AMD - Piledriver chitchat

[DanceswithUnix]

AMD Naples apparently has 512MB of L3 cache:

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

WTF??

Interesting if true. A cache that size won't be fast though, and at that size I have to wonder if that is a 4 CPU dies on an interposer and the L3 is some external chip also on the interposer.

Edit: Or it could be a misinformation leak to mess with Intel's heads

[badass]

AMD Naples apparently has 512MB of L3 cache:

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

WTF??

I always thought that only 4 Channel DDR4 for 32 cores was not enough. Now I know why they only need 4 Channels. They will only be hitting that memory once in a blue moon!

[scaryjim]

Hmmm, odd that it's listed as 64MB x8 ... from the table further down the 8 core summit ridge is listed as having 16MB cache - that's 2MB per core, or 64MB total on a 32 core chip. Could easily just be an incorrectly displaying cache amount.

OTOH setting up a modular design where 4 cores share a chunk of L3 cache makes a lot of sense: it's just a question of whether 4 cores sharing 64MB of L3 cache makes sense...

[CAT-THE-FIFTH]

Interview with Mark Papermaster about Zen:

http://semiengineering.com/the-zen-o.../?sf37016781=1

[CAT-THE-FIFTH]

A really small Lenovo BR system with a 35W A12-9800E:

https://www.computerbase.de/2016-10/...-tiny-amd-pro/
http://techreport.com/news/30857/len...k-amd-pro-apus
http://www.marketwired.com/press-rel...md-2169115.htm

Edit!!

Spec sheet:

http://isby.s3.amazonaws.com/lenovop...-Datasheet.pdf

Second Edit!!

Just saw this:

Performance labs. PC manufacturers may vary configurations yielding different results. 3DMark 11 Performance is used to simulate graphics performance, and Cinebench R11.5 1T Performance is used to simulate single threaded CPU performance; the 7th Generation AMD PRO A12-9800 at 65W scored 3521.25 and 1.21 while the AMD PRO A10-8850B at 95W scored 2880 and 1.06 respectively.

That CB single thread score is higher than any other AMD CPU apart from the FX9590 AFAIK and is around the same as my old Core i3 2100.

[DanceswithUnix]

Wccftech play BF1 on just an APU: http://wccftech.com/battlefield-1-amd-a10-apu/

Wonder how much better that would be on a desktop A12 if AMD would just pull their damned finger out and sell them.

[CAT-THE-FIFTH]

Wccftech play BF1 on just an APU: http://wccftech.com/battlefield-1-amd-a10-apu/

Wonder how much better that would be on a desktop A12 if AMD would just pull their damned finger out and sell them.

By the time AMD actually get off their arse,Kaby Lake Core i3 chips will be out with even a faster IGP,and then for DIY builds nobody wants to get FM2+ when they know AM4 is coming out.

[CAT-THE-FIFTH]

Looks like Polaris is in the new Macbooks:

http://arstechnica.com/apple/2016/10...-macbook-pros/

http://radeon.com/en-us/radeon-pro-g...w-macbook-pro/

The GPU is very thin:

Die thinning is an elaborate process that reduces the thickness of each wafer – those thin slices of material used in circuits — from 780 microns to 380 microns, or 0.38 millimeters. This allows for impossibly thin, beautiful designs.

[badass]

I'm surprised no one's mentioned the Embeded Polaris GPU's that seem to have the same performance at 2/3 of the power budget http://www.anandtech.com/show/10710/...on-e9260-e9550

Also referred to as 50% more performance per watt in other articles:
http://wccftech.com/amd-polaris-revi...ance-per-watt/
http://vr-zone.com/articles/amd-roll...tt/115217.html

It looks like there's a performance boost to be had as well.
I do believe that makes the Polaris 10 based board both more powerful and more efficient than the Geforce 1060. As well as being more future proof.

[DanceswithUnix]

I'm surprised no one's mentioned the Embeded Polaris GPU's that seem to have the same performance at 2/3 of the power budget http://www.anandtech.com/show/10710/...on-e9260-e9550

It had been in other threads.

It is a fairly standard trick though, using the fully enabled part and reducing the clock speed, that then allows the voltage to be reduced and the power drops proportional to the square of the voltage drop.

[badass]

It had been in other threads.

It is a fairly standard trick though, using the fully enabled part and reducing the clock speed, that then allows the voltage to be reduced and the power drops proportional to the square of the voltage drop.

This is not what they are doing - the clock speed is not reduced at all. 1.266GHz is the boost clock for both the RX480 and the embedded version and both are fully enabled Polaris 10 GPU's. The actual GPU has been revised - more metal layers or something similar. Think Athlon XP Thoroughbred A-B transition or 4870 vs 4890. Huge improvement in efficiency by adding a metal layer on the same process. https://en.wikipedia.org/wiki/Athlon...d_.28T-Bred.29
or 4870 vs 4890

The GPU will perform computationally identically at the same clock speed but use substantially less power.

