AMD - Piledriver chitchat

[watercooled]

Once the 10nm process is more mature, I expect Intel to release the desktop and server Cannonlake CPU's.

But Cannonlake is 'only' a die shrink of the Skylake architecture, so even itself probably not worth upgrading to from Skylake on the desktop, and that's if it makes it to a full desktop release at all.

I mean it's not like Skylake is slow, and it does at least give some AMD some breathing room if we have to take away some positives, but it's getting to be an awfully long time between worthwhile CPU upgrades.

In other news, I know Geekbench isn't *the* standard benchmark but they do seem to have improved things a fair bit with Geekbench 4, and it looks like Apple are getting within ballpark performance of the Core CPUs, and at mobile CPU power (in which I don't include Core m as they 'boost' their power consumption orders of magnitude higher to reach their peak performance). http://www.macrumors.com/2016/09/15/...n-macbook-air/

Again, I'm not claiming Geekbench is more than it is, but it's impressive nonetheless. I'd really like to see some more non-mobile benchmarks tried on the CPU, Openbenchmarking.org would be a good place to start, and maybe some video encoding thrown in for good measure.

[badass]

Allowing for the usual "the numbers aren't comparable", reports are that Intel are being caught up (http://www.fudzilla.com/news/process...ping-over-10nm) and even things like their new flash replacement isn't working out as it should (http://semiaccurate.com/2016/09/12/i...y-much-broken/)

On the whole "The numbers aren't comparable" thing, my understanding is that the front end if Samsung/Globalfoundries 14nm process is. It's the back end that isn't.
I also believe that the performance of a chip is based on the front end and the die size is based on the back end.

So even though TSMC/Samsung/Globalfoundries used their 20nm back end for their 16/14nm process, it's speed and efficiency is. It's just that the dies are bigger.

Whilst it's easy to get excited about Globlefoundries catching Intel up, I wouldn't hold my breath. They had test wafers out for each node at similar times to intel. They were still over a year behind on mass production. Hopefully this changes

[badass]

But Cannonlake is 'only' a die shrink of the Skylake architecture, so even itself probably not worth upgrading to from Skylake on the desktop, and that's if it makes it to a full desktop release at all.

I mean it's not like Skylake is slow, and it does at least give some AMD some breathing room if we have to take away some positives, but it's getting to be an awfully long time between worthwhile CPU upgrades.

Agreed, but they aren't targeting Skylake users. The will probably target Westmere and Sandy Bridge users as It will have a decent speed bump over them.

I've got an ivy Bridge and I compared it to the equivalent Skylake. It's only 20-30% faster at best.

[watercooled]

On the whole "The numbers aren't comparable" thing, my understanding is that the front end if Samsung/Globalfoundries 14nm process is. It's the back end that isn't.
I also believe that the performance of a chip is based on the front end and the die size is based on the back end.

So even though TSMC/Samsung/Globalfoundries used their 20nm back end for their 16/14nm process, it's speed and efficiency is. It's just that the dies are bigger.

Whilst it's easy to get excited about Globlefoundries catching Intel up, I wouldn't hold my breath. They had test wafers out for each node at similar times to intel. They were still over a year behind on mass production. Hopefully this changes

Samsung were shipping 14nm processors in volume arguably before or around the same time as Intel - Broadwell 'launched' a couple of months before but you couldn't expect to actually buy one for quite some time after that, and they started with just a couple of the low power mobile chips. They're not the same node/products but it was a surprise nonetheless. Even if you consider that a proper launch, Samsung were only ~3 months after, tops.

[DanceswithUnix]

In other news, I know Geekbench isn't *the* standard benchmark but they do seem to have improved things a fair bit with Geekbench 4, and it looks like Apple are getting within ballpark performance of the Core CPUs, and at mobile CPU power (in which I don't include Core m as they 'boost' their power consumption orders of magnitude higher to reach their peak performance). http://www.macrumors.com/2016/09/15/...n-macbook-air/

I always look forward to Apple APU analysis when new products come out, most interesting thing about their range
They are already fast, and seem to just get faster. I presume they are beating Intel way down on price with a threat that they just need to knock out a chip optimised for performance rather than power, and then they don't really need x86 any more. Switching would be a hassle and expense, but they have done it twice before.

Or Apple just let the Arm based stuff slowly display the x86. Perhaps the sales are already unbalanced to the point x86 isn't important to them in terms of profit?

