AMD Raven Ridge 'Zen APU' Thread

[Glo.]

Now Im starting to really hate AMD for not saying a word about even possibility of using HBM on those APUs, and reserving full dies for this chip.

4C/8T, 12 CU's(768 GCN cores), 2 GB's of HBM with lets say 160 GB/s is enough for low-end computers, and this chip in 65-95W Thermal envelope would have performance of GTX 1050 Ti if clocked properly(1.2-1.3 GHz).

 

[CatMerc]

Score differences between 2500U and 2700U. 11 CUs is 704 GCN cores. 8 CUs - 512 Cores. We are looking at around 40% increase in core count, higher clock speeds potentially, and only 400 pts in 3dMark GPU score difference(10%). Its too low score difference between them to indicate its bigger, full part.

And yes, the naming scheme U suggest that it is really at best 35W TDP part. So it might be pretty nice chip, after all.

It could entirely be bandwidth constraints rearing their ugly heads.

 

[Qwertilot]

Don't hate - I'm sure they'd like to do something like this but just don't have the R&D.

 

[LightningZ71]

From looking at the slides that AMD released, BAnded Kestrel and Horned Owl were basically the raven ridge die packaged as integrated SoC parts. The specs were a complete match to a full fat RR chip for HO, and BK was just cut down from that it seemed. For how tight the margins must be in that space, I would not expect AMD to do anything else but recover faulty RR dies in the BK market. With so much of the die disabled for that market (BK), with modest clocks, power draw would be dramatically lower.

 

[Jan Olšan]

Partially disabled/harvested dies don't get their own codenames. Note that Banded Kestrel was supposed to support LPDDR4, Raven Ridge not, in those early embedded slides.
Pretty sure it would have been a separate chip, because what do you replace the Bristol Ridge lowend with? And Stoney Ridge BGA chips for cheap laptops?
Both Raven Ridge and Summit Ridge are over 200mm2, that is insane to use in 50-100$ MSRP price segment - you want some actually cheap-to-make chip to sell there. They made extra Excavator implementation just for that market, surely it is economically sound do make a Zen for it, when Zen should sell better than XV.

You think they don't need something new to offer against the G4560s and the like? Most people disagree with that one.

 

[raghu78]

Now Im starting to really hate AMD for not saying a word about even possibility of using HBM on those APUs, and reserving full dies for this chip.

4C/8T, 12 CU's(768 GCN cores), 2 GB's of HBM with lets say 160 GB/s is enough for low-end computers, and this chip in 65-95W Thermal envelope would have performance of GTX 1050 Ti if clocked properly(1.2-1.3 GHz).

I don't think we will see HBM in APUs until 7nm APUs launch in 2020. The problem is that High Bandwidth Memory technology and packaging technology is not yet mature and cost effective for mass market products like APUs. I do think by 2020 memory and packaging technology will be mature from a yield perspective and there should be cost effective packaging technologies available.

 

[Abwx]

Take this with a grain of salt, R7 2700U on 3DMARK11

Comparison

Source

The CPU score suggest that the CPU test is at 2GHz or so, that s what is required to keep on a 15W TDP, FTR a 1C/2T Zen core use about 2.3-2.5W/2GHz, with 4 cores and the uncore this make about for this power range.

https://www.futuremark.com/hardware/cpu/AMD+Ryzen+5+1400/review

Also Carrizo s 1792pts GPU score is at 15W (IrisPro is at 30W+..) and 14nm LPP is 2x the efficency of GF s own 28nm, this and the 40% higher SP count should yield theoricaly 2.8x the GPU score within the same 15W, or 2x at same CU count, and of course at same frequencies in all cases.

https://images.techhive.com/images/article/2015/06/amd-carrizo-gaming-100588759-orig.png

 

[PeterScott]

Partially disabled/harvested dies don't get their own codenames. Note that Banded Kestrel was supposed to support LPDDR4, Raven Ridge not, in those early embedded slides.
Pretty sure it would have been a separate chip, because what do you replace the Bristol Ridge lowend with? And Stoney Ridge BGA chips for cheap laptops?
Both Raven Ridge and Summit Ridge are over 200mm2, that is insane to use in 50-100$ MSRP price segment - you want some actually cheap-to-make chip to sell there. They made extra Excavator implementation just for that market, surely it is economically sound do make a Zen for it, when Zen should sell better than XV.

You think they don't need something new to offer against the G4560s and the like? Most people disagree with that one.

G4560s are not low power laptop parts. I thought the Argument for BK (Banded Kestrel) was Ultra Low power tablet part? If BK exists it certainly won't be a G4560 competitor.

BK hasn't been seen on a roadmap since early 2016. There is no sign whatsoever that the idea was carried forward.

