why RR will obliterate Kaby in laptop and there is no doubt about it. Will have much better perf (MT) and better power as you spread the load on more cores and gain efficiency.
The Intel mobile SKUs https://en.wikipedia.org/wiki/Kaby_Lake#Mobile_processors
Look at core and thread count, at TDP and pricing. Intel's strategy has been to push ASPs up while lowering costs(offering less) as they had no share left to gain.That's why Kaby Lake mobile is dual core centric.
They have a few quads, priced pretty high and at 45W and then lots of dual cores at crazy prices at lower TDP.
Raven Ridge can cover all price points and TDP ranges )12 to 45W) with quads.

With Cannon Lake Intel is highly likely to follow AMD and offer quads instead of dual as they can't just give the entire market to AMD.
Intel said 10nm will ship in volume in the first half of 2018, we'll see if they can do that in Q1 or at least before back to school or things go wrong with the process ramp.

On the GPU side hard to guess as we don't know much about Vega but right now their top laptop APU clocks the 512 cores to 900MHz at 35W.
AMD claimed a 40% increase in GPU perf at half the power.
So 37.5% more cores, a process shrink, an improved architecture and higher clocks.
512 cores at 900MHz is 0.92 TFLOPS at boost and RR needs to do 40% better at half the power. How much can it do at same power? Guess part of the gain is increase utilization not TFLOP gains.
The more efficient CPU cores should also give the GPU some more TDP room, up to a certain level of perf.
I don't quite dare to make a guess for boost clocks at 35-45W TDP with so little info but you cross 1.5TFLOPs at 1080MHz and i hope they can reach that at least.
Couple that with some utilization gains and Radeon Chill and it's not bad at all at 1080p.
EDIT: improved memory bandwidth utilization matters a lot from a TDP perspective too
EDIT 2: Just took this ss on Newegg, long live dual core laptops LOL.

 

[Atari2600]

Looks like no HBM2 for Raven Ridge, just as I predicted oh so long ago...

To be fair you did.

Very surprised (and somewhat crestfallen) AMD aren't doing at least one at the high end with HBM2. I'd have thought it the ideal weapon for accelerated compute in scenarios where customers couldn't justify or utilise racks of GPUs, but being able to pack into a 2 socket blade one socket having a conventional Naples and the other having a high end Zen+Vega APU utilising HBM.

Not to mention the obvious high end ultrabook etc.

 

[coercitiv]

Vs Kaby Lake, KbL is dual core centric while almost all (if not all) RR SKUs will be quad cores. RR will have a 4 cores CCX so every die has 4 cores and maybe they have some lower end SKUs with cores disabled. but they gain more financially if almost all are quads.
This is why RR will obliterate Kaby in laptop and there is no doubt about it. Will have much better perf (MT) and better power as you spread the load on more cores and gain efficiency.

With Coffee Lake Intel is highly likely to follow AMD and offer quads instead of dual as they can't just give the entire market to AMD.

AFAIK Intel will have 4c/8t 15W TDP mobile CPU in H2 2017. Whether they call it Kaby Lake Refresh or something else, we know they are low wattage quads built on 14nm++.

Computerbase reference
Sisoft Ranker

So, while I do agree a 4c/8t RR would have the upper hand against 2c/8t KBL, it won't be that easy after refresh hits. Intel claims 30% performance improvement over current 2c/4t CPUs.

The 30 percent boost came in one benchmark—SYSmark 2014 version 1.5—and applies to 15W U-series mobile processors. The comparison pits an i7-7500U (2.7GHz base, 3.5GHz turbo) with two cores and four threads against an unnamed next generation chip. The new chip has an unspecified base clockspeed, a 4GHz turbo, and doubles the number of cores and threads to four and eight. The 8th generation chip is built on a refined iteration of Intel's 14nm process.

source: ars technica

 

[mtcn77]

They'll eventually do 20000 of those for the fastforward 2 exascale project, as CTO Mr. Mark Papermaster said it best: "The goal of the research is eventual commercial applications, so the IP created as part of this research will make its way into various future AMD products."
I'm okay with AMD's semi-custom projects. Each step of the way, the console timeline shifted 2 years forward. It will eventually pay off to make HBM-Zen.

 

[Valantar]

Very surprised (and somewhat crestfallen) AMD aren't doing at least one at the high end with HBM2. I'd have thought it the ideal weapon for accelerated compute in scenarios where customers couldn't justify or utilise racks of GPUs, but being able to pack into a 2 socket blade one socket having a conventional Naples and the other having a high end Zen+Vega APU utilising HBM.
Not to mention the obvious high end ultrabook etc.

That doesn't make much sense. Raven Ridge is Raven Ridge, not a server/workstation APU. If they make a compute-centric APU for server/multi socket use, I would be pretty shocked if it was a direct die transplant of RR. Why not just use a 2-zeppelin Naples implementation with a separate GPU die connected over IF? Do you think RR will be able to scale to 125W+ with decent efficiency? Because if it can't, those servers are leaving a lot of potential performance on the table.

Now, do I want a high-end Raven Ridge chip with HBM? YES. But geared for mobile and SFF desktop <95W, not as some odd duck kinda-mobile-kinda-server-not-really-good-at-anything weirdo chip. No thanks.

