Question Zen 6 Speculation Thread

[adroc_thurston]

If they ever want a Zen classic + dense part like the cursed Medusa point SKU.

Dawg, PHX2 onwards is a mix of Classic and Dense.

 

[Nothingness]

No? what if you throw SIMD heavy code at the problem of large vectors of 512bit they would be stupidly slow they are good for client code like i have said not for SIMD code

I didn't say otherwise. Now what widely used client code benefits from 512-bit vectors that isn't already accelerated by some external block? There are obviously niche workloads that benefit from such wide vectors that aren't HW accelerated, but they are niche. I'm pretty sure AMD added AVX-512 only for some HPC workloads, and having them in client was just a side effect.

 

[511]

Dawg, PHX2 onwards is a mix of Classic and Dense.

medusa point 4 different PnP cores will surely be easy to schedule

 

[511]

I didn't say otherwise. Now what widely used client code benefits from 512-bit vectors that isn't already accelerated by some external block? There are obviously niche workloads that benefit from such wide vectors that aren't HW accelerated, but they are niche. I'm pretty sure AMD added AVX-512 only for some HPC workloads, and having them in client was just a side effect.

VAES? Crypto that everyone uses more or less. JSON parsing ? the issue with AVX-512 is the availability rate is low so no one targets it on client.

 

[poke01]

Crypto that everyone uses more or less

For client use cases you currently don’t the massive throughput advantage that AVX-512 provides.

There’s some encryption benchmarks below, a M4 vs Zen4 desktop comparison. Where AVX2 or AVX512 is used it’s a massive advantage for AMD. Where it’s not used, the M4 wins.

GitHub - hiae-aead/benchmarks: Benchmark of high-performance authenticated encryption algorithms.

Benchmark of high-performance authenticated encryption algorithms. - hiae-aead/benchmarks

github.com

JSON parsing ?

this isn’t a client workload lol.

 

[adroc_thurston]

this isn’t a client workload lol.

It is.
Not the most frequent one, though.

 

[gdansk]

this isn’t a client workload lol.

Don't check network requests in your browser right now for this very website.

 

[poke01]

Don't check network requests in your browser right now for this very website.

Do you need AVX-512 for that?

Edit: let me rephrase, do people who buy client CPUs ever need the performance that AVX-512 provides for JSON parsing e.g simdjson?

 

[gdansk]

Do you need AVX-512 for that?

It'll be faster for larger objects, however.

 

[gdansk]

Edit: let me rephrase, do people who buy client CPUs ever need the performance that AVX-512 provides for JSON parsing e.g simdjson?

I'll add that it's one of those elements - like TLS acceleration, gzip vectorization, etc - that are very helpful for servers so it'll be pursued there first. Yet it can add small wins in client browsers (in decade or so when AVX512 is actually ubiquitous).

 

[511]

Do you need AVX-512 for that?

Edit: let me rephrase, do people who buy client CPUs ever need the performance that AVX-512 provides for JSON parsing e.g simdjson?

Clearly you have never used GB of JSON 🤣🤣

 

[Joe NYC]

medusa point 4 different PnP cores will surely be easy to schedule

By that, do you mean Medusa Premium?

They probably have some "escalation" algorithm.
1. keep everything on LP cores
2. add dense cores
3. add classic cores
4. switch everything to CCD
5. if all 12 cores of CCD at 100% utilization, add SoC cores.

 

[Nothingness]

Clearly you have never used GB of JSON 🤣🤣

Perhaps because no client workload needs to parse GB of json files? And even then, parsing is only part of what you do with your json. Again I'm not trying to dismiss how useful wide vectors can be, but I think it's relevant for cases that don't matter that much for most people; I'm not in that case, and I blame Intel for having slowed down AVX2 and even more AVX-512 adoption with their braindead market segmentation, as much as I blame Arm and Apple for not supporting SVE more, but at the same time I understand their reasoning.

 

[ETI4711]

Perhaps because no client workload needs to parse GB of json files?

Why would anyone bother using AVX-512 when many CPUs don’t support it at all?

I blame Intel for having slowed down AVX2 and even more AVX-512 adoption with their braindead market segmentation

This, alongside Intel’s initial suboptimal implementations, is the reason why AVX-512 has been so slow to catch on.

 

[Win2012R2]

Why would anyone bother using AVX-512 when many CPUs don’t support it at all?

And many CPUs now support it very nicely, so software runs more efficiently, which can safe battery life too, so why not?

Base libraries also get updated to use more efficient path if correct CPU is detected, so software that uses them will automatically benefit.

 

[511]

Perhaps because no client workload needs to parse GB of json files? And even then, parsing is only part of what you do with your json. Again I'm not trying to dismiss how useful wide vectors can be, but I think it's relevant for cases that don't matter that much for most people; I'm not in that case, and I blame Intel for having slowed down AVX2 and even more AVX-512 adoption with their braindead market segmentation, as much as I blame Arm and Apple for not supporting SVE more, but at the same time I understand their reasoning.

I do as well there shouldn't be non AVX-2 Processor still in manufacturing like AVX2 Should be 100% on all newly launched CPU.

 

[DrMrLordX]

Pardon my ignorance but in the future, won't software developers be targeting something like AVX10 or whatever anyway?

 

[adroc_thurston]

Pardon my ignorance but in the future, won't software developers be targeting something like AVX10 or whatever anyway?

AVX10 is just AVX512 prettymuch.

 

[511]

AVX10 is just AVX512 prettymuch.

that makes the dev's life easier

AVX10 is just AVX512 prettymuch.

also APX as well but it would take at least 5 Year+

 

[adroc_thurston]

also APX as well but it would take at least 5 Year+

APX is a meme.

 

[Nothingness]

I do as well there shouldn't be non AVX-2 Processor still in manufacturing like AVX2 Should be 100% on all newly launched CPU.

You can check yourself: it took Intel almost 10 years from the first AVX2 CPU (Haswell 2013) until all of their CPU had support for it (Gracemont 2021). As I wrote, this is what slowed down adoption, and for AVX-512 it's even worse.

 

[adroc_thurston]

You can check yourself: it took Intel almost 10 years from the first AVX2 CPU (Haswell 2013) until all of their CPU had support for it (Gracemont 2021). As I wrote, this is what slowed down adoption, and for AVX-512 it's even worse.

Well this is primarily down to Atom guys being really really annoying about it.

 

[511]

APX is a meme.

It's not if it was a meme AMD would have pushed back

 

[511]

You can check yourself: it took Intel almost 10 years from the first AVX2 CPU (Haswell 2013) until all of their CPU had support for it (Gracemont 2021). As I wrote, this is what slowed down adoption, and for AVX-512 it's even worse.

AVX-512 was supposed to be mainstreamed but ala 10nm happened.

 

[adroc_thurston]

It's not if it was a meme AMD would have pushed back

They did!
But a lot of times, you gotta compromise.