AMD Raven Ridge 'Zen APU' Thread

[The Stilt]

And how performance is? 15w/35w/45w parts catch up with Intels 15w/35w/45w parts on CPU performance?

No idea, obviously.
I see no reason why they wouldn't be competitive, since Ryzen desktop parts most of the time are.

I'm not too optimistic about the iGPU side of the chip thou.
It's definitely faster than anything we've seen in the past, but the absolute performance is still totally murdered by the lack of bandwidth and partially by the power restrictions (which are always present on mobile platforms).
For laptops anything faster than DDR-2400 isn't really realistic, despite at least G.Skill has 2800MHz+ (?) rated SO-DIMMs available.

I would personally like to see AMD releasing a true mobile CPU die (i.e. a Raven without the iGPU) so that they could be paired with a nVidia Pascal based mobile GPU, without wasting any silicon or power on the CPU die itself.

Without a new memory technology which would ~double the bandwidth available for the iGPU, they will remain as jack of all trades but master of none.
Sure, a full Raven with it's 11CUs will offer decent performance (for an iGPU), but still they will be a total joke against even the most entry level dGPUs with their dedicated GDDR5 VRAM.

 

[SpaceBeer]

How large is mid/high-end mobile dGPU market? I mean, most of sub $800 notebooks have (had) 2c/4t ULV CPU and some crappy low-end dGPU, which isn't much faster than IGP

In my opinion, upcoming AMD's IGP are more than enough for this market ($400-$800 notebooks), and suitable for everything but (AAA) gaming. Regarding bandwidth, if it's improved in Vega arch, than ~40 GB/s shouldn't decrease performance (significantly). I was surprised to see GT 1030 is competitive with 48 GB/s. And I doubt new Intel's 4-core ULV i5/i7 plus nVidia's GP1080 are cheaper than single AMD's APU, even if it is ~240mm^2 chip.

As I said before, if AMD makes such PRO chip (certified for CAD) it could make big impact on (entry-level) mobile/SFF/AiO workstation market
https://www.youtube.com/watch?v=jcrH6l8JzVw
And of course it would be great for bunch of other tasks (for office users, students, developers...)

 

[ahimsa42]

No idea, obviously.
I see no reason why they wouldn't be competitive, since Ryzen desktop parts most of the time are.

I'm not too optimistic about the iGPU side of the chip thou.
It's definitely faster than anything we've seen in the past, but the absolute performance is still totally murdered by the lack of bandwidth and partially by the power restrictions (which are always present on mobile platforms).
For laptops anything faster than DDR-2400 isn't really realistic, despite at least G.Skill has 2800MHz+ (?) rated SO-DIMMs available.

I would personally like to see AMD releasing a true mobile CPU die (i.e. a Raven without the iGPU) so that they could be paired with a nVidia Pascal based mobile GPU, without wasting any silicon or power on the CPU die itself.

Without a new memory technology which would ~double the bandwidth available for the iGPU, they will remain as jack of all trades but master of none.
Sure, a full Raven with it's 11CUs will offer decent performance (for an iGPU), but still they will be a total joke against even the most entry level dGPUs with their dedicated GDDR5 VRAM.

would not LPDDR4 go a long way in solving the iGPU bandwidth limitation? any chance it will be utilized for RR and if not, why?

 

[The Stilt]

Regarding bandwidth, if it's improved in Vega arch, than ~40 GB/s shouldn't decrease performance (significantly).

Sure, the DCC in Raven is no doubt more efficient than the one featured in GCN Gen. 2-4.
However since the performance improvement from going 38.4GB/s to 80.0GB/s bandwidth was up to ~43.7% (the scaling didn't even stop there) on a ~887GFlops GCN Gen. 2 GPU, a DCC no matter how effective it is won't be able to remedy the lack of bandwidth.
And the fastest Raven SKUs (DT) are expected to have almost twice the grunt from the HW side...

 

[vinhom]

How much area would have been necessary to include a single GDDR5 controller together with the dual channel DDR4 controllers to provide bandwidth for the GPU? I always imagined that would have been the best solution. Laptop manufacturers would have the flexibility to offer laptops with better GPU performance just by adding a few more motherboard traces and some GDDR5 memory.