Edit: Nvidia seem to have a new chip as well with their latest Denver CPU on it. Those usually end up in development boards, so I am hopeful we can get some interesting benchmarks on that. The old Denver better than the Intel laptop chips in a few benchmarks but still had a way to go.

[badass]

Samsung were shipping 14nm processors in volume arguably before or around the same time as Intel - Broadwell 'launched' a couple of months before but you couldn't expect to actually buy one for quite some time after that, and they started with just a couple of the low power mobile chips. They're not the same node/products but it was a surprise nonetheless. Even if you consider that a proper launch, Samsung were only ~3 months after, tops.

I have faith in Samsung delivering the goods. Just not globalfoundries. Theyve ruined several AMD launches with delays then yield issues.

[watercooled]

I have faith in Samsung delivering the goods. Just not globalfoundries. Theyve ruined several AMD launches with delays then yield issues.

Ah yeah just realised you were talking specifically about GloFo with that comment, not all of them.

@DanceswithUnix: I agree, especially for things like the Macbook Air there's no real need for x86; in fact it's not really needed at all but it would mean breaking app compatibility (like you say, this is something they've worked through before) and perhaps more problematic is Parallels i.e. you wouldn't be able to run Windows - the Windows-x86 stranglehold has even spread over to Apple!

Maintaining two incompatible versions of OSX (e.g. ARM on the Air) would be messy. So even though an ARM architecture processor could be technically a great fit for some/all OSX devices, that alone probably isn't enough of a motivation to switch as long as an x86 processor is good enough for their needs and not absurdly expensive.

WRT Denver - IIRC it was good in some workloads but not so good in others. I do look forward to seeing how the new one performs but it might never make it to consumer electronics by the looks of it.

[DanceswithUnix]

Interesting article on S/A comparing the usability of an APU vs RX-460: http://semiaccurate.com/2016/09/19/r...y-upgrade-apu/

The 460 is definately an upgrade, but on average only about a doubling in performance. Makes you wonder if there is much of a market these days for anything slower.

[watercooled]

Yeah the market for entry-level GPUs must be getting seriously eroded by APUs now - even Intel's HD 530 has very respectable performance. There's still a market for basic GPUs though e.g. systems based around workstation CPUs without integrated GPUs, but where GPU performance isn't as important.

[DanceswithUnix]

Some 9800 benchmarks, looks like a nice chip. I'm sure it would sell well if you could actually buy one.

http://wccftech.com/amd-bristol-ridg...m-performance/

[DanceswithUnix]

Fudzilla claim that Vega 10 is a monster with more flops than a P100 and 16GB of HBM2: http://www.fudzilla.com/news/graphic...ores-24-tflops

[CAT-THE-FIFTH]

Fudzilla claim that Vega 10 is a monster with more flops than a P100 and 16GB of HBM2: http://www.fudzilla.com/news/graphic...ores-24-tflops

http://videocardz.com/63700/exclusiv...a10-and-vega20

Looks like Vega 11 is replacing Polaris 10 next year and Vega 20 will be arriving at some point.

[CAT-THE-FIFTH]

Looks like a Korean website briefly tested the A12-9800:

http://wccftech.com/amd-bristol-ridg...m-performance/

However,the RAM was only tested in single channel with the A12-9800 based system it appears. The power consumption drop is quite large too.

Edit!!

Also the A12-9800 was overclocked to 4.8GHZ on the stock cooler in this video too:

https://forums.overclockers.co.uk/sh...postcount=1702

[DanceswithUnix]

Looks like a Korean website briefly tested the A12-9800:

http://wccftech.com/amd-bristol-ridg...m-performance/

Look a few posts above

Also the A12-9800 was overclocked to 4.8GHZ on the stock cooler in this video too:

https://forums.overclockers.co.uk/sh...postcount=1702

Seems to be a Wraith cooler, which would answer the question of whether the cooler standard has changed.

[DanceswithUnix]

Interesting throw away comment here: "Since all AM4 CPUs have an unlocked multiplier, ..."

http://wccftech.com/amd-x370-chipset...ge-processors/

[CAT-THE-FIFTH]

AMD Naples apparently has 512MB of L3 cache:

http://wccftech.com/amd-zen-32-core-...hmarks-leaked/

WTF??