New Zen microarchitecture details

[ElFenix]

measuring pixels as if the marketing flunkie that put that chart together (several times) was at all that accurate is like converting the bible to numbers to predict the date of the apocalypse

 

[JoeRambo]

measuring pixels as if the marketing flunkie that put that chart together (several times) was at all that accurate is like converting the bible to numbers to predict the date of the apocalypse

And if Bible was written by AMD marketing, then we'd be @ ~5th coming of Jesus by now.

This thread keeps going without any real info and leaks. Wake me up when we have a solid estimate of clocks, then we can start making educated guesses about perf if AMD did not screw up design. Remember, any Intel gen since Sandy will destroy Zen with ease if it has clock advantage that is substantial enough.

 

[guskline]

Polaris is on the radar screen now and AMD needs to maximize it, but ZEN will soon need to release more concrete info.

Fortunately for AMD, Broadwell E, though an improvement has been somewhat stymied by price.

Hopefully we will see some more concrete info on Zen in the next 60 to 90 days.

Hey, I have an idea! How about AMD "hiring" the guy that went livestream this morning from China showing what purported to be benchmarking a RX480!:'cool:

 

[Dresdenboy]

And if Bible was written by AMD marketing, then we'd be @ ~5th coming of Jesus by now.

This thread keeps going without any real info and leaks. Wake me up when we have a solid estimate of clocks, then we can start making educated guesses about perf if AMD did not screw up design. Remember, any Intel gen since Sandy will destroy Zen with ease if it has clock advantage that is substantial enough.

Of course, more leaks would be nice. But to your example: same is true for ARM, MIPS, PIC, AVR at high enough clocks. So no point has been made.

 

[nenforcer]

While we all wait for some more interesting news on AMD Zen processors and AM4 Promontory chipset

It's has been announced that Zendaya

will be the official spokesperson for the upcoming chips. You heard it hear first.

 

[The Stilt]

Based on the Polaris overclocking and power draw results, the 14nm LPP process looks to be significantly worse than I expected (and my expectations were non-existent to begin with) :'(

The reference cards are at least somewhat limited by the cooling, however the < 1350MHz (average, based on reviews) Fmax is still extremely disappointing. Power delivery wise the reference cards are absolutely fine since they feature a six phase digital VRM. Also a card featuring GPU this small, running at clocks as low as ~1266MHz should not violate the PCI-E device power delivery specifications D: I'm mentally fully prepared to see Zen failing,l however I'll still be extremely gutted if it happens because of the manufacturing process. And frankly I don't see any other way Zen to fail, since I'm pretty confident about the design itself.

 

[coercitiv]

I'm mentally fully prepared to see Zen failing,l however I'll still be extremely gutted if it happens because of the manufacturing process.

Their power consumption numbers simply do not match their perf/w claims.

Starting from now, any new claim they make about Zen better be backed by a live demo.

 

[ShintaiDK]

Based on the Polaris overclocking and power draw results, the 14nm LPP process looks to be significantly worse than I expected (and my expectations were non-existent to begin with) :'(

The reference cards are at least somewhat limited by the cooling, however the < 1350MHz (average, based on reviews) Fmax is still extremely disappointing. Power delivery wise the reference cards are absolutely fine since they feature a six phase digital VRM. Also a card featuring GPU this small, running at clocks as low as ~1266MHz should not violate the PCI-E device power delivery specifications D: I'm mentally fully prepared to see Zen failing,l however I'll still be extremely gutted if it happens because of the manufacturing process. And frankly I don't see any other way Zen to fail, since I'm pretty confident about the design itself.

Welcome to the world of reality

 

[ShintaiDK]

Their power consumption numbers simply do not match their perf/w claims.

Starting from now, any new claim they make about Zen better be backed by a live demo.

Saying they outright lied would be an understatement. Its FX all over.

 

[Arachnotronic]

Saying they outright lied would be an understatement. Its FX all over.

I never thought that Polaris would match Pascal in terms of performance/watt since NVIDIA has had a pretty solid lead there. But a 14LPP product, given how nice FinFETs are, failing to compete on performance/watt with a TSMC 28nm part is really unfortunate and disappointing.

I wouldn't put the blame entirely on the process. As I have said on these forums before, 16FF+ is a more efficient process than 14LPP, but GTX 1080 is ~82% more efficient than the RX 480 according to TPU. The majority of that delta likely comes from architecture, not process.