Would dual channel CPU memory be really as necessary as it is today if you had some GDDR5 to go with it?

 

[Bouowmx]

How large is mid/high-end mobile dGPU market? I mean, most of sub $800 notebooks have (had) 2c/4t ULV CPU and some crappy low-end dGPU, which isn't much faster than IGP

In my opinion, upcoming AMD's IGP are more than enough for this market ($400-$800 notebooks), and suitable for everything but (AAA) gaming. Regarding bandwidth, if it's improved in Vega arch, than ~40 GB/s shouldn't decrease performance (significantly). I was surprised to see GT 1030 is competitive with 48 GB/s. And I doubt new Intel's 4-core ULV i5/i7 plus nVidia's GP1080 are cheaper than single AMD's APU, even if it is ~240mm^2 chip.

As I said before, if AMD makes such PRO chip (certified for CAD) it could make big impact on (entry-level) mobile/SFF/AiO workstation market
https://www.youtube.com/watch?v=jcrH6l8JzVw
And of course it would be great for bunch of other tasks (for office users, students, developers...)

It really isn't hard to take a gander, Intel 4-core H-series and NVIDIA GeForce GTX 1050, sub-$800: https://www.newegg.com/Product/Prod...98&IsNodeId=1&bop=And&Order=PRICE&PageSize=96

AMD Raven Ridge is likely a suitable successor to $500-level laptops with NVIDIA GeForce 940MX, which for some reason has not been replaced with MX150 in price yet.

 

[blublub]

No idea, obviously.
I see no reason why they wouldn't be competitive, since Ryzen desktop parts most of the time are.

I'm not too optimistic about the iGPU side of the chip thou.
It's definitely faster than anything we've seen in the past, but the absolute performance is still totally murdered by the lack of bandwidth and partially by the power restrictions (which are always present on mobile platforms).
For laptops anything faster than DDR-2400 isn't really realistic, despite at least G.Skill has 2800MHz+ (?) rated SO-DIMMs available.

I would personally like to see AMD releasing a true mobile CPU die (i.e. a Raven without the iGPU) so that they could be paired with a nVidia Pascal based mobile GPU, without wasting any silicon or power on the CPU die itself.

Without a new memory technology which would ~double the bandwidth available for the iGPU, they will remain as jack of all trades but master of none.
Sure, a full Raven with it's 11CUs will offer decent performance (for an iGPU), but still they will be a total joke against even the most entry level dGPUs with their dedicated GDDR5 VRAM.

I think what u would like, and I like to have that too, is Raven Ridge with HBM video memory - on mobile with 1GB would enough with HBCC expansion to drr4

 

[raghu78]

I think what u would like, and I like to have that too, is Raven Ridge with HBM video memory - on mobile with 1GB would enough with HBCC expansion to drr4

HBM2 in an APU is not going to happen till 2020 atleast. I think the first 7nm APUs and AMD's 25x20 vision will coincide along with HBM on APUs. Right now HBM2 has serious cost, yield and packaging issues. I think we will see improvements in HBM2 yield , cost and packaging in the 2019-2020 timeframe. AMD has the IP to disrupt the low end discrete gpu market with Zen 2/Navi APU. Its a matter of memory and packaging technology maturing to be introduced in a high volume mainstream product at reasonable cost.

 

[cbn]

HBM2 in an APU is not going to happen till 2020 atleast. I think the first 7nm APUs and AMD's 25x20 vision will coincide along with HBM on APUs. Right now HBM2 has serious cost, yield and packaging issues. I think we will see improvements in HBM2 yield , cost and packaging in the 2019-2020 timeframe. AMD has the IP to disrupt the low end discrete gpu market with Zen 2/Navi APU. Its a matter of memory and packaging technology maturing to be introduced in a high volume mainstream product at reasonable cost.

But what if Raven Ridge has 1/2 rate DP floating point?---> http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=thread...s-today-as-a-completed-design-at-amd.2519054/

 

[IntelUser2000]

But what if Raven Ridge has 1/2 rate DP floating point?---> http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=thread...s-today-as-a-completed-design-at-amd.2519054/

It shouldn't. Half rate DP would mean unnecessary die size increase.

