AMD CDNA Discussion Thread

Trento + MI200?

 

Yea.

 

big boye

 

CoWoS-3?



https://videocardz.com/newz/amd-instinct-mi200-mcm-aldebaran-gpu-pictured

 

Spoiler: Big image

 

Yes, that lidded 3x reticle CoWoS option for HPC was made for MI200s.

 

boom

 

N6 process.

 

All AMD parts are N6 now

 

AWWWWWWWWWWW YIS IT'S THE FIRST -LSI part



Or maybe it's an ASE part.
Time to annoy people.

 

no cowos? EFB.

 

I think it's ASE/SPIL's FOEB.
AMD is a big-big ASE house.
https://www.3dincites.com/2020/07/iftle-456-spil-fan-out-embedded-bridge-foeb-technology/

 

This isn't a one single wholesome GPU though, It's a two die GPU and appears as two GPUs to the system, juggling data between the dies is way slower than a single monolithic die, as the link between the two dies can only handle 200GB/s between each die, so aggregate performance numbers are theoretical numbers and should not be taken at face value.

Also, on contrast to NVIDIA's A100, this isn't primarily a matrix/AI GPU, but rather is a vector GPU primarily. As a result, AI enhancements are very modest compared to A100 even when combining the two dies together .. however, vector throughput is through the roof. AMD and NVIDIA is diverging here, AMD is still not fighting NVIDIA in the AI market, rather focusing solely on the HPC market.

Power consumption is 560W and is liquid cooled, which is curious considering the 6nm process.

 

That's all DC GPUs going forward due to NUMA.
A100 programming in some cases is already that due to wonky as balls split L2.

More, like 400.

No, this thing is just purpose-built.
They'll make a blingee matrix engine when the wider market wants their stuff.

500, 560 is the actual Frontier SKU as per stuffed into that Shasta blade.

 

https://www.amd.com/en/graphics/server-accelerators-benchmarks
AMD benchmarks.

On average MI250 is up to 2.3x faster in comparison with A100 in DP workloads with 1 GPU config mostly (except for LAMMPS).
The largest gains are in LINPACK of course.
A100 still has an edge in GPU to GPU communication bandwidth, so the difference with real multi-GPU configs will likely be smaller.

 

MI200 series is old.
Cool folks are talking about MI300 already

Probably for delivery to El Cap

 

Yessss.

Obviously, big DOE heads quoted the El Cap timeframe already.
It's essentially 5Q after Frontier.

 

I wonder if they will revise the perf target given latest exascale showdown with far east contenders.
Already saw some govt. head honchos on twitter debating.

 

El Cap is like puny 33MW so throwing moar nodes at the problem is totally feasible.
It's a closed off literally simulate nukes system so do they really have to?