Question Zen 6 Speculation Thread

[poke01]

really? I jumped from a 8250u to a 185H

Me too dude, I jumped from a horse to a combustion car just last week and and the difference was massive but it looks like you went from crap to less crap.

 

[igor_kavinski]

Making a product for the one or two lunatics using a consumer CPU to transcode is an interesting business decision if real.

I have a 128 thread Zen 2 Epyc so I guess I'm lunatic²?

 

[511]

Me too dude, I jumped from a horse to a combustion car just last week and and the difference was massive but it looks like you went from crap to less crap.

when i bought the 8250u it was not crap lol i bought it in 2019 i think.

 

[StefanR5R]

[approximately 52 CPU threads on the desktop]

I wish I had a dollar for every time someone said we don't need this much of something.

Let's turn to Gene Myron Amdahl to ask what we actually gain from 52 (read 48) threads, compared to for example sixteen threads, in client workloads.
Also let's think about what our needs are, and what our wants are, and how little they are connected anymore in the so-called developed countries.

I have a 128 thread Zen 2 Epyc

Me too (among else; alas only at home but not at work). But at home, I execute Firefox and the likes on a 4c/8t Haswell, not on the 128t 16-channel dual-Rome. I am even more of a lunatic than you.

 

[Markfw]

have you seen the cost of a 64 Zen 4 thread ripper it is $4400 even accounting for $1200 next gen motherboard and $1000 NVL-S CPU I would have $2200 to spare for the rest of the build.
7980WX Amazon

Here is the W880 MB

â Supermicro X14SAE W880 ATX Motherboard, LGA1851 Socket V1, DDR5-5600MHz | eBay

Cloud Project is located in the Silicon Valley of California, USA. 2 PCIe 5.0 x16 slots, 1 PCIe 4.0 x4 slot.

www.ebay.com

X14SAE | Motherboards | Products | Supermicro

I am a bit confused here 😅.

Your question, what is the best AM5 motherboard could be answered by the best AM5 expert in high CPU usage/overclocking.

re-reading, maybe that was not the source of your comment, but at least maybe we can get an answer to the AM5 motherboard question from detx

 

[igor_kavinski]

Let's turn to Gene Myron Amdahl to ask what we actually gain from 52 (read 48) threads, compared to for example sixteen threads, in client workloads.

People don't really need the 4090/5090 but still getting it coz their perf/$ calculations tell them that the GPUs are worth it and they will last them 5 or 6 years easily. Same for these halo CPUs. They may age better than the 32 thread or lower count CPUs and end up saving the users money in the long run as they will feel less need to upgrade.

Me too (among else; alas only at home but not at work). But at home, I execute Firefox and the likes on a 4c/8t Haswell, not on the 128t 16-channel dual-Rome. I am even more of a lunatic than you.

Same situation to a tee almost!

I'm trying to see if I can pay for a cheap two socket 48 thread Haswell workstation to replace my 4770 desktop at work coz the ones in charge are buffoons in tech matters and don't want to give me a Threadripper or something with more memory channels.

My main browsing laptop has a i7-3720QM. I have a Comet Lake laptop to replace it with but need to figure out which of the RAM sticks on it is bad first as y-cruncher errors out. I can't use the Epyc because I like to use it to mess around with stuff so can't use it for my browsing needs. It would need to stay on all the time with a UPS and I would need to save my work every time I would need to restart it. But yeah, blessed are those who have a server for messing around, a server for work and a server for browsing needs!

 

[OneEng2]

I wish I had a dollar for every time someone said we don't need this much of something.

Examples:

HDD and SSD storage
CPU GHz
CPU cores
RAM (We are already at 128GB with Strix Halo and some people, including me, want more!)
GPU VRAM

Once you have experienced something, you always want more of it and because the companies want to sell thousands or millions of units of something new and more exciting than what's already on the market. I guess Intel feels threatened by 48 Zen 6 threads so they were like, hey, let's go for 52 cores instead of just threads coz they have admitted defeat that they can't do HT better than AMD's SMT.

Agree ... however, we have long since reached the point of diminishing returns with cores. Lots of programs benefit from 4-8 cores. Fewer (by a lot) benefit from 16 cores.