RX Vega with 600mm2 die size doesn't have it, why would an iGPU that's hampered by die size/power, and memory bandwidth have it?

 

[ao_ika_red]

But what if Raven Ridge has 1/2 rate DP floating point?---> http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=thread...s-today-as-a-completed-design-at-amd.2519054/

1/8 or 1/16 of FP64 should be more sensible to allow it fully utilizes its limited bw..

 

[ao_ika_red]

I think what u would like, and I like to have that too, is Raven Ridge with HBM video memory - on mobile with 1GB would enough with HBCC expansion to drr4

IIRC, AMD's original mission with their APU (Fusion) project is to provide an affordable dan powerful package to the mass with the available resources. That's why they're pushing HSA and its zero-copy concept so hard.
The only problem is, their GPU arm is not performing as expected. Their compression tech is lacking behind its competitor, hence the need of larger bandwidth.
By adding HBM, it will raise the cost considerably and (probably) turn people attention into the less powerful one (with standard DDR4 memory).

 

[cbn]

But what if Raven Ridge has 1/2 rate DP floating point?---> http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=thread...s-today-as-a-completed-design-at-amd.2519054/

It shouldn't. Half rate DP would mean unnecessary die size increase.

RX Vega with 600mm2 die size doesn't have it, why would an iGPU that's hampered by die size/power, and memory bandwidth have it?

Bristol Ridge APU actually does have 1/2 rate DP FP---> http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=threads/cpu-for-floats-crunching.2518734/#post-39077849

So I think having 1/2 rate DP FP on Raven Ridge is definitely possible....and with one stack of HBM2 (@ up to 256 GB/s) the 1/2 rate DP FP would actually be fully usable (unlike the 1/2 rate DP FP on Bristol Ridge).

 

[HurleyBird]

Raven Ridge + HBM would have uniquely high perf/w and perf/area and therefore would be able to command quite a premium.

There's definitely a market for it. Hard to say if AMD built in support or not.

 

[Qwertilot]

Surely very unclear about the perf/watt? The iGPU bit helps but they’re giving such a big handicap to NV to start with.....

Perf/area quite possibly.

 

[blublub]

IIRC, AMD's original mission with their APU (Fusion) project is to provide an affordable dan powerful package to the mass with the available resources. That's why they're pushing HSA and its zero-copy concept so hard.
The only problem is, their GPU arm is not performing as expected. Their compression tech is lacking behind its competitor, hence the need of larger bandwidth.
By adding HBM, it will raise the cost considerably and (probably) turn people attention into the less powerful one (with standard DDR4 memory).

But still an APU with 1GB or even 512MB HBM would likely outperform any availability APU currently available - and with HBCC more vram isn't an issue

For high-end notebooks ala Apple or ThinkPad and alike there would be a market

 

[PeterScott]

But still an APU with 1GB or even 512MB HBM would likely outperform any availability APU currently available - and with HBCC more vram isn't an issue

For high-end notebooks ala Apple or ThinkPad and alike there would be a market

It would likely raise the price so much that a dGPU would be similar priced and better performing.

 

[The Stilt]

Like the good folks at SA pointed out, my "info" about the RV die size was wrong.
The real size is 19.x x 11.x mm, i.e anything between 209mm² or so.

As I said I don't have a chip available, so I can't be entirely sure about it.

Those measurements seem to match quite well with the picture thou: https://www.purepc.pl/notebooki/procesor_amd_ryzen_mobile_uchwycony_na_zdjeciach

 

[LightningZ71]

I still think that RR finally having an L3 will be the single biggest improvement over BR, followed closely by the improved cpu cores and additional available threads. As has been demonstrated in the DC community, bandwidth contention between the CPUs and the iGPU absolutely slaughters performance once saturation hits. Having even a 4MB L3 should be enough to keep the CPUs reasonably happy while the GPU is leaning hard on the DRAM. An 8MB L3 would be even better of course.

