Tangent from 1660Ti Reviews Thread

[AtenRa]

Compared again to Vega VII's jump over Vega 64, which had an entire node increase to use for efficiency improvements, TU116 it actually did a better job at improving efficiency over GP106, which was already an extremely efficient design.

Vega 20 is just a Vega 10 with a few adjustments ported from 14nm GloFo to 7nm TSMC. Transistor count between Vega 10 and Vega 20 is almost the same (12.5B vs 13.23B). Even with the same transistor count and less CUs than Vega 64 , the Radeon VII managed to gain 25% higher gaming performance at lower power than Vega 64.

Not only that but if you directly compare Vega 20 to TU104 they are competing in every metric. Same amount of Transistors, Same gaming performance, Vega 20 has higher compute performance FP64(except Tensor matrix) and TU104 may have higher RT performance (we still dont know how fast Vega 20 can do DXR).

On the other hand , the TU 116 (GTX1660Ti) has 50% more transistors than GP106 (GTX1060). Its 2.5 year later and uses 12nm vs 16Nm. After 2.5 years NVDIA managed to reach parity with Polaris in FP16 (mm2) and still have way lower FP64 throughput. Not to mention Polaris has less transistors than TU116 (5.7B for Polaris vs 6.6B for TU116).

From now on, both AMD and NVIDIA are allocating more and more transistors for compute (Machine learning etc) , so the perf/mm2 in Gaming will go down , much like Vega and Turing cards.




Moved all this tangential crud out of the 1660 review thread.

AT Moderator ElFenix

 

[insertcarehere]

Vega 20 is just a Vega 10 with a few adjustments ported from 14nm GloFo to 7nm TSMC. Transistor count between Vega 10 and Vega 20 is almost the same (12.5B vs 13.23B). Even with the same transistor count and less CUs than Vega 64 , the Radeon VII managed to gain 25% higher gaming performance at lower power than Vega 64.

A bit more than 25% perf/Watt improvement for what is effectively 2 process node jumps.

In an environment where Qualcomm can make their 7nm SD855 something like 50% perf/Watt better than 10nm SD835 and still get their fair share of criticism. It takes some seriously warped logic to think that Radeon VII was a good product given the circumstances.

 

[AtenRa]

A bit more than 25% perf/Watt improvement for what is effectively 2 process node jumps.

One process jump from GF14nm to TSMC 7nm.

Perf/watt will be 25% or 50% vs Vega 64 depending on performance.

Its 25% higher perf at the same power or
50% less power at the same perf.

All this with the same architecture and almost same number of transistors, you only get what 7nm gives you over GF 14nm.
Also, Vega 20 has even more perf/watt vs Vega 10 and TU104 in many compute workloads.

 

[sontin]

Wow, do you really need to defend a worse product? TU116 is 40% more efficient than GP106 while having 50% more transistors. And this
is without 7nm. Vega7 has 3,2% more transistors and is only ~35% better.
And TU116 uses 50% less power at the same perf, too.

And no, Vega20 has no better perf/w in "many compute workloads". In fact it is even multiple times worse in DL and Raytracing workloads. It is the definition of medicore.

 

[AtenRa]

And no, Vega20 has no better perf/w in "many compute workloads". In fact it is even multiple times worse in DL and Raytracing workloads. It is the definition of medicore.

Get your facts straight

https://www.anandtech.com/show/13923/the-amd-radeon-vii-review

Also,

https://www.phoronix.com/scan.php?page=article&item=radeon-vii-linux&num=1

Here is a look at the overall AC system power consumption during the course of the various OpenCL benchmarks carried out for this launch-day testing. The power efficiency of the 7nm Vega was extremely good with a total AC system power average of 125 Watts and a peak of 337 Watts.

 

[railven]

ITT: GTX 1660 Ti so good, Radeon 7 is being defended.

Why is anyone even talking about Radeon 7 in a <$350 card topic?



On the comparison of the RX 580 - my god are these little things loud. My RX 580 Red Devil is LOUDER than my Powercolor Radeon 290X. I don't even know how that is possible. Eitherway, $250 would have been a good spot for this card (GTX 1660 Ti). Nice card, at $280 with no freebies, I'd just pay up for the RTX 2060 Ti personally.