Question Zen 6 Speculation Thread

[poke01]

For actual real world loads that need these operations (not Geekbench), I believe (we will see) the iGPU in Zen 6 will outperform AMX in Intel CPU's.

sometimes i feel as though we repeat this a lot but here we go again

for matrix workloads SPEED: GPU > CPU
for matrix workloads BOARD COMPATIBILITY: CPU > GPU

i don't get why this has be a competition. Both can be useful but for 99% of usecases GPU wins.
Why? cause speed always wins

 

[adroc_thurston]

for matrix workloads BOARD COMPATIBILITY: CPU > GPU

Ehhhh.
CUDA/ROCm are broadly compatible and usually have high perf kernels available.
The rest, OTOH.

 

[MS_AT]

for matrix workloads SPEED: GPU > CPU

Depends on the case. For example in 9950x case the gpu will be slower than the cores but if you have big enough matrix and big enough gpu, then for sure.

 

[OneEng2]

Depends on the case. For example in 9950x case the gpu will be slower than the cores but if you have big enough matrix and big enough gpu, then for sure.

I guess the question I am posing is: Will next gen iGPU's be powerful enough that they render APX pointless?

I am currently of the belief that they will.

 

[Win2012R2]

Will next gen iGPU's be powerful enough that they render APX pointless?

AMX

RDNA5 likely, but since AMD appear to plan to shift crappy RDNA35 for years to come that might be a problem, Intel will ship nvidia iGPUs.

 

[MS_AT]

Will next gen iGPU's be powerful enough that they render APX pointless?

Depends on the SKU i guess. I wouldn't count on that for direct desktop chips replacements, the igpu there is not meant for anything more than displaying desktop and handling hw video decode. The premium mobile parts are supposed to be GPUs with attached CPUs, so there for sure

Also keep in mind that with AMX you don't leave the CPU, it might be quite competitive with pushing stuff to dGPU for right problem sizes. Or it can be used to pump Geekbench scores. Seeing people obsess over GB or CB it makes sense from marketing point of view

 

[OneEng2]

My bad (AMX) .

 

[reaperrr3]

AMX

RDNA5 likely, but since AMD appear to plan to shift crappy RDNA35 for years to come that might be a problem, Intel will ship nvidia iGPUs.

I suspect that AMD will simply (begrudgingly) price Medusa Premium (with AT4) a bit lower than originally planned if they have to, and that's that.
RDNA3.5-only SoCs are mostly for office users or to be paired with a dGPU.

Besides, NVL will still be Xe3(P), I wouldn't expect NV iGPUs before 2028/2029 and that's probably close to the release of the Medusa base SoC's successor with RDNA5+, nvm that it'll probably have AMX via Zen7, too.

AMD's choice to stick with RDNA3.5 for mainstream looks wrong, but more so over its anemic gaming perf vs. the bigger Intel Xe3(P) iGPUs.
Not sure if putting only 8 CUs into the Medusa base SoC was a good idea, regardless of higher clocks.

 

[adroc_thurston]

I suspect that AMD will simply (begrudgingly) price Medusa Premium (with AT4) a bit lower than originally planned if they have to, and that's that

No, it's a discrete compete part at discrete prices.

AMD's choice to stick with RDNA3.5 for mainstream looks wrong, but more so over its anemic gaming perf vs. the bigger Intel Xe3(P) iGPUs.
Not sure if putting only 8 CUs into the Medusa base SoC was a good idea, regardless of higher clocks.

Cost cost cost cost cost.