And, no, I still don't think that RR will be a substitute for a decent dGPU, it will definitely obsolete most of the low end ones. It can only get better as companies develop faster SODIMMs for DDR4. Really, what could the performance of an R7 RR with 8MB L3 and DDR4-3800 dual channel look like? I can't imagine that it would be slower than an RX550 or Nv 1030 in most any case.

 

[PeterScott]

And, no, I still don't think that RR will be a substitute for a decent dGPU, it will definitely obsolete most of the low end ones. It can only get better as companies develop faster SODIMMs for DDR4. Really, what could the performance of an R7 RR with 8MB L3 and DDR4-3800 dual channel look like? I can't imagine that it would be slower than an RX550 or Nv 1030 in most any case.

I expect it will be slower than an RX550. The RX550 has 128 bit GDDR5 with 112 GB/S of bandwidth, almost double dual channel DDR4-3800. Also, RR is a laptop chip. It will be running DDR4 2400 at best.

If the RX550 could have got by with half the memory bandwidth, AMD would have given it a 64 bit bus, and built a less expensive card.

I expect RR will slot in under the RX550/GT 1030.

 

[maddie]

I expect it will be slower than an RX550. The RX550 has 128 bit GDDR5 with 112 GB/S of bandwidth, almost double dual channel DDR4-3800. Also, RR is a laptop chip. It will be running DDR4 2400 at best.

If the RX550 could have got by with half the memory bandwidth, AMD would have given it a 64 bit bus, and built a less expensive card.

I expect RR will slot in under the RX550/GT 1030.

Not saying that there's a 100% improvement in compression but you are drawing conclusions between Polaris and Vega IP without proper analysis.

 

[PeterScott]

Not saying that there's a 100% improvement in compression but you are drawing conclusions between Polaris and Vega IP without proper analysis.

Clock for clock including matched memory bandwidth. Vega is about 5% faster than Fury X, so I am not seeing a huge boost in memory BW efficiency.
https://www.hardocp.com/article/2017/09/12/radeon_rx_vega_64_vs_r9_fury_x_clock_for/15

Especially given that Fury X was said to be sensitive to memory bandwidth.

Like most things surrounding Vega, the memory BW efficiency improvements also seem to be largely exaggerated.

 

[cbn]

If the RX550 could have got by with half the memory bandwidth, AMD would have given it a 64 bit bus, and built a less expensive card.

I think the reason AMD gave Polaris 12 a 128 bit bus (the same bus width as the twice as large Polaris 11) was so that it could eventually use DDR4.

Otherwise, it (Polaris 12) could have gotten a 64 bit bus because it has half the stream processors of the 128 bit bus Polaris 11.

 

[AtenRa]

And, no, I still don't think that RR will be a substitute for a decent dGPU, it will definitely obsolete most of the low end ones. It can only get better as companies develop faster SODIMMs for DDR4. Really, what could the performance of an R7 RR with 8MB L3 and DDR4-3800 dual channel look like? I can't imagine that it would be slower than an RX550 or Nv 1030 in most any case.

With 3800MHz memory on 128bit it will have a theoretical maximum memory bandwidth of 60,8GB/sec, in comparison GT1030 only has 48GB/sec.
Im expecting RR at 65W TDP with 11CUs and 3200MHz+ DDR-4 memory to come very close to GT1030.

 

[PeterScott]

With 3800MHz memory on 128bit it will have a theoretical maximum memory bandwidth of 60,8GB/sec, in comparison GT1030 only has 48GB/sec.
Im expecting RR at 65W TDP with 11CUs and 3200MHz+ DDR-4 memory to come very close to GT1030.

NVidia is MUCH more efficient with memory BW. I think this is what drove AMD to HBM first.

GT1030 is a match for the RX550 despite having less than half the memory BW.

Again, how many RR will be paired with even DDR4-3200? ZERO from OEMs. RR will be an OEM chip well into next year.

Eventually it will be available for enthusiasts doing self builds. But they aren't likely to lavish much attention on tweaking a 4 core APU to chase the lowest of low end GPUs. They will just get an inexpensive GPU if they care about GPU performance at all.