AMD - Piledriver chitchat

[kalniel]

Not going to rise to it!

Sorry, no baiting intended! I shouldn't have included the nVidia comment, but I am genuinely hopeful AMD can get more out of the silicon, speed-wise. I was expecting a little more but that's why I think they are being conservative with early silicon (=good volume - I think they're going to get a LOT of orders).

edit: Do we know what AMD's boost strategy is for Polaris yet?

[CAT-THE-FIFTH]

I didn't expect AMD would beat Nvidia in clockspeeds for two reasons:
1.)On the same node GCN could not hit the clockspeeds Maxwell could
2.)GF/Samsung 14NM seems to be more density and power consumption orientated and TSMC 16NM FF+ more performance orientated
3.)Nvidia was hitting nearly 1.5GHZ with Maxwell for normal overclocks and pre-overclocked cards. AMD was lucky to see even 1.2GHZ with later GCN cards with a ton of voltage which makes it a 25% increase.

My GTX960 in a mini-ITX case usually hovers at 1.4GHZ which is like 350MHZ more than a R9 380.

The GTX1080 seems to be pushed almost to its max boost clocks at 2GHZ which is a 33% max clockspeed increase over Maxwell.

AMD has managed to gain at least 21% over the other GCN cards which were around 1.05GHZ at max for stock clockspeeds,but we have no clue whether AMD has tuned the clockspeeds for performance or performance/watt or what the max clocks might be.

But they will be less.

[kalniel]

AMD has managed to gain at least 21% over the other GCN cards which were around 1.05GHZ at max for stock clockspeeds,but we have no clue whether AMD has tuned the clockspeeds for performance or performance/watt or what the max clocks might be.

Early GCN (well, my 7870 ) could easily hit 1.266ghz on OC, 7970s even higher (phage's about 1.5ghz) so hopefully Polaris has similar headroom

[watercooled]

I think it's still funny the amount of people making out that AMD not having their high-end part out as being catastrophic. I mean I admittedly guessed P10 (and GP104) would be more like 250-300mm2 IIRC but gunning for the mainstream and mobile markets is completely sensible especially given its an area where they'll have no competition for a while. FinFET design is expensive so aside from marketing and fanboy fuel it makes a lot of sense.

Granted the very top-end have higher ASP but bigger dies are much more expensive to produce. Going by the cost being in line with the square of die size (feel free to correct me if I'm misusing that guideline), a ~300mm2 die would be more than twice as expensive as a 200mm2 die. It gets more complicated when you factor in die harvesting but it also looks like the 480 might not be quite fully-enabled either so it's likely much cheaper to produce than the 1070 - if it can get close in performance then it's surely a win for AMD?

AMD have at least 4 dies that we know of running through 14nm at the moment; Polaris 10+11, Zen, and Vega 10 - that's a lot for a company of AMD's size, not to mention any semi-custom stuff they have lined up. Up until a couple of years ago, I remember £150 being the sweet spot for good value GPUs (roughly what I got my 4850 and 5870 for) so you won't hear me complaining if they manage to return to that! I'm perfectly happy at 1080p and 60fps, and don't care if I can't use indistinguishable 'ultra' settings, so I don't care much about the top-end stuff.

[Bagnaj97]

Early GCN (well, my 7870 ) could easily hit 1.266ghz on OC, 7970s even higher (phage's about 1.5ghz) so hopefully Polaris has similar headroom

My 7870XT is unstable above 1125, so that's not always the case!

[kalniel]

Granted the very top-end have higher ASP but bigger dies are much more expensive to produce. Going by the cost being in line with the square of die size (feel free to correct me if I'm misusing that guideline), a ~300mm2 die would be more than twice as expensive as a 200mm2 die.

Sounds close or thereabouts. Simplistically, you can get 50% more 200mm^2 chips out of the same wafer size compared to a 300mm^2 chip. So the cost per 300mm^2 chip is unsurprisingly 1.67x higher. Add a little for increased wastage around the edge of a wafer (you can get closer to the edge with a smaller chip). Subtract a little if you can tolerate more faults because you're gating more aggressively (we know 1070 is very cut down, 1060ti further more probably).

Then add a bunch of unknowns because they might actually edge fill with much smaller chips. Who knows?!

AMD have at least 4 dies that we know of running through 14nm at the moment; Polaris 10+11, Zen, and Vega 10 - that's a lot for a company of AMD's size, not to mention any semi-custom stuff they have lined up. Up until a couple of years ago, I remember £150 being the sweet spot for good value GPUs (roughly what I got my 4850 and 5870 for) so you won't hear me complaining if they manage to return to that! I'm perfectly happy at 1080p and 60fps, and don't care if I can't use indistinguishable 'ultra' settings, so I don't care much about the top-end stuff.

Agreed!

My 7870XT is unstable above 1125, so that's not always the case!

To be fair though, the XT was made of parts that should have been thrown away

[Biscuit]

Anyone seen this?


I mean the info is out all over the place now but... hehe

[CAT-THE-FIFTH]

Early GCN (well, my 7870 ) could easily hit 1.266ghz on OC, 7970s even higher (phage's about 1.5ghz) so hopefully Polaris has similar headroom

Most of the gcn1.1 parts onwards could barely break 1.20ghz - and we are talking everyday clocks not suicide runs.

Remember not a single standard GCN SKU gas had reference clocks above 1.05ghz - the optimal clocks for gcn1.1 onwards seen to be around 900mhz.

