AMD - Zen chitchat

[watercooled]

Apparently (although I can't find much to back it up yet, although AIDA didn't deny it) the AGESA patch decreases Ryzen's memory latency to ~60ns, which would put it right in line with Intel. I wonder if Hexus could test this?

https://twitter.com/tekwendell/statu...24729791434752

[kompukare]

Hardware.fr did their Overclocking en pratique for their Ryzen R5 reviews too.
And with that it's possible to answer the question: what kind of board do I need to overclock an R5 1600?

My guess was going to be that any board which can run stock 1800X should be fine with moderately overclocking a 1600, and this more or less confirms this:

For the 1600, even 3.9GHz @ 1.3V should be fine (less than 2W more at the mains than stock 1800X).
For the 1600X, 4GHz @ 1.35V might be a bit too much (5.1 more than stock 1800X).
The 1500X and 1400 are all fine at the max sane clocks and voltages.
So a B350 board @ £75 should be fine for most users. Those B350 prices have gone down @£15 since I last looked at them so there is no reason why even good B350 boards shouldn't come down to £60 or so in a few months given very few reasons to chose an i5.

[CAT-THE-FIFTH]

For some weird reason they still seemed obsessed with the X370 motherboards over on OcUK - I pretty much was telling people looking at previous AMD CPUs,like the FX6300,etc you could get away with a cheaper motherboard.

Personally buying a nearly £170 motherboard for a £220 CPU seems a bit excessive!!

[kompukare]

Totally skews the value proposition too.
R5 1600 + B350 is good value at @ £300 compared to an i5-7600K + Z270 and cooler @ £320.
But if people are going to crazy with X370 boards even for R5 1500X or 1400 then those compares badly to a i5-7500 + cheap B250 board @ £236 or so.

(Of course, as usual Intel didn't do themselves any favours by not making H110 boards forward compatible with Kabylake without a new BIOS. Can't really see any good reason why boards shouldn't be able to take future CPUs and at least be able to update the BIOS with them. Open Firmware has been around since 1994 after all.)

[CAT-THE-FIFTH]

If you buy an R5 1600 and a £170 motherboard that is nearly £400. You can buy a Core i7 7700K and a cheapo B250 motherboard for not much more.

I can appreciate some think the more expensive boards will have a greater chance of getting Ryzen 2,etc,but TBH you might as well either wait for Ryzen 2 in the first place or put the money into a faster graphics card.

[watercooled]

There's a more recent AMA over at Tom's: http://www.tomshardware.com/reviews/...-ama,5018.html

The answers are quite genuine and non-evasive IMO, apart from the ones they're outright not allowed to answer of course.

Apart from the first question which has a strange answer anyway, about TDP. I think I kinda get what he's trying to say, but it comes across like electrical watts != thermal watts, which is nonsense because at any given time they're obviously identical. But with short periods of 'boost' it can allow some room for play because of thermal mass etc. E.g. even with a 95W-capable thermal solution, you could still boost to >120W or so for short amounts of time provided you average out at 95W, much like Intel do to much greater extents with their optimistically-spec'd Core M parts.

[Terbinator]

For some weird reason they still seemed obsessed with the X370 motherboards over on OcUK - I pretty much was telling people looking at previous AMD CPUs,like the FX6300,etc you could get away with a cheaper motherboard.

Personally buying a nearly £170 motherboard for a £220 CPU seems a bit excessive!!

Especially when there isn't even much OC room available.

[Xlucine]

There's a more recent AMA over at Tom's: http://www.tomshardware.com/reviews/...-ama,5018.html

The answers are quite genuine and non-evasive IMO, apart from the ones they're outright not allowed to answer of course.

Apart from the first question which has a strange answer anyway, about TDP. I think I kinda get what he's trying to say, but it comes across like electrical watts != thermal watts, which is nonsense because at any given time they're obviously identical. But with short periods of 'boost' it can allow some room for play because of thermal mass etc. E.g. even with a 95W-capable thermal solution, you could still boost to >120W or so for short amounts of time provided you average out at 95W, much like Intel do to much greater extents with their optimistically-spec'd Core M parts.

It's not completely true to say that power into a CPU ≡ thermal power dissipated, because there's capacitances and downstream components (as well as the upstream inefficiencies)

[watercooled]

Depending on how pedantic one is, you could also factor in EM and audible emissions, but they're usually safe to consider negligible (I usually attach that disclaimer any time I say this but thought it would be safe to leave it out by now). DRAM has its own power phase so can be considered separately.

With DC power, reactance from capacitance and inductance doesn't come in to it.

For all intents and purposes, aside from PCB trace thermal losses, power supplied from the VRM is dissipated almost entirely by the CPU as heat. Power lost via signals is also negligible, not least because any sort of higher power signals get their own off-package amplification.

