[News] AMD Shares ZEN info and talks about based 32 core 64 thread product

[vario]

Looks like AVX instructions are indeed executed at half rate...

http://www.pcper.com/reviews/Process...ction-and-More

That's a very unfortunate design decision, IMO.

Well, i dont know, i think it may be good for us.
Zen still has 8 real cores, so even if they process AVX at half the speed, still it gives similar performance as 4 core intel.
AVX isnt used very much so it shouldnt be a BIG problem ,but its one more thing that makes me feel they will price zen against intel mid end aka not BD-E.Anyhow, we still dont know the real performance penalty as zen microarchitecture is in some ways different then intels.It also has bigger caches.

[qcmadness]

Looks like AVX instructions are indeed executed at half rate...

http://www.pcper.com/reviews/Process...ction-and-More

That's a very unfortunate design decision, IMO.

http://www.anandtech.com/show/10591/...-parallelism/4

The FP Unit uses four pipes rather than three on Excavator, and we are told that the latency in Zen is reduced as well for operations (though more information on this will come at a later date). We have two MUL and two ADD in the FP unit, capable of joining to form two 128-bit FMACs, but not one 256-bit AVX. In order to do AVX, the unit will split the operations accordingly. On the counter side each core will have 2 AES units for cryptography as well as decode support for SSE, AVX1/2, SHA and legacy mmx/x87 compliant code.

Even weaker than half-speed.

[zalbard]

Zen still has 8 real cores, so even if they process AVX at half the speed, still it gives similar performance as 4 core intel.

I suspect that the power consumption may have also influenced AMD's decision to limit AVX capabilities of the chip.

[Darakian]

2 AES units per core... so all 16 threads should be able to be doing heavy crypto work at the same time. Me like

[drmrlordx]

According to some work The Stilt did comparing Blender results with and without AVX2 enabled, AVX2 support is not currently doing much for Blender performance. So I retract my statements earlier in this thread that indicated that Summit Ridge challenging Broadwell-E in a Blender benchmark was somehow indicative that Summit Ridge had acceptable AVX/AVX2 performance.

Despite Blender having support for AVX2 officially, it's not clear how that helps with anything.

[qcmadness]

http://dresdenboy.blogspot.hk/2016/0...hot-chips.html

Guess this is one of the reason that Zen cannot do native AVX for each of the FPUs because it is derived from Jaguar FPUs (?).