Thanks largely to AMD, the market for 2 cores is dwindling rapidly. I think AMD realized this, so they won't bother taping out a 2 core part. The savings of 2 cores worth of dies space is not going to be that large, especially going forward at 7nm.

Using one tapeout with disabled cores, fits with AMD strategy for over a decade. It will continue for Raven Ridge.

 

[DrMrLordX]

Intel's Core M for tablets is dual core and Zen dual core will scale from 4w tablets to 12w ultrathins. There is a huge market for good enough performance with emphasis on portability and battery life. If all you do is run MS Office, Adobe, IE,Chrome and watch video on your PC (which is what the majority of users do) you don't need 4 cores with HT. A dual core with HT is more than enough.

It's a nice thought, but I have the suspicion that A11x will do a real number on Zen in that power envelope. At that point it's only the software ecosystem/form factors preventing A11x from completely dominating the field. Well that and Apple's boutique pricing. Feh.

 

[Yotsugi]

A11x will do a real number on Zen in that power envelope

If everything you do is running Geekbench, then yes.

 

[DrMrLordX]

If everything you do is running Geekbench, then yes.

Laugh all you want. I would not want to be AMD in that fight. There are probably some areas where Apple has yet to optimize for more desktop-like workloads, but overall, I think a11x would handle itself very well. Still not really sure how Apple's SIMD implementation (uh, NEON?) stacks up against stuff like AVX/ADX on Zen. But in that power envelope, I don't know that SIMD will be that big of a deal.

 

[CatMerc]

The CPU score suggest that the CPU test is at 2GHz or so, that s what is required to keep on a 15W TDP, FTR a 1C/2T Zen core use about 2.3-2.5W/2GHz, with 4 cores and the uncore this make about for this power range.

https://www.futuremark.com/hardware/cpu/AMD+Ryzen+5+1400/review

Also Carrizo s 1792pts GPU score is at 15W (IrisPro is at 30W+..) and 14nm LPP is 2x the efficency of GF s own 28nm, this and the 40% higher SP count should yield theoricaly 2.8x the GPU score within the same 15W, or 2x at same CU count, and of course at same frequencies in all cases.

https://images.techhive.com/images/article/2015/06/amd-carrizo-gaming-100588759-orig.png

According to this naming scheme, U is standard mobile, while M is low power and H is high performance mobile.
Raven Ridge is supposed to span 15W-45W, so I imagine the U part is between 25W-35W.

Laugh all you want. I would not want to be AMD in that fight. There are probably some areas where Apple has yet to optimize for more desktop-like workloads, but overall, I think a11x would handle itself very well. Still not really sure how Apple's SIMD implementation (uh, NEON?) stacks up against stuff like AVX/ADX on Zen. But in that power envelope, I don't know that SIMD will be that big of a deal.

Until we see both A11 running on the same OS and software stack as other SoC's, we can't really tell performance in apples to apples comparison. The OS makes a huge difference in tests.
A large part of Apple's SoC's strength is that they're running a top to bottom show, meaning their software stack is incredibly optimized for their hardware, and their hardware is incredibly optimized for their software stack.

 

[Yotsugi]

U is standard mobile

Sounds 15W tbh.

while M is low power

Sounds like Core m competitor, i.e. 4.5W SKUs.

H is high performance mobile

This could be anything 28W-onwards.

 

[CatMerc]

Sounds 15W tbh.

Sounds like Core m competitor, i.e. 4.5W SKUs.

This could be anything 28W-onwards.

Would be true had Raven Ridge been intended for 4.5W. That's Banded Kestrel (the dual core) territory.

 

[Abwx]

According to this naming scheme, U is standard mobile, while M is low power and H is high performance mobile.
Raven Ridge is supposed to span 15W-45W, so I imagine the U part is between 25W-35W.


.

Doesnt make sense, if U is 25-35W it wouldnt be competitive at all since we re talking of 2GHz, and then what would be high performance, 65W..?

FTR i mentioned what is a Zen core power comsumption, at 3.2GHz it use 7.25W when loaded with two threads and power/frequency will decrease at a 2.56 polynomial rate (wich is due to the characteristics of GF s finfets), this can be seen on an AT s Ryzen review, the uncore will be the same as the one in Bristol Ridge and its proportion in the total TDP (at 15W) should be much lower than in Summit Ridge, somewhere like 20-30% of the chip total TDP at full CPU load.

Besides IIRC RR is designed for TDPs as low as 4W...

https://www.kitguru.net/components/...d-cpu-roadmap-appears-online-zen-coming-2017/

 

[Yotsugi]

Would be true had Raven Ridge been intended for 4.5W. That's Banded Kestrel (the dual core) territory.

