Current state of AMD APU landscape?

[leoneazzurro]

There is some improved power management in BR, which might allow for somewhat higher CPU and GPU clocks, than with CZ.


One poster says: same number, the other says: more CUs.... ^^
BTW, there is DDR4 instead of DDR3.

I meant, if they will put more CU maybe they'll end anyway being limited by the RAM bandwidth. Yes, it is DDR4, maybe even dual channel but it is nothing like having even a 128 bit GDDR5 like in the lowest-end Radeon based card, the RX460. And it's shared with the CPU.

 

[Dresdenboy]

Which of course makes no (positive) difference performance wise, since it cannot be clocked any further than DDR3 (without increasing the BCLK).
The increased latency of DDR4 meanwhile makes the already abysmal memory latency on Excavator even worse. It is already so high on DDR3 that it kills the performance in certain workloads (can be seen in Anand review too).

That's true. It's definitely time for new memory technologies here.

I meant, if they will put more CU maybe they'll end anyway being limited by the RAM bandwidth. Yes, it is DDR4, maybe even dual channel but it is nothing like having even a 128 bit GDDR5 like in the lowest-end Radeon based card, the RX460. And it's shared with the CPU.

They could have upped the L2 cache in the GPU, but that didn't happen. Anything else is even more cost driven. But would GDDR5 fit the CPU use cases?

 

[leoneazzurro]

They could have upped the L2 cache in the GPU, but that didn't happen. Anything else is even more cost driven. But would GDDR5 fit the CPU use cases?

Graphical workloads are more often than not limited by the bandwidth, also because cache costs die size (especially if it has to be "wired" to a large bus) and for being effective with such a low (with respect to the workload, of course) RAM bandwidth it should be quite larger than now. Anyway, mine was a mere observation, not a criticism. If they had put the same CU number as in the 460 being so BW limited, they will probably have gained maybe a few percent performance but at an higher cost due to die size increase. And also going for i.e. quad channel RAM is not a viable solution for budget systems. So definitely I believe they did the right thing and that yes, for future, if there is need for high performing (in the graphical sense) APUs they'll need to improve the memory interconnection. So far, I think that only EDRAM and HBM or similar technlogies are viable but they will also come at a cost.
So, maybe in the future we'll see different APU lines, a budget one with only DDR4 and what follows, and another with better interconnects, maybe with 460-like performance, which could be very good for laptops... Maybe.

 

[The Stilt]

GDDR5 should work pretty well as general purpose DRAM on APUs. The latency seems to be comparable to DDR4 at the same clocks. Streamroller has exclusive 2x2 GDDR5 and 2x DDR3 controllers in the die after all.

 

[SpaceBeer]

I think there is enough space to place Polaris 11 in 4-core Zen based APU, considering its die size, and size of current R7 IGPs. But there is no chance to see such a chip with HBM/GDDR5, since it would be to expensive for mid range segment, and not powerful enough for high-end

But even if they put Polaris 11 in Raven Ridge, with better memory controller and bandwidth (let's say it could fully utilize DDR4 at 3000/3200 MHz), it will still be much slower than RX 460, maybe similar to R7 260(X). Though even then it would still be good for APUs, especially in laptops and ITX/HTPC systems

 

[Sweepr]

This is the current state of AMD's APUs where it matters the most, power limited mobile products:

Kaby Lake-U i7-7500U (15W) vs Bristol Ridge FX-9800P (15W)

Cinebench R11.5 - CPU Single 64Bit
FX-9800P: 0.91
i7-7500U: 1.68

Cinebench R11.5 - CPU Multi 64Bit
FX-9800P: 2.86
i7-7500U: 4.03

Cinebench R15 - CPU Single 64Bit
FX-9800P: 74
i7-7500U: 145

Cinebench R15 - CPU Multi 64Bit
FX-9800P: 236
i7-7500U: 363

Geekbench 3 - 64 Bit Single-Core Score
FX-9800P: 2057
i7-7500U: 3706

Geekbench 3 - 64 Bit Multi-Core Score
FX-9800P: 5496
i7-7500U: 7835

3DMark 11 - Performance 1280x720
FX-9800P: 1794
i7-7500U: 1864

3DMark 11 - Performance GPU 1280x720
FX-9800P: 1852
i7-7500U: 1668

https://forums.anandtech.com/threads/intel-skylake-kaby-lake-thread.2428363/page-302#post-38456429