 

[antihelten]

I never thought that Polaris would match Pascal in terms of performance/watt since NVIDIA has had a pretty solid lead there. But a 14LPP product, given how nice FinFETs are, failing to compete on performance/watt with a TSMC 28nm part is really unfortunate and disappointing.

I wouldn't put the blame entirely on the process. As I have said on these forums before, 16FF+ is a more efficient process than 14LPP, but GTX 1080 is ~82% more efficient than the RX 480 according to TPU. The majority of that delta likely comes from architecture, not process.

As you say it looks to be mainly an issue with the Polaris architecture not the 14LPP process.

GTX 1080 is 60% more efficient than the GTX 980.
RX 480 is 76% more efficient than the 285*.

So it would seem that AMD has gotten at least as big an increase from their node jump as Nvidia (just doesn't help much when they were as far behind as they were in the first place).

*This is based on TPU's numbers, they don't have any power numbers for the 380 unfortunately, so I used the 285 instead.

 

[KTE]

Reality is enthusiasts always make sweeping oversimplified comparisons and look at process to create hugely unrealistic power/performance metrics.

14nm or even if they had 7nm, wouldn't mean anything. They need Skylake clocks and power brackets at launch, or, we'll be looking at another too little too late product.

It needs to be on par IPC wise but 25% short on clocks and power to just match the Phenom launch. I personally cannot see it happening.

Sent from HTC 10
(Opinions are own)

 

[The Stilt]

Welcome to the world of reality

I don't think I've ever been anywhere else :sneaky:

 

[The Stilt]

There appears to be huge variations in the leakage between the different Polaris 10 ASIC specimens. The factory programmed ASIC leakage value (SIDD) can be read from the fuse registers (interpreted as ASIC "Quality" by GPU-Z) and here is what some of the review samples had:

High = Higher SIDD (lower voltage operation, runs hotter)
Lower = Low SIDD (higher voltage operation, runs cooler)

Note: a rough definition.

Muropaketti: 87.1%
TPU: 83.0%
Sweclockers: 74.3%
Nordichardware: 75.6%
Hexus: 75.2%

The sample Muropaketti received ran so hot that the card throttled at stock, despite the fan was running at 100% speed.

If TPUs sample was already exceeding the PCI-E power specifications by > 11%, I wonder by how much do the highest leaking samples violate it

Also the reviewed cards (at least TPU) weren't even using prototype or ES silicon, but the very same stuff the retail cards are shipped with. This is pretty unusual since it's been a good while since I've seen a day 1 review featuring anything but ES silicon. Sometimes even the very first retail cards have ES marked silicon on them.

 

[Qwertilot]

Ummm, wow. Surely they should they be testing for this sort of thing and stamping very firmly on the worst offenders?

A bit scary to think how consoles would go with this sort of stuff.

 

[KTE]

^Means poor process maturity and yields, currently.

Sent from HTC 10
(Opinions are own)

 

[Qwertilot]

I guess so

Just seems surprising that they're happy pushing something quite that out of spec out of the door. Is it just too expensive to test every individual chip? Maybe it is for such a relatively low GPU.

 

[Doom2pro]

High = Higher SIDD (lower voltage operation, runs hotter)
Lower = Low SIDD (higher voltage operation, runs cooler)

Are you trying to suggest that higher voltage leakage = higher operating voltage = runs cooler? Pretty sure you got the thermal parts backwards.

Isn't it pretty common knowledge that better ASIC quality means less voltage required to operate at stable clocks due to lower voltage leakage and thus higher ASIC quality cards run cooler due to their lower operating voltage but overclock worse than less ASIC quality due to less headroom?

The more voltage leakage the higher the operating voltage needs to be in order to compensate for the excessive voltage drop associated with lower leakage resistance and thus the higher operating temps.

Unless you are privy to knowledge that the majority is ignorant to and you would like to share and or elaborate on further?

 

[DrMrLordX]

It isn't voltage that leaks, it's current . . .

 

[sm625]

Are you trying to suggest that higher voltage leakage = higher operating voltage = runs cooler? Pretty sure you got the thermal parts backwards.

The way I interpret it is "higher voltage operation" means it is capable of reaching higher voltage at the same power/heat level.