[Xlucine]

Chips aren't really DC though. Yes, there is a fairly constant voltage across them, but they don't drain current at a constant rate. A CMOS logic gate will pass no current in normal operation, only letting through a wodge of current when it flops to a different state - this is why power consumption is proportional to FV^2, rather than V^2. So with current behaviour that is pretty different to a DC resistive load, inductances and capacitances are going to be significant

[watercooled]

It's irrelevant - DC (very tightly-regulated DC for that matter) is supplied by the VRM to the CPU. You cannot simply 'lose' energy to capacitance - a perfect capacitor stores and releases energy, a real capacitor has a resistance, so you're either using the energy somewhere else or dissipating it in the resistance.

It's also irrelevant to my point, but the transistors founds in modern CPUs are not perfect switches either, they leak even when they're 'off'. What you're referring to is dynamic power which is only one aspect of a CPU's power consumption; there's also short-circuit dissipation and leakage to factor in. The end result is heat, regardless.

Even in an AC circuit, you still never, ever lose energy to capacitance itself. Capacitance can lead to losses e.g. additional resistive losses because of the higher induced current.

Just remember conservation of energy - the energy always has to end up somewhere, it cannot simply cease to exist.

[CAT-THE-FIFTH]

AMD is also now using Samsung to make Polaris GPUs:

https://translate.google.co.uk/trans...%2F&edit-text=

Plus why no RX550 and RX560 reviews?? These are the parts which are the biggest upgrades - Nvidia has no decent $80 graphics card from the GTX1000 series either!!

[kompukare]

AMD is also now using Samsung to make Polaris GPUs:

https://translate.google.co.uk/trans...%2F&edit-text=

Plus why no RX550 and RX560 reviews?? These are the parts which are the biggest upgrades - Nvidia has no decent $80 graphics card from the GTX1000 series either!!

So it was LPP (Low Power Plus) all along and not LPE (Low Power Early). Guess that does at least explain why there are no efficiency gains. Since the process and the design get very inefficient past 1400MHz, you'd really think AMD would have learned and tried pawning off these pre-overclocked cards as the 500 series.

Wonder what it is the optimal speed for Polaris? That is did anyone do a perf/watt chart at various frequencies and voltages like we saw with, for example, The Stilt's Ryzen charts?

Has to be said that GP107 on 14nm (undisclosed but rumoured to be Samsung) runs slower than GP106 so it looks like TSMC's 16nm is the better process.

But as GPUs are nicely parallel, it should be possible to run narrow and fast or slow and wide for a similar performance, and since GF's process appears to be worse AMD must be paying less for their wafers. So they could have designed a larger chip, ran it slower and get far better perf/watt. But then again they are already loosing badly in terms of perf/area and total BOM (GTX 1060 only has a 192-bit bus, consumes around 40W less than RX480 meaning partners can use cheaper VRMs, fans etc.), so that kind of approach would only improve perf/watt but make the other metrics worse.

[scaryjim]

... Wonder what it is the optimal speed for Polaris? That is did anyone do a perf/watt chart at various frequencies and voltages like we saw with, for example, The Stilt's Ryzen charts? ...

On first release CAT posted a voltage v clock chart that showed Polaris sitting at ~ 0.8v up to between 900MHz and 1GHz, so somewhere in that range. There's a FirePro polaris card that uses Polaris 10 Pro (2048 shaders) in a 75W envelope, which I guess is clocked at around that peak efficiency figure...

At least the Polaris refresh is reaching higher stable clocks, although these cards should totally have been the 485 and 475 - when AMD announced the new naming scheme they specifically set aside the final digit of the number for a refreshed version of the same card! May as well not bother with an advertised naming scheme and just make it up as they go along.... *sigh*

[shaithis]

Especially when there isn't even much OC room available.

and I wouldn't be surprise to see Zen2 require AM4+

[kompukare]

On first release CAT posted a voltage v clock chart that showed Polaris sitting at ~ 0.8v up to between 900MHz and 1GHz, so somewhere in that range. There's a FirePro polaris card that uses Polaris 10 Pro (2048 shaders) in a 75W envelope, which I guess is clocked at around that peak efficiency figure...

Always hard to find FirePro reviews, but it seems there were two Polaris 10 cards with the one being a huge cut-down. WX7100 was 2304:128:32@900(boost 1240) while WX5100 was 1792:112:32@926 (but boost 1090). The WX5100 has the 75W TPU. All according to the wikipedia (no ark.amd.com so finding specs from AMD is a lot more difficult than for Intel):
https://en.wikipedia.org/wiki/Radeon_Pro#Chipset_Table
Did find one place reviewing a WX5100 but of course no games:

Seems only the fourth test down shows a bit difference.
This place
http://techgage.com/article/amd-rade...-gpu-review/5/
reviewed it too and tested some 'gaming' but only Unigine Heaven and 3DMark and besides they have no gaming cards in there to compare

At least the Polaris refresh is reaching higher stable clocks, although these cards should totally have been the 485 and 475 - when AMD announced the new naming scheme they specifically set aside the final digit of the number for a refreshed version of the same card! May as well not bother with an advertised naming scheme and just make it up as they go along.... *sigh*

Damned if you do, and damned if you don't I guess. OEMs want shine 'new' 500 series so the announced xx5 would only have been a factor if they had a respin within 3-6 months of the initial launch.