Different dies that still fit into AMD's naming scheme.

 

[CatMerc]

Doesnt make sense, if U is 25-35W it wouldnt be competitive at all since we re talking of 2GHz, and then what would be high performance, 65W..?

FTR i mentioned what is a Zen core power comsumption, at 3.2GHz it use 7.25W when loaded with two threads and power/frequency will decrease at a 2.56 polynomial rate (wich is due to the characteristics of GF s finfets), this can be seen on an AT s Ryzen review, the uncore will be the same as the one in Bristol Ridge and its proportion in the total TDP (at 15W) should be much lower than in Summit Ridge, somewhere like 20-30% of the chip total TDP at full CPU load.

I'm not particularly keen on such math, as it tends to rarely be accurate enough for such predictions.

Different dies that still fit into AMD's naming scheme.

Here's how I see it:
Raven Ridge H - 35W-45W
Raven Ridge U - 25W-35W
Raven Ridge M - 15W
Banded Kestrel H - 15W
Banded Kestrel U - 7.5W
Banded Kestrel M - 4.5W

 

[Yotsugi]

Raven Ridge H - 35W-45W
Raven Ridge U - 25W-35W
Raven Ridge M - 15W
Banded Kestrel H - 15W
Banded Kestrel U - 7.5W
Banded Kestrel M - 4.5W

This is retarded and makes no sense.
15W 2/4? Current year?
Besides, 2.0/3.6 25W SKU?
Makes even less sense.

 

[Abwx]

I'm not particularly keen on such math, as it tends to rarely be accurate enough for such predictions.


Here's how I see it:
Raven Ridge H - 35W-45W
Raven Ridge U - 25W-35W
Raven Ridge M - 15W
Banded Kestrel H - 15W
Banded Kestrel U - 7.5W
Banded Kestrel M - 4.5W

I edited my post with a link, RR will be 4-35W, a 35W-45W SKU doesnt make sense since that s a range that is too limited, anyway AMD didnt release anything mobile that was more than 35W since 2011...

As for the maths they are based on measurements, not on predictions...

 

[CatMerc]

I edited my post with a link, RR will be 4-35W, a 35W-45W SKU doesnt make sense since that s a range that is too limited, anyway AMD didnt release anything mobile that was more than 35W since 2011...

As for the maths they are based on measurements, not on predictions...

Raven's TDP range varies quite a bit depending on source. I don't think AMD nailed it down completely until a few months ago.
Different slides say different things.

 

[Dayman1225]

15W 2/4? Current year?

I don't see the problem with a 15w 2/4 part if it has 1. Higher boosts 2. Holds those boosts for longer. IF they do not they don't make sense. Oh an 3. Are substantially cheaper

Besides, 2.0/3.6 25W SKU?

Its not just the CPU cores taken into count here, the Vega Graphics 8/11 CUs from I last heard, would likely take some power. Either way it would still be a lot more efficient than Intel's IRIS PRO parts. So I don's see the issue with his "prediction"

15w for 2.0/3.6 4c/8t 8/11CU parts is optimistic IMO, but not out of the realm of possibility, and if they hit that, thats very good

 

[CatMerc]

Yeah, as Dayman said, that kind of performance in 15W sounds optimistic. Even at 35W it would be good.

If we get that in 15W, that's just absolutely incredible. Not outside the realm of possibility, but not the cautious "predicition" either.

 

[Abwx]

Raven's TDP range varies quite a bit depending on source. I don't think AMD nailed it down completely until a few months ago.
Different slides say different things.

Surely, but if say 25W were necessary for 4C/8T at 2GHz then a chip like the R7 1700 would be at least 125W.

Think about it, 8C/16T@3GHz fit within 65W, and half this core amount would require 25W@2GHz.?..

 

[Yotsugi]

1. Higher boosts

2500U peak clock already sits close to 14LPP cap.

2. Holds those boosts for longer.

They all will hold these single core turbos for about the same amount of time.

Are substantially cheaper

Possibly, though makes no sense with AMD pushing moar corez into every corner of the market.

the Vega Graphics 8/11 CUs

These will sip power since they are clocked very conservatively.
15W SKU it is. Besides, cTDP exists.

If we get that in 15W, that's just absolutely incredible.

What's special about that? Zen and Vega are both incredibly efficient when not overvolted trying to reach higher clocks.

 

[Dayman1225]

2500U peak clock already sits close to 14LPP cap.

I wasn't just referring to a single turbo burst, I meant higher turbo steps too.

They all will hold these single core turbos for about the same amount of time.

We don't really know this, the products aren't even out yet..

Your other points are completely valid, we'll just have to wait and see!