 

[Abwx]

The more voltage leakage the higher the operating voltage needs to be in order to compensate for the excessive voltage drop associated with lower leakage resistance and thus the higher operating temps.

Higher leakage means that the transistors s threshold voltage is lower, hence they conduct (leak) more when switched off but at the same time since they are closer to the conduction state a lower voltage is required to make them switch as fast (conduct the same current) as low leakage siblings.

 

[IEC]

There appears to be huge variations in the leakage between the different Polaris 10 ASIC specimens. The factory programmed ASIC leakage value (SIDD) can be read from the fuse registers (interpreted as ASIC "Quality" by GPU-Z) and here is what some of the review samples had:

High = Higher SIDD (lower voltage operation, runs hotter)
Lower = Low SIDD (higher voltage operation, runs cooler)

Note: a rough definition.

Muropaketti: 87.1%
TPU: 83.0%
Sweclockers: 74.3%
Nordichardware: 75.6%
Hexus: 75.2%

The sample Muropaketti received ran so hot that the card throttled at stock, despite the fan was running at 100% speed.

If TPUs sample was already exceeding the PCI-E power specifications by > 11%, I wonder by how much do the highest leaking samples violate it

Also the reviewed cards (at least TPU) weren't even using prototype or ES silicon, but the very same stuff the retail cards are shipped with. This is pretty unusual since it's been a good while since I've seen a day 1 review featuring anything but ES silicon. Sometimes even the very first retail cards have ES marked silicon on them.

EightySix Four on this forum has three RX 480s with ASIC scores of 70%, 77%, and 83%.

GPU-Z BIOS dump here:
https://1drv.ms/u/s!AqniqrcgThMVr6Io8ncCCsExJ67YZQ

 

[Doom2pro]

Higher leakage means that the transistors s threshold voltage is lower, hence they conduct (leak) more when switched off but at the same time since they are closer to the conduction state a lower voltage is required to make them switch as fast (conduct the same current) as low leakage siblings.

So transistors that leak more (I assume you mean Drain-Source leakage) in cutoff state have a lower threshold voltage (Saturation voltage).

Is that unique to FinFETs or is that true for all MOSFETs? I always assumed MOSFETs with a lower threshold voltage had lower RDS(on) as well as lower DS leakage.

 

[KTE]

^Vth is beside the point here.

Higher leakage chips require lower voltage to fit within the TDPs. It's one way we can spot them, and what The Stilt is talking about. It's been this trend for at least 15 years now.

Sent from HTC 10
(Opinions are own)

 

[The Stilt]

Are you trying to suggest that higher voltage leakage = higher operating voltage = runs cooler? Pretty sure you got the thermal parts backwards.

Isn't it pretty common knowledge that better ASIC quality means less voltage required to operate at stable clocks due to lower voltage leakage and thus higher ASIC quality cards run cooler due to their lower operating voltage but overclock worse than less ASIC quality due to less headroom?

The more voltage leakage the higher the operating voltage needs to be in order to compensate for the excessive voltage drop associated with lower leakage resistance and thus the higher operating temps.

Unless you are privy to knowledge that the majority is ignorant to and you would like to share and or elaborate on further?

Lower voltage is only better if the actual power / current draw is also lower. And in case of high leaking semiconductors they never are.

Higher leakage ASICs require lower voltage to operate and generally have significantly better voltage scaling than ASICs with lower leakage characteristics. The trouble is that since they consume the same or even slightly higher amounts of power as the ASICs with low leakage characteristics, the currents will be higher. The higher currents cause temperature rise not just within the ASIC itself but throught the entire system. Higher current draw will stress the power delivery further and increase the power consumption by resulting in a lower conversion efficiency and in higher conduction losses. On ASICs with lower leakage characteristics it is completely the other way around. They require higher voltages to operate, but draw significatly lower amounts of current and therefore run significantly cooler.

A ASIC with high leakage characteristcs is only desireable when you have basically an infinite amount of cooling (i.e phase change, LN2) and power delivery capacity available and the ASIC in question has a certain Vmax you need to work with. Usually the absolute voltage is relatively similar between the highest and the lowest leaking ASICs. Due their worse voltage scaling, the lower leaking ASICs might actually run into the Vmax barrier before reaching their Fmax.

For the normal consumers only products with low or average leakage characteristics are desireable, as it provides the best overall system efficiency and the lowest temperatures.