They'll eventually do 20000 of those

I might be missing something here, but what are "those"?

 

[mtcn77]

I might be missing something here, but what are "those"?

Exascale APU's. Each come with 10 teraflops according to the draft. They will be made 100,000 in total, sorry. I mistook the power budget(20,000 kW for the whole project).

 

[imported_jjj]

AFAIK Intel will have 4c/8t 15W TDP mobile CPU in H2 2017. Whether they call it Kaby Lake Refresh or something else, we know they are low wattage quads built on 14nm++.

Computerbase reference
Sisoft Ranker

So, while I do agree a 4c/8t RR would have the upper hand against 2c/8t KBL, it won't be that easy after refresh hits. Intel claims 30% performance improvement over current 2c/4t CPUs.

source: ars technica

So you are claiming a Coffee Lake 14nm refresh in August or later in mobile followed by 10 nm Cannon Lake just 3-6 months later? How does that make sense.
Intel's 30+% claim is vs a quad but a reasonable assumption would be that the quad is on 10nm and that's next year, as soon as they can deliver it.
If there is a laptop Coffee Lake, that would imply further delays for 10nm so a rather big discovery and positive news for RR. Or a split under 10W for 10nm and 14 nm above

EDIT: I was assuming 14nm refresh in desktop only but if 10nm only addresses very low power, that's not good news for the 10nm ramp and AMD gets some extra time. AMD's very low power SKUs are likely one year away so 10nm would not impact them much at all.
A 14nm refresh for Intel would reshape the lineup to better compete but hit them on costs and if they take that hit, 10nm is in more troubles than thought.

 

[jpiniero]

So you are claiming a Coffee Lake 14nm refresh in August or later in mobile followed by 10 nm Cannon Lake just 3-6 months later? How does that make sense.

Intel has cut the scope of Cannonlake to only one DC U/Y model; and at this point it's probably just going to be Y. So you'd have the QC U Coffee Lake and the DC Y Cannonlake.

 

[imported_jjj]

Intel has cut the scope of Cannonlake to only one DC U/Y model; and at this point it's probably just going to be Y. So you'd have the QC U Coffee Lake and the DC Y Cannonlake.

It's possible and that shows troubles.
For Raven Ridge the uncore power will matter a lot, if they do great there , it can buy some extra TDP for the cores.They have a lot less I/O than on Summit Ridge and great SMT.. we'll see.

EDIT: If it goes this way, Intel will try to shift the market to 5W but then they are exposed to ARM solutions.They are running out of room to maneuver.

 

[Atari2600]

That doesn't make much sense. Raven Ridge is Raven Ridge, not a server/workstation APU. If they make a compute-centric APU for server/multi socket use, I would be pretty shocked if it was a direct die transplant of RR. Why not just use a 2-zeppelin Naples implementation with a separate GPU die connected over IF? Do you think RR will be able to scale to 125W+ with decent efficiency? Because if it can't, those servers are leaving a lot of potential performance on the table.

2 modules.
module1 = 2x Zeppelin
module2 = 1xCCX + Vega GU

Those two modules go onto one package along with HBM. Module2 is essentially RavenRidge with slightly different interfacing. Module2 can be repackaged standalone without the accompanying Zeppelin.

In the other socket on the motherboard is a conventional Naples (which is already an MCM).

 

[leoneazzurro]

That doesn't make much sense. Raven Ridge is Raven Ridge, not a server/workstation APU. If they make a compute-centric APU for server/multi socket use, I would be pretty shocked if it was a direct die transplant of RR. Why not just use a 2-zeppelin Naples implementation with a separate GPU die connected over IF? Do you think RR will be able to scale to 125W+ with decent efficiency? Because if it can't, those servers are leaving a lot of potential performance on the table.

Now, do I want a high-end Raven Ridge chip with HBM? YES. But geared for mobile and SFF desktop <95W, not as some odd duck kinda-mobile-kinda-server-not-really-good-at-anything weirdo chip. No thanks.

I might be missing something here, but what are "those"?

From the cost/benefits point of view it makes little sense (maybe only in High-end Ultrabooks) to pair a RR with HBM. Costs can be higher than a discrete low-end GPU solution with no advantage in performance. It may have more sense to couple a full Ryzen CPU with a mainstream Vega and one HBM2 stack dedicated to SFF PCs, i.e. for All-in-one or Apple.

 

[imported_jjj]

2 modules.
module1 = 2x Zeppelin
module2 = 1xCCX + Vega GU

Those two modules go onto one package along with HBM. Module2 is essentially RavenRidge with slightly different interfacing. Module2 can be repackaged standalone without the accompanying Zeppelin.

In the other socket on the motherboard is a conventional Naples (which is already an MCM).

What would be the point in adding a low perf GPU and pairing it with HBM? What market does it serve and why does it need HBM?
To MCM a mobile Vega 10 (interposer only for the GPU+HBM) and 8-16 CPU cores might make some sense. No idea if there are Z height issues with HBM and the Naples socket.
And if we are going MCM bonanza, why not go for 4xRR in a Naples socket with just the normal 8 channel memory? OFC assuming RR has GMI but that seems wasteful.