Low-cost emerging market SoC/phone discussion

[Ailuros]

The A8X uses 128bit LPDDR3 (25GB/s) for an 8-cluster GPU capable of ~230GFLOPs at the estimated 450MHz.
If you assume that apple/Intrinsity usually produce very balanced designs, you have 25GB/s for a 3-core CPU and a 230GFLOPs GPU.
An 8CU (512sp) GPU at ~400MHz would do ~410GFLOPs.

...at 450MHz it's 230.4 GFLOPs FP32 or 460.8 GFLOPs FP16, so pick your poison It being a DR it can save more bandwidth under conditionals. That shouldn't mean that I believe that MTK has integrated anything with 8 clusters/CUs/whatever...I believe it when I see it. Anything with 4 clusters would be already "high end" for their "standards".

If the Helios X20 uses a 128bit bus with LPDDR4, we'd have twice the memory bandwidth for an iGPU with almost twice the FP32 throughput. It doesn't sound unreasonable.

There's never anything like "enough" bandwidth especially in a power constrained ULP SoC environment where a high number of units are struggling for the very same memory bandwidth. It doesn't matter what it'll contain in the end, the bandwidth won't go wasted.

 

[Deleted member 13524]

Where did I write "enough"?
All I wrote was "it's not unreasonable".

 

[Ailuros]

Where did I write "enough"?
All I wrote was "it's not unreasonable".

Of course isn't it unreasonable and I didn't contradict what you said either per se. Bandwidth is never enough whether they use 2x, 3x times or more of it. You don' t have to have a super GPU if you're Mediatek unless they've changed their so far philosophy by quite a few degrees. I'd love to see it, but it still sounds like wishful thinking at this stage.

 

[Nebuchadnezzar]

128bit AXI refers to the interconnect to the other bus, not memory controller.

 

[aaronspink]

So 4 + 4 + 2 would be an option.

Or generally speaking, n + m + k, where n, m, k ≤ 4 and n + m + k = 10.

A solution with four clusters would also be possible, but probably stupid.

Any phone soc with 10 cores is already stupid so might as well go full retard.

 

[Ailuros]

big.LITTLE has always been a power saving initiative. Since the two A53 clusters can run at different frequencies it's considering real time usability rather a 4+2 config than anything else. The two A72 cores for milliseconds for things like boot up, application start, browser page changes etc. and the majority of work swapping between the two A53 quad clusters depending on how heavy the remaining tasks are.

However A53 aren't as tiny (in a relative sense) as A7 cores anymore. Is it really a benefit considering die area in the end, or should ARM maybe consider aSMP for future CPU IP clusters in the end?

 

[Deleted member 13524]

- 2*A72 @ 2.5GHz + 4* A53 @ 2.0GHz + 4*A53 @ 1.4GHz
- Mali T880 MP4 @ 700MHz
- dual-channel LPDDR3 @ 933MHz for 15GB/s

Some caveats: by not adopting LPDDR4, memory bandwidth didn't seem to go keep up with the rest of the SoC. WiFi is restricted to 1x1, though 802.11ac is supported.

How does the T880 compare to the T760 core-for-core at same clocks? ARM claims 1.8x better performance, but suggests 850MHz for the T880 and 650MHz for the T760.
If we take down these clock speed differences, a T880 core might be around 38% faster than a T760 one. This means the T880 MP4 @ 700MHz won't reach Exynos 7420's T760MP8, but should be pretty close to Exynos 5433.

 

[Blazkowicz]

Are they claiming 110dB SNR on the headphone output or I guess that's theoretical performance when paired with some capacitors etc. that won't ever get fitted to a cell phone .

Anyway it crosses many checkboxes such as VP9/H265, useless camera megapixels (that's going into 8-blade razor territory, and I'm scared to check if an 8-blade razor actually exists)
Just one channel of 5GHz wifi is awesome already, as most everyone are stuck to 2.4GHz on their devices (even desktops/laptops)

I honestly wouldn't give a damn about memory bandwith and GPU performance ; the stupid CPU can make use of the A72 on pig web pages I guess.
Missing is the max memory amount supported (2GB, 3GB, 4GB etc.)

