AMD: Navi Speculation, Rumours and Discussion [2019-2020]

[Cat Merc]

Still haven't seen anything regarding actual power consumption in any of the documentation...

Can anyone make sense of this slide along with the endnotes?

View attachment 3149


It would appear that performance/watt part was based on a Division 2 run at 2560x1440 Ultra settings, with a 40CU Navi vs a "Vega 64" with 40CUs enabled. So that 14% more performance doesn't seem great vs a 40CU Vega 64 but the 23% less power is interesting if a Vega56 is around 210w.
Edit- the 14% increase in performance is the exact same as the clock difference between Navi's gaming clockspeed and Vega 64's boost clockspeed...

Then the area comparison of a 14nm "Vega10" is Vega 56 vs Navi. (I assume they used a 40CU Navi for bigger numbers rather than the 36CU Navi)

It's 64 CU Vega vs 40 CU Navi.

Power calculated from Vega 64 results -23% also appears to be 225~W, which is exactly what AMD stated as the TBP.

So as it says, 14% higher perf, 23% lower power, results in 1.5x perf/watt gain over Vega 64.

Perf/mm^2 increased by 20% over Vega VII, which is impressive considering smaller chips tend to be worse in that regard due to the media engines and other blocks being a fixed cost, and I didn't calculate it but from eyeballing the GDDR6 PHY's appear to be a bit larger than Navi's?

That said Vega VII also has half rate DP, so hard to compare.

Perf/watt appears to be 10% or so better than Vega VII.

 

[DavidGraham]

It would appear that performance/watt part was based on a Division 2 run at 2560x1440 Ultra settings,

I am afraid one game shouldn't be enough.

The end notes had several other caveats, they never specified which Navi GPU they are comparing to. They also tested some 3D mark 11 tests @ 720p.

It's 64 CU Vega vs 40 CU Navi.

Not it's not, it's a Vega 64 with 40 CUs enabled. It states so quite clearly in the endnotes.

 

[anexanhume]

Random things I noticed on later skimming of the LLVM commit I linked to earlier:
While the dot product instructions do have a lot of commonality between GFX1011 and GFX1012, there are some differences like GFX1012 showing extra mentions for atomic instructions that the other GFX10 variants did not. This may have more importance in a compute scenario.

GFX1011 has a sparse set of entries for errors with some instructions and little documentation. I suppose it'd be fun to mention a few lines here:

Since this is a skeleton of an error file, I am not sure what prompted their mention here at this time.

It’s a hat tip. They know people are looking.

 

[Cat Merc]

I am afraid one game shouldn't be enough.

The end notes had several other caveats, they never specified which Navi GPU they are comparing to. They also tested some 3D mark 11 tests @ 720p.

Not it's not, it's a Vega 64 with 40 CUs enabled. It states so quite clearly in the endnotes.

Where?

Considering the performance they're claiming for 5700XT, it would make zero sense for that to be anything but a full Vega 64.

 

[DavidGraham]

Where?

here:

AMD Next Horizon: Gaming PDF

Considering the performance they're claiming for 5700XT, it would make zero sense for that to be anything but a full Vega 64.

and yet here we are

 

[Cat Merc]

here:


and yet here we are

That's... Not good looking for performance. Good looking for power though.

Maybe the "Navi GPU" in that slide isn't a 5700XT, but rather some other configuration at a more efficient point in the voltage/frequency curve, which would explain the power difference vs a 40CU Vega 64.
23% lower than a full Vega 64 would be 225W as it is, if that's a 40CU Vega 64 then power would have to be lower.

 

[fehu]

Wait! We now have the responses.

 

[Kaotik]

Ca you provide any context around the graphs provided in the second one. Are they showing showing a comparison of relative performance between the two systems listed, and if so is there anything noteworthy about the tests?

Which one exactly did you mean?

 

[Gubbi]

AMD claim 2.3 x performance per area, a Vega 64 is 495mm², Navi us 251mm².

That comes out to 5700XT (40 RDNA CUs) being 16% faster than a Vega 64 (64 GCN CUs).

