AMD Raven Ridge 'Zen APU' Thread

[NTMBK]

It's very frustrating, because AMD obviously can scale up their APU tech if they want to- the current crop of games consoles proves it. The problem is the memory bus, every single time.

Just look at the XBox One S motherboard. This is a console which used quad-channel DDR3-2133 (and was still hampered with insufficient bandwidth):

Look at that area around the APU, between the processor and the memory chips. That isn't dead space- that is all the traces required for a quad-channel bus. That adds a heck of a lot of complexity, cost, and area to a motherboard. And that's not even with socketed DIMMs/SODIMMs, which would make the situation even worse!

I'm sad that the APU support for GDDR5 never came to anything. Raven Ridge in a laptop with soldered down GDDR5 could actually be a very good solution for gaming.

 

[Qwertilot]

That's no R&D money for you Look how much Intel are asking for the skull trail stuff say - there's definitely a market for this kind of thing and it will stand quite a premium.

 

[SPBHM]

It's very frustrating, because AMD obviously can scale up their APU tech if they want to- the current crop of games consoles proves it. The problem is the memory bus, every single time.

Just look at the XBox One S motherboard. This is a console which used quad-channel DDR3-2133 (and was still hampered with insufficient bandwidth):

Look at that area around the APU, between the processor and the memory chips. That isn't dead space- that is all the traces required for a quad-channel bus. That adds a heck of a lot of complexity, cost, and area to a motherboard. And that's not even with socketed DIMMs/SODIMMs, which would make the situation even worse!

I'm sad that the APU support for GDDR5 never came to anything. Raven Ridge in a laptop with soldered down GDDR5 could actually be a very good solution for gaming.

it looks less dramatic with slots
quad DDR4:

I think triple channel with DDR4 3200 would be nice enough, you could probably have 1024SPs @ 900MHz or something and it would perform close enough to the RX 460, which is OK for 1080P medium-low gaming.

cost is higher no doubt, but, if stores can sell dual channel boards for less than $40...
well DDR4 is expensive, but 3x4GB is appealing compared to 2x8GB.

 

[nathanddrews]

Broken record time: anyone know if AMD is going to get support for Netflix, etc. for 4K streaming and UHD Blu-ray like Intel did for Kaby Lake? When Polaris launched, AMD discussed "adding support later" but they never did. Not a peep about it with Vega. So far, it's only Kaby Lake and Pascal (limited support).

This is the number one selling point of Raven Ridge for my use case (very small HTPC + light gaming).

 

[PeterScott]

it looks less dramatic with slots
quad DDR4:

http://www.overclock3d.net/gfx/articles/2015/08/24124227157l.jpg

I think triple channel with DDR4 3200 would be nice enough, you could probably have 1024SPs @ 900MHz or something and it would perform close enough to the RX 460, which is OK for 1080P medium-low gaming.

cost is higher no doubt, but, if stores can sell dual channel boards for less than $40...
well DDR4 is expensive, but 3x4GB is appealing compared to 2x8GB.

The XBox could spare the area, so the likely went with less circuit board layers, the quad channel ITX likely required more circuit board layers driving up cost that way.

The problem for RR is that most people only really care about the mainstream 2 channel design, and any kind of alternate design is going to AMD tens of millions of dollars to tape out, for a niche market. That might work if you could charge a BIG premium AND sell quite large numbers of them, but once you get to the people willing to pay a big premium for more powerful GPU, they are in the market for dGPU designs, and you have big competition that can likely do a better by having two independent devices.

RR will be a great product because it will maintain/expand AMDs lead over Intel iGP, while catching up on the CPU front and that will be enough for most people.

But unfortunately, I don't think serious GPU performance (say even GTX 1050 level) inside an APU/iGP is coming in the foreseeable future.

 

[Gideon]

2500U Geebench leak (@ 2Ghz). Multi threaded seems pretty decent , but single seems lackluster. Boost not working properly perhaps?:
http://www.guru3d.com/news-story/geekbench-results-for-raven-ridge-2500u.html

EDIT:
There actually seems to be a longer list of the results (Mandolin and Tambourine):
https://browser.geekbench.com/v4/cpu/search?utf8=✓&q=2500U
The best ST result is 3730 and MT: 12432 (not from the same run)
Supposedly the same 2500U

EDIT2:
Never mind, it seems the Tambourine (seems to be the mobo name) is running on Android 8.1 and is therefore not really comparable apples-to-apples. Interesting nevertheless

 

[psolord]

Scores are pretty close to a stock 2500k. Not bad at all.

 

[LTC8K6]

Hard to make anything of it with that reported clock speed.

 

[Bouowmx]

More single-core score than, but similar multi-core score as Intel Core i7-4702[HM]Q 2.9-3.2 GHz, 37 W. Doesn't compare to Intel Core i5-8250U 3.4 GHz, 15-44 W.

Does there exist an AMD equivalent to Turbo Boost Short Power Max?

 

[lixlax]

Compared to the official Ryzen 5 1400 result on their website, that CPU must have been running 3,3-3,4GHz for the single threaded benchmark and ~2,5GHz for tthe multithreaded.
And as compared to Intel, isn't Zen relatively weak in Geekbench anyway?