 

[Abwx]

This is the current state of AMD's APUs where it matters the most, power limited mobile products:

Kaby Lake-U i7-7500U (15W) vs Bristol Ridge FX-9800P (15W)



Cinebench R15 - CPU Single 64Bit
FX-9800P: 74

The KBL is set at 25W as i posted it in the relevant thread, as for NBC they obviously rigged the numbers, as usual when it s AMD in this site, prove is the score that i bolded, that is the CB R15 ST score, wich is supposed to be at 3.6GHz...

So they reduced the score by 18% or so, surely to please a given sponsor, indeed since they massacred AMD products and even some brand that sold an AMD based laptop they are no more delivered any AMD equipped laptop neither by AMD nor by any OEM, they got the last ones from German retailer Notebooksbilliger who will soon realise, if it s not already the case, that he s feeding some shills nest..

 

[Sweepr]

The KBL is set at 25W as i posted it in the relevant thread, as for NBC they obviously rigged the numbers, as usual when it s AMD in this site, prove is the score that i bolded, that is the CB R15 ST score, wich is supposed to be at 3.6GHz...

So they reduced the score by 18% or so, surely to please a given sponsor..

Stop spreading false information. After a period of maxing out PL2, it must go below PL1 for a while to stay within the PL1 envelope. Here's a hint from the pic you used as 'proof' in the other thread, but actually shows the opposite: CPU Package Power = 14.898W

Until you provide data that supports the idea that Kaby Lake-U can sustain a 25W PL2 state in all benchmarks (and during their entire run) - don't bother to reply. Just accept that Kaby Lake-U is a better mobile product than their current Bristol Ridge products at these power levels.

 

[cbn]

Virtual Larry,

Regarding APU vs. dGPU power consumption and efficiency I did recently buy a A8-7670K APU...so what I will do is compare idle on Athlon x 4 860K + GT730 GDDR5, Athlon x 4 860K + GT710 and A8-7670k (All using the same motherboard, PSU, RAM, storage, etc.)

I had problems with my A8-7670K/board combo, but was able to source a A10-7860K board combo for results:

https://forums.anandtech.com/thread...x-4-860k-dgpu-and-apus.2488541/#post-38518300

(Keep in mind the A10-7860K is a 65W APU as compared to the A8-7670K being a 95W APU)

 

[DrMrLordX]

Surprised I missed this topic.

The 7860k was the best deal in APUs unless you scummed eBay like I did for a ~$70 7870k. The 7600 is old hat and costs too much now, just as it did a month ago. Not worth the trouble there.

Anyway the 7860k is a downbinned 7870k that throttles to 2.5 GHz during iGPU usage (meaning get one of The Stilt's GeAPM-disabling UEFIs/BIOSes). Otherwise it's a full 512 shader iGPU + GV-A1 chip that can pretty much clock where you want it, fully unlocked etc. It's 65W only due to the base clocks and low throttle points.

With Bristol Ridge out there I wouldn't seriously consider it unless you're looking for a drop-in replacement in an FM2+ system.

Please ignore pointless posts about HEVC encode/decode since you aren't getting anything that modern anyway, and please ignore useless diversions regarding mobile APU performance when we all know you're getting a desktop part.

 

[cbn]

Heads up to anyone considering A10-7860K, according to reviews it comes with this cooler (shown below).

But mine came with the conventional (non-heatpipe) 58mm tall copper core 95W Quiet cooler (Same cooler as found on the Athlon x4 845).

 

[The Stilt]

They are both HS65 (95W) class heatsinks, from different supplier. The one with heatpipes is supplied by Coolermaster and the solid aluminum one is from AVC. You can get either, depending on the batch.