Cheers

 

[LordEC911]

AMD claim 2.3 x performance per area, a Vega 64 is 495mm², Navi us 251mm².

That comes out to 5700XT (40 RDNA CUs) being 16% faster than a Vega 64 (64 GCN CUs).

Cheers

Even more confusion... end notes says they compared benchmark scores of "Navi" vs Vega56 and had the Vega 10 die size at 486mm2, but the slide says 495mm2...
Who the heck fact checks their presentations?
Are they cutting out 9mm2 for the Vega56's disabled parts?

 

[del42sa]

AMD claim 2.3 x performance per area, a Vega 64 is 495mm², Navi us 251mm².

That comes out to 5700XT (40 RDNA CUs) being 16% faster than a Vega 64 (with 40 GCN enabled CUs).

Cheers

fixed it for you

.....had the Vega 10 die size at 486mm2, but the slide says 495mm2...

which is more close to reality

https://forum.beyond3d.com/posts/2046674/

 

[DavidGraham]

Maybe the "Navi GPU" in that slide isn't a 5700XT, but rather some other configuration at a more efficient point in the voltage/frequency curve

They mention the configuration of Navi is 40CU, so it is the 5700XT, but maybe with different clocks.

New AMD claim 2.3 x performance per area, a Vega 64 is 495mm², Navi us 251mm².

Nope, the comparison was made against a Vega 56 (486 mm2), while running some weird tests like 3D Mark 11 @720p.

Navi scored 140 (251 mm2)
V56 scored 113 (486 mm2)

This is where you get your 2.3X per/area.

 

[LordEC911]

fixed it for you

Nope, the Vega64 with 40CUs is only for the perf/watt metric.
The perf/area metric is against a Vega56.

which is more close to reality

https://forum.beyond3d.com/posts/2046674/

Well at least that part sorta makes sense now.

 

[del42sa]

Nope, the Vega64 with 40CUs is only for the perf/watt metric.
The perf/area metric is against a Vega56.

Vega 56 or Vega 64 doesn´t matter if it comes to chip area, because both dies are physically same size

 

[LordEC911]

Vega 56 or Vega 64 doesn´t matter if it comes to chip area, because both dies are physically same size

It does when a person is trying to extrapolate Navi performance with a performance/area metric.

 

[DavidGraham]

A more revealing set of benches, from the slides ..

FireStrike GT1 @1440p:
Navi: 49
Vega56: 41

Navi is 19% faster

FireStrike GT2 @1440p:
Navi: 37
Vega56: 32

Navi is 15% faster

Unigine Heaven @1080p:
Navi: 84
Vega56: 72

Navi is 16% faster

In these common artificial tests, Navi is barely any faster than a Vega 64.

 

[iroboto]

A more revealing set of benches, from the slides ..

FireStrike GT1 @1440p:
Navi: 49
Vega56: 41

Navi is 19% faster

FireStrike GT2 @1440p:
Navi: 37
Vega56: 32

Navi is 15% faster

Unigine Heaven @1080p:
Navi: 84
Vega56: 72

Navi is 16% faster

In these common artificial tests, Navi is barely any faster than a Vega 64.

Seems fairly significant. Barely is a word
I use for 1-2% difference.

15-20% is a fairly significant difference.

 

[DavidGraham]

15-20% is a fairly significant difference.

These results are against a Vega 56, my statement is for Vega 64, which is usually 15% faster than a Vega 56.

 

[Globalisateur]

A more revealing set of benches, from the slides ..

FireStrike GT1 @1440p:
Navi: 49
Vega56: 41

Navi is 19% faster

FireStrike GT2 @1440p:
Navi: 37
Vega56: 32

Navi is 15% faster

Unigine Heaven @1080p:
Navi: 84
Vega56: 72

Navi is 16% faster

In these common artificial tests, Navi is barely any faster than a Vega 64.

Are those tests done with the same number of CUs and same frequency ?

 

[Rootax]

These results are against a Vega 56, my statement is for Vega 64, which is usually 15% faster than a Vega 56.

Is this surprising ? Navi 10 is a mid range gpu. At least it seems too match with previous amd "kind of" high end. And with that price, it the least they can propose.