 

[LightningZ71]

Not necessarily for MT at 2.5Ghz. Remember that APUs become memory starved when the GPU starts demanding memory access. Those 8 active MT threads will be hitting that memory bus pretty hard, and as I recall, the display will still require constant memory bus traffic. This may be a situation where maximum system performance will actually be better with SMT disabled. The other weak link for APUs is thermal and power envelope. With 8 threads running full load, the CPU is going to be drawing a lot of power at once. The GPU will still be drawing its normal load on top of that. This could easily lead to frequency throttling to maintain the power and/or thermal envelope.

With SMT disabled, the cores will actually have momentary rest periods when there are memory stalls. This will mitigate thermal loading from the CPU side. Also, with memory access demands down to just 4 threads, there will be more memory bandwidth available for the GPU. I was initally disappointed that AMD had opted to have only 4MB L3 cache, but I think that upping it to 8MB would not help much as the threads will just be even more active and thermal loading will be that much worse. It would help for the GPU's memory needs, but if the CPUs are eating up a lot of heat and power capacity, it might not even be able to use that extra bandwidth.

I really wonder how it would perform on a good overclocking board that's got plenty of power. With high speed DDR4 ram and an open power budget and sufficient cooling, this could be quite an interesting part. But, then, you get into a money issue. If you're spending enough on a cooling solution, high end motherboards, and memory, why not just get a discrete GPU and less expensive base components? The only place where that may be interesting is SFF/USFF setups, and those tend to have thermal management issues and smaller power supplies.

 

[The Stilt]

Does there exist an AMD equivalent to Turbo Boost Short Power Max?

Sure, Core Performance Boost (CPB) and XFR.

I expect Raven to have rather respectable MSCXFRC (single core) frequency for it's rated TDP and relative to the base and MACXFRC (all core, TDP limited) frequencies.
Definitely not pushed as far as on Zeppelin, but respectable regardless.

Unless the halved L3 size (compared to Zeppelin) penalizes the performance, the single core frequency for that alleged Geekbench 2500U result (3561 points) has been =< 3.1GHz.

 

[Arachnotronic]

Sure, Core Performance Boost (CPB) and XFR.

I expect Raven to have rather respectable MSCXFRC (single core) frequency for it's rated TDP and relative to the base and MACXFRC (all core, TDP limited) frequencies.
Definitely not pushed as far as on Zeppelin, but respectable regardless.

Unless the halved L3 size (compared to Zeppelin) penalizes the performance, the single core frequency for that alleged Geekbench 2500U result (3561 points) has been =< 3.1GHz.

More like 3.4GHz. Zeppelin at ~4GHz scores around 4200 single core in GB4.

 

[The Stilt]

More like 3.4GHz. Zeppelin at ~4GHz scores around 4200 single core in GB4.

Do the scores between the build versions differ that significantly?

My ST score at fixed 3.6GHz speed is 4122 points.
The very same result I sent you on 3/3/2017: https://browser.geekbench.com/v4/cpu/1993360

 

[Arachnotronic]

Do the scores between the build versions differ that significantly?

My ST score at fixed 3.6GHz speed is 4122 points.
The very same result I sent you on 3/3/2017: https://browser.geekbench.com/v4/cpu/1993360

Very interesting. I just searched GB4's database for recent 1800X results that looked like they were at stock for my comment above.

Could you by any chance re-run the test at fixed speed with the latest version of GB4?

 

[VirtualLarry]

Scores are pretty close to a stock 2500k. Not bad at all.

Which is, in turn, nearly equivalent to a G4560. Not really sure how great that is, for a true quad-core APU based on Zen cores.

Edit: Oh, I guess the frequency was ~2Ghz? That's not too shabby then.

 

[Gideon]

An unconfirmed 3Dmark 11 score as well. Caution, could easily be fake.
https://www.reddit.com/r/Amd/comments/707xz1/3dmark_11_performance_amd_ryzen_5_2500u/

 

[NewFatMike]

Interesting that the GPU is still clocked at 300MHz

Sent from my Pixel using Tapatalk

 

[AtenRa]

And HD640 at 0 mhz

 

[NewFatMike]

HA! Totally missed that

Sent from my Pixel using Tapatalk

 

[The Stilt]

Interesting that the GPU is still clocked at 300MHz

Sent from my Pixel using Tapatalk

It isn't.
It lists the boot frequency.

 

[NewFatMike]

It isn't.
It lists the boot frequency.

Is that the case with some of the other benches we've seen as well?

Sent from my Pixel using Tapatalk

 

[The Stilt]

Is that the case with some of the other benches we've seen as well?

Sent from my Pixel using Tapatalk

Depends on the bench.

 

[NewFatMike]

Depends on the bench.

GFXbench dropped this:

CL_DEVICE_MAX_CLOCK_FREQUENCY - 300

On the OpenCL info dump

Sent from my Pixel using Tapatalk

 

[cbn]

I'm sad that the APU support for GDDR5 never came to anything. Raven Ridge in a laptop with soldered down GDDR5 could actually be a very good solution for gaming.

Or how about LPDDR4?

https://news.samsung.com/global/samsung-rolls-out-industrys-first-8gb-lpddr4-dram-package

At 4266 speed it is 2x as fast as DDR4 2133.....and it comes in 8GB packages.