[News] AMD 2nd Gen Ryzen Threadripper MCM Configuration Confirmed, More Details

[StyM]

https://www.techpowerup.com/246557/a...d-more-details

Tech Day slides leaked to the web by Kitguru provide a confirmation of how AMD is wiring out the additional two dies on the 24-core and 32-core Ryzen Threadripper 2000-series MCMs on the socket TR4 platform. We had speculated that because the quad-channel DDR4 memory and PCIe interfaces are wired to two diagonally-opposite dies on AMD X399 chipset motherboards; in the interest of backwards compatibility, AMD could wire out memory and PCIe from just two out of four dies on the multi-chip module, and have the two additional dies seek memory and PCIe over the InfinityFabric interfaces.

The obvious trade-offs with this design choice is that latencies to the dies with indirect memory/PCIe access are higher, and that reflects heavily in AMD's own performance figures for comparing the 32-core 2990WX with the 16-core 2950X from the same generation. The 2990WX is "only" up to 64 percent faster than 2950X at Cinebench R15 nT, despite having double the number of cores. To its credit, the 2950X has higher clock-speeds (3.50 GHz nominal with 4.40 GHz boost) than the 2990WX (3.00 GHz nominal with 4.00 GHz boost). The presentation also puts out interesting bits of information such as AMD's own performance numbers showing 10-15 percent performance gains between the 2950X and the 1950X; and performance gains of the 2990WX over Intel Core i9-7980XE.

[kromosto]

first slide is really mind blowing. graphs says 2950 is faster than 7980 but the writings says 7980 is faster than 2950 when we compare them to 2990...

[Particle]

first slide is really mind blowing. graphs says 2950 is faster than 7980 but the writings says 7980 is faster than 2950 when we compare them to 2990...

Maybe you misread some model numbers? I'm not finding what you are referring to.

[kromosto]

Maybe you misread some model numbers? I'm not finding what you are referring to.

if I am not missing something very obvious which is very possible for me here is what I see

On the graph 7900x = 1 2950x = 1.41 which means it is %41 better performing than 7900x and it is also written on the left side. 7980xe = 1.38 which means it is performing %38 better than 7900x but it is slower than 2950x because it is %41 better. But on the left it is written that 2990x is %51 better than 7980xe and %64 better than 2950x.

Also if you divide the numbers on the graph to find the given percentages it is also not possible.

2950x/7900x = 1.41/1 = 1.41 = 2950x is %41 better than 7900x
2990x/7980xe = 2.33/1.38 ~= 1.68 = 2990x is %68 better than 7980xe
2990x/2950x = 2.33/1.41 ~= 1.65 = 2990x is %65 better than 2950x (they write %64 so this is also acceptable)

if we change 1.38 to 1.54 then everything is corrected. Again the difference can be negligible but if I am not wrong 7980xe is performing a little better than 1.54 against 7900x in cinebench.

[Particle]

Ok, I think I understand the confusion now. Take note that the slide is showing multiple comparisons. The numbers on the left aren't referring to the Cinebench results on the right. The Cinebench scores are already normalized, so there wouldn't be any need to write out the results in text for a percentage comparison. Each performance comparison on the left is qualified with a superscript that refers to an end note describing the conditions that comparison was made under.

[kromosto]

Ok, I think I understand the confusion now. Take note that the slide is showing multiple comparisons. The numbers on the left aren't referring to the Cinebench results on the right. The Cinebench scores are already normalized, so there wouldn't be any need to write out the results in text for a percentage comparison. Each performance comparison on the left is qualified with a superscript that refers to an end note describing the conditions that comparison was made under.

While %41 and %64 are similar with the graph results, what you are saying can be right but a condition that 2950x beats 7980xe on Cinebench nT as graph shows must be a very very special condition.

[kromosto]

And also don't get me wrong I am not trying to disprove your point or trying to argue just hate these nonsense marketing slides with obvious and intentional MISTAKES.