If you look at HWBot,the hd7970 top out at 1.2 GHz:

http://hwbot.org/hardware/videocard

The same as the HD7870:

http://hwbot.org/hardware/videocard/...pitcairn_core/

If you look at the other cards they tend to be just under that and that is based on 100s or 1000s of submissions.

So AMD has managed to get an uplift of 21% on a process which is meant to be denser and more power efficient but apparently does not clock as high.

My gtx960 will hit just shy of 1.45 ghz stable and tends to boost close to 1.4ghz.

That gives Nvidia close to 35% better clocks on the same node and nearly 25%to 30% higher than an r9 380 when both are overclocked.

Even if Nvidia maintain the same advantage,that mean AMD probably hitting 1.4ghz to 1.5ghz max.

[kalniel]

Don't ask me what Phage did to get his 7970 to ~1.5ghz, I'm not sure I want to know! The HWBot for the 7870 shows average OC of 1.21ghz for the 7870, so my 1.265ghz seems within a reasonable sample variance (that was still at stock volts for eg.).

But yes, your analysis sounds about right. AMD cards seem to punch above their weight over time too, so hopefully some of the gap will be reduced.

[Bagnaj97]

To be fair though, the XT was made of parts that should have been thrown away

In particular my XFX model - The memory is only 1250MHz (all other 7870xts have 1500+) and whilst the cooler looks like one used on the 7870 DD it's a heatpipe down (you can see all the cutouts where it should be).
I think my card is literally XFX clearing out whatever they had left in the parts bin!

[CAT-THE-FIFTH]

In particular my XFX model - The memory is only 1250MHz (all other 7870xts have 1500+) and whilst the cooler looks like one used on the 7870 DD it's a heatpipe down (you can see all the cutouts where it should be).
I think my card is literally XFX clearing out whatever they had left in the parts bin!

I told you to buy an HD7950 for a bit moar,lol!

Even the other HD7870LE cards like the Joker Edition were decent overclocked.

Anyone seen this?


I mean the info is out all over the place now but... hehe

Fast forward to 3 minutes 50 seconds and see Raja Koduri nearly drop the card.

Lol.

[CAT-THE-FIFTH]

It appears Bristol Ridge and Stoney Ridge are new chips:

http://www.anandtech.com/show/10362/...-for-notebooks

The 7th Generation APU, in silicon, is by and large similar to the 6th Generation platform called Carrizo with some updates to help drive efficiency and improve performance. This is due to the updated Excavator core, primarily supporting DDR4 memory through the controller (similar to the R-series embedded AMD parts announced earlier), but also tweaks to the 28nm Bulk CMOS physical design that improved the voltage/frequency characteristics of the silicon. This combined with new AVFS (adaptive voltage frequency scaling) algorithms via improved through internal sensors enables AMD to offer a higher frequency for the same Thermal Design Power on the 7th Generation. With this, AMD is claiming that a move from a Kaveri ‘15W’ APU in 2014 to a new Bristol Ridge APU in 2016 can afford a 56% increase in raw CPU performance, or +10% over last year’s Carrizo.

Additional 6/1: AMD has shared with us die sizes and transistor counts. The Bristol Ridge dual module die, including the integrated graphics, comes in at 250.4 mm2 with 3.1 billion transistors. The Stoney Ridge single module die (including IGP) is 124.5mm2 and 1.2 billion transistors.

So there are know two different dies - one which is 250MM2 and another which is 125MM2.

Carrizo only had one die which was around 245MM2.

[badass]

Shocked at the pricing TBH, now very interested to see reviews. Particularly power consumption as 1x6 pin may be 150W max but they are unlikely to leave no headroom. Maybe we'll see some passive versions out after the launch?

[CAT-THE-FIFTH]

Shocked at the pricing TBH, now very interested to see reviews. Particularly power consumption as 1x6 pin may be 150W max but they are unlikely to leave no headroom. Maybe we'll see some passive versions out after the launch?

I think I read somewhere the card was consuming 120W or thereabouts but I don't know whether that is card power consumption or at the wall.

If we say card power consumption that is around the level of my GTX960. If AMD do hit around R9 390X level performance,and going by the TPU figures it should be around the same as Pascal or double the performance per watt.

Interestingly enough,the GTX1080 is 2.8X time better performance per watt than a R9 390 according to TPU.

If AMD does hit that level of performance per watt,I will be impressed by the fact that GCN1.1 onwards was much less efficient than Maxwell,and AMD with a design using hardware scheduling has gotten close again to a design using software scheduling.

Edit!!

Some renderings of the PCB.

[DanceswithUnix]

Is there still nothing on Polaris 11? I have a case in my collection that can only take a half height single slot card, so will be interested to see what gets made of the lower power part.

[CAT-THE-FIFTH]

Not ATM,but it seems shots were fired by Nvidia:

http://www.trustedreviews.com/news/n...-20-years-away

“Oh my gosh, we’re missing so much,” he said. “First of all, VR displays are a little too cumbersome. It has to be much more elegant, being connected by a wire has to be solved. The resolution has to be a lot higher. The physical worlds do not behave according to the laws of physics. The environment you’re in isn’t beautiful enough. We’re going to be solving this problem for the next 20 years.”

He added: “Having 20 years for employment is a good thing.”

His comments won’t surprise many people who’ve used VR so far, but what’s interesting is how little interest in actually solving this side of the problem the CEO expressed.

They talked a lot about VR and the GTX1080 and the week AMD decides to launch cheaper cards for VR,it suddenly becomes a bit meh.

Got to love how AMD and Nvidia PR chop and change what is relevant every week or so!!