I wonder about the flash storage. Let's say you're transfering some gigabytes of data (music is a good test as it's many smallish files) and it goes like this : PC or NAS -> gigabit ethernet -> router -> 5GHz wifi -> phone.
Then phone's write speed has to at least match the transfer rate, I mean even if it's just a real 20MB/s or such it has to keep up (while possibly using the phone or device for some other task). I occasionally copy music at 4MB/s to USB flash drives.
So, does it support something like the latest eMMC versions? then the question is probably about what grade or kind of flash is integrated to device. Not related to the SoC by itself but perhaps we'd need some certification for phones/tablets (like SD card class 10) to reassure the buyer and user?

 

[Nebuchadnezzar]

How does the T880 compare to the T760 core-for-core at same clocks? ARM claims 1.8x better performance, but suggests 850MHz for the T880 and 650MHz for the T760.
If we take down these clock speed differences, a T880 core might be around 38% faster than a T760 one. This means the T880 MP4 @ 700MHz won't reach Exynos 7420's T760MP8, but should be pretty close to Exynos 5433.

We'll have to wait on actual devices before we ever find out. ARM made similar claims about the T760 and we know how that ended up in terms of perf/clock/core. They told us they use perf comparisons to gen -2 at who knows what configuration.

 

[Ailuros]

What no AMD in the Helio20? *cough* LOL

Other than that the GPU is more or less on the level I expected it. Nice, but nothing to write home about either. It's Mediatek after all.

 

[Deleted member 13524]

What no AMD in the Helio20? *cough* LOL

I'm putting my money on Helio X30! :sly:

 

[Ailuros]

I'm putting my money on Helio X30! :sly:

Are you also planning to change your name to FudoTranz?

 

[tangey]

IMG results yesterday listed Spreadtrum as a new licensee. In the post results conference call, IMG's CEO confirmed they'd signed a graphics deal with a new partner which is a tier-1 mobile player. He wouldn't be drawn on whether this was the Spreadtrum license. However to me it looks very likely.

Spreadtrum are growing at an enormous rate in china, Intel have a 20% stake and also some sort of co-development agreement which sees Spreadtrum having access to the Intel foundries, which is different to the rockchip/Intel agreement.

 

[Rys]

While it's fresh in my head, I wanted to mention that I had a look at MT6735V earlier in the week (physically, die analysis). It's probably the most impressively integrated SoC I've seen since I started looking at SoC dice up close. Less than 50mm2 in 28LP yet still packing the digital part of an advanced LTE modem, almost zero dead area or routing yet still with simple layout blocks, and the smallest A53s I've seen to date.

MediaTek really know what they're doing when it comes to chip integration and the physical side of things.

 

[Deleted member 13524]

So the MT673x are not MT675x chips with a disabled quad-core module?


But damn, 50mm^2.. Nintendo should definitely hire those guys for the NX

 

[Nebuchadnezzar]

and the smallest A53s I've seen to date.

On 28nm I suppose?

 

[Rys]

On 28nm I suppose?

Yeah, 28LP.

 

[Rys]

So the MT673x are not MT675x chips with a disabled quad-core module?

Nope, they're completely separate chips. There's a few MT6735 variants, but they're just bins of the same silicon I looked at.

 

[tangey]

A lot of the mediatek discussion has taken place on here, so thought I'd use this thread.

There has been some rumours that Mediateks X30 chip will be switching out Mali graphics in favour of PowerVr.

Is imgtec.eetrend.com officially associated with IMG ? It appears to be given it's wide use of IMG livery and logo, and the fact it bills itself as "Imagination, in association with eetrend"

Anyhow, it's carrying an article that states Mediatek's X30 has 4 core Series 7, which if it's on an officially sanctioned IMG website, would lend quite some authority to the rumours.

HTTP://imgtec.eetrend.com/news/7611

 

[tangey]

The article I referenced above, having been on said website since 31st Mar, has disappeared a couple of hrs after I queried it on twitter @ImaginationPR

Clearly, a coincidence.

Here's an image from google cache for posterity. The relevant translation reads:-

"In addition, Helio X30 will be powered by 4 core PowerVR 7XT GPU, which is high-end PowerVR GPU, support 26 million pixel camera and dual ISP support VR connections, and LTE Cat.13."