[Xlucine]

This is not what they are doing - the clock speed is not reduced at all. 1.266GHz is the boost clock for both the RX480 and the embedded version and both are fully enabled Polaris 10 GPU's. The actual GPU has been revised - more metal layers or something similar. Think Athlon XP Thoroughbred A-B transition or 4870 vs 4890. Huge improvement in efficiency by adding a metal layer on the same process. https://en.wikipedia.org/wiki/Athlon...d_.28T-Bred.29
or 4870 vs 4890

The GPU will perform computationally identically at the same clock speed but use substantially less power.

Not with more cores and less FLOPS. The mobile 460 has 14% more cores but only 87% of the performance, suggesting a clock speed of ~900MHz

It had been in other threads.

It is a fairly standard trick though, using the fully enabled part and reducing the clock speed, that then allows the voltage to be reduced and the power drops proportional to the square of the voltage drop.

Power consumption of a CMOS circuit is proportional to v^2 and frequency! So running at lower volts and Hz gets you more power savings than v^2 would suggest

[DanceswithUnix]

The GPU will perform computationally identically at the same clock speed but use substantially less power.

Ah right, I had only read about the 460 where the desktop part wasn't fully enabled. Mobile 480 just hadn't registered with me for some reason. New metal masks? Nice.

Interesting to see articles wondering if we will see this silicon in the desktop variants at some point. I doubt that AMD have a warehouse full of unused chips at this point, and I can't imagine them producing two different chips if one is clearly superior and they share the same expensive silicon patterning. Really, I can't imagine how they will stay out of desktop cards, though binning might still see less change for us.

[badass]

Ah right, I had only read about the 460 where the desktop part wasn't fully enabled. Mobile 480 just hadn't registered with me for some reason. New metal masks? Nice.

Interesting to see articles wondering if we will see this silicon in the desktop variants at some point. I doubt that AMD have a warehouse full of unused chips at this point, and I can't imagine them producing two different chips if one is clearly superior and they share the same expensive silicon patterning. Really, I can't imagine how they will stay out of desktop cards, though binning might still see less change for us.

It looks like what you said is exactly what they have done to the 460 more cores, less clocks. Nothing to see here - yawn My interest was in the 480 because that is the GPU I am more interested in. (460 esque performance is no better than my current GPU but the current 480 is just too hot for my case to extract heat quietly)

In terms of desktop variants, I expect it depends on the cost of the extra metal layers - if it costs anything extra at all. If the extra cost is non existent or minimal, they will probably look to transition to the newer spins as quickly as practical. Probably release RX xx5 variants with the clock boosted a little bit and the power between the old and new revisions or something.

I'm still interested in this also because it shows Polaris reaching the perfect Trifecta over the GTX 1060 - more powerful, less power and less cost.

[CAT-THE-FIFTH]

It looks like what you said is exactly what they have done to the 460 more cores, less clocks. Nothing to see here - yawn My interest was in the 480 because that is the GPU I am more interested in. (460 esque performance is no better than my current GPU but the current 480 is just too hot for my case to extract heat quietly)

In terms of desktop variants, I expect it depends on the cost of the extra metal layers - if it costs anything extra at all. If the extra cost is non existent or minimal, they will probably look to transition to the newer spins as quickly as practical. Probably release RX xx5 variants with the clock boosted a little bit and the power between the old and new revisions or something.

http://www.anandtech.com/show/10521/...olaris-to-pros
http://www.anandtech.com/show/10710/...on-e9260-e9550








I'm still interested in this also because it shows Polaris reaching the perfect Trifecta over the GTX 1060 - more powerful, less power and less cost.

Thats the thing the TDP reduction must be massive to get a full Polaris 11 GPU into a single slot and low profile card. Even getting a cut-down Polaris 10 into a single slot card is pretty impressive IMHO OFC.

Also,you might want to check out the reviews of the XFX RX480 GTR - it seems to have much lower power consumption and higher clockspeeds and was the last of the AIB models to be launched.

The Powercolor Devil RX480 also seems to drop power consumption a reasonable amount.

[DanceswithUnix]

If the extra cost is non existent or minimal, they will probably look to transition to the newer spins as quickly as practical.

It has to cost something, more layers means more patterning. The thing is, the up-front cost of generating and testing masks seems to be vast these days so they have already invested the big money. If they run the old and new lines together, well what if they get a big order and all they have on hand is the other chips? For inventory control it just makes sense to make one chip, and bin the output for best use, but keep some flexibility.

There is an issue that with much reduced power consumption will come boards with fewer VRM phases. That could well make overall product costs neutral or cheaper, but you don't want to put an old chip in a low power board unless you want it to go bang like an EVGA 1070! The BIOS would want to spot that, so I presume new chips will come with a different PCI ID at least.

Hopefully the lessons learnt here have fed into Vega to give that a more impressive launch. The 480 is faster than my 380, but I am thinking for the money not enough of an upgrade so I am waiting for a Vega based 580 now.