I'm not saying that NO ONE will pay for the top core count processors, just not very many (percentage wise in the market).

Of course, as I was saying, we may be entering another marketing war (like with Mhz). "My Dad's core count is bigger than YOUR Dad's core count"!

have you seen the cost of a 64 Zen 4 thread ripper it is $4400 even accounting for $1200 next gen motherboard and $1000 NVL-S CPU I would have $2200 to spare for the rest of the build.
7980WX Amazon

Here is the W880 MB

â Supermicro X14SAE W880 ATX Motherboard, LGA1851 Socket V1, DDR5-5600MHz | eBay

Cloud Project is located in the Silicon Valley of California, USA. 2 PCIe 5.0 x16 slots, 1 PCIe 4.0 x4 slot.

www.ebay.com

X14SAE | Motherboards | Products | Supermicro

I am a bit confused here 😅.

Trivial $ for any business.... especially if it is considered an essential tool for the job.

I have a 128 thread Zen 2 Epyc so I guess I'm lunatic²?

No Igor, not at all ..... but definitely in the minority .

 

[511]

Your question, what is the best AM5 motherboard could be answered by the best AM5 expert in high CPU usage/overclocking.

re-reading, maybe that was not the source of your comment, but at least maybe we can get an answer to the AM5 motherboard question from detx

Thanks

 

[MS_AT]

What really? I jumped from a 8250u to a 185H and the performance difference for me is pretty insane for blender or when I use FFmpeg is it possible that windows was assigning things wrong way?

Something between Windows, Dell tweaks and Intel Thread Director was prioritizing E cores when launching the builds in particular way (no need to say my favourite way). Unlocking the special knobs in registry and setting everything to prefer P cores according to Windows docu did not change a thing. So to enjoy 3x increase in performance I found myself a new favourite way to launch the builds as I did not have time to dig deeper.

 

[igor_kavinski]

Of course, as I was saying, we may be entering another marketing war (like with Mhz). "My Dad's core count is bigger than YOUR Dad's core count"!

The side effect is that some developers may decide to try to use those extra cores to sell their applications to the wealthy owners of these halo beasts so it will lead to better core utilization for everyone.

 

[igor_kavinski]

So to enjoy 3x increase in performance I found myself a new favourite way to launch the builds as I did not have time to dig deeper.

Had you or your company waited to give you an ARL-H or ARL-U laptop, you probably wouldn't have noticed the slowdown because the Skymonts are so performant. Gracemont is just ughhhhh what the heck lame ugly core matching only Skylake.

 

[OneEng2]

First to address the expectations of a MT matchup between NVL and Zen 6 24c/48t:

Intel's "mont" cores are weak with vector operations. Very weak compared to Zen 5. They may get better, but don't expect miracles. In engineering you don't get something for nothing, and their small size has come at the expense of good vector operations and SMT. I don't expect "mont" AVX10 to be even close to Zen 6 AVX512.

Also, as stated, ARL is really a 48c product. The 4 LPE will likely be intentionally idled in the scheduler in heavy MT workloads.

Still, it is not reasonable to expect that Zen 6 24c/48t will win all MT benchmarks over a 48c NVL.... but I expect it will win some.

@OneEng2 I wouldn't be so sure that consumer Zen7 is going to have the same, wide data paths of Zen5. There's a lot of rumor traffic that the DC CCDs will be different from the consumer desktop/mobile ones. While I expect full instruction compatibility, I do NOT expect that the consumer desktop CCDs will get the full fat avx512 pathways. Instead, I think that they will get the 256 but wide ones like Strix Point and Kraken have. It'll make the costs of those 12 core CCDs more manageable by keeping their size down.

Remember, there was a recent strong leak that Medusa point is going to be essentially a regular monolithic die from the bottom of the stack, with the top of the stack having a single 12 core CCD added on in a larger package. No way are they having different core logical topologies between the two.

I don't think you get DC CCDs until you get to Threadripper and above. It makes sense from a throughput point of view as the limited memory bandwidth on desktop/mobile will certainly hold AVX512 throughput back anyway.

I am assuming you meant Zen 6, and yes, I fully expect AMD to maintain its full 512b AVX path.

The monolithic die will likely have a combination of P and E cores for Zen 6 along with some LPE.

As for the memory bandwidth, I think that a new IOD and support for DDR8000 will certainly keep Zen 6 fed just fine even with AVX512.

And you may well be correct. Or they may have misinterpreted data, as was the case with the Zen5 40%+ claim. They're not going to change their tunes in the face of counter-arguments.

It seems reasonable that AMD would prioritize a newer, expensive node for high-margin products. But, up until now, even their premium products have shared CCDs with consumer desktop (exceptions being Turin dense and Bergamo). Zen6 may change things a bit.

Oh, I was mostly kidding. I should have put an indicator. I could easily be wrong and MLID could be right.

It just doesn't make sense to me, and it didn't make sense to AMD with Zen 5.

Just a note to all.....

Just because a companies COST goes up does NOT mean that they can raise their price. Price is determined by the market demand and the supply.

Right now, there is plenty of supply of CPU's. People are getting squeezed by inflation, and a "fast CPU" is definitely a luxury item. Luxury items are not purchased by the masses when times are tough.

I don't expect to see the price for the top end processors go up in the next generation.

 

[MS_AT]

Had you or your company waited to give you an ARL-H or ARL-U laptop, you probably wouldn't have noticed the slowdown because the Skymonts are so performant. Gracemont is just ughhhhh what the heck lame ugly core matching only Skylake.

I did not have a choice to wait, only to pick Meteor Lake which was slower. And the reason I started digging was that my compiler benchmark was giving the expected boost but real builds did not.

think that a new IOD and support for DDR8000 will certainly keep Zen 6 fed just fine even with AVX512.

For specific algos with dataset fully fitting in L1 cache I can get about 250-300 GB/s running on single core. When running from memory it drops to 30GB/s at best. So nope, it won't be well fed anytime soon, seeing single CCD has 8 cores and soon 12. But of course every improvement helps.

 

[Joe NYC]

If Medusa Halo has quad channel 256GB 10,000 MT/s RAM in addition to being attractively priced with PCIe 5.0 x16 slot, it will solve multiple problems at once:

1) Superb iGPU

2) Enough bandwidth for LLMs and AVX-512 workloads

3) No buying DIMMs and hoping they will hit the needed speed.

4) Less heat generation and lower noise and compact chassis

I will pay $2000 for it today if AMD wants a beta tester

My interpretation of Zen 6 leaks is that there may be a Medusa Halo(ish) with Zen 6 cores and Strix Halo (LPDDR5x) IOD, and a real Medusa Halo, with new IOD supporting LPDDR6.

Medusa may need fast PCIe connectivity, but not dGPU but for networking, to potentially put it into datacenter rack solutions for AI inference.

 

[gdansk]

I have a 128 thread Zen 2 Epyc so I guess I'm lunatic²?

It isn't a consumer CPU so the whole sentence does not apply.
But you might be.

 

[Joe NYC]

If they can get the same MT perf cheaper from 52C Nova Lake-S than from TR, why would they go with TR? If the performance is the same, they won’t get any increased productivity for going with TR. They will just waste money.

Sure, 64C TR will be faster, but 16/32C TR will likely not be a sensible option unless AMD drops the price significantly on those. But then the price jump for going to 64C will be insane if the price remains unchanged on only that variant.

Performance gains in client, for each additional thread beyond ~32 threads approaches zero.

What you are discussing here (and managed to suck a number of people into this discussion) is equivalent to the age old argument of how many angels can dance on the head of the pin.

 

[Joe NYC]

Just because a companies COST goes up does NOT mean that they can raise their price. Price is determined by the market demand and the supply.

Products, such as 9800x3d, 9950x3d already have astronomical margins (especially boxed products). $10-$20 cost difference is a margin of error.

Much bigger question is, between Intel and AMD, who will dominate the premium market.

Right now, AMD dominates DIY end of the spectrum, while Intel has a near monopoly on premium CPUs for corporate PCs.

For AMD, the goal is to move the needle on the premium corporate PCs. You don't get there by saving nickels and dimes on cost, but by demonstrating clear product superiority.

Right now, there is plenty of supply of CPU's. People are getting squeezed by inflation, and a "fast CPU" is definitely a luxury item. Luxury items are not purchased by the masses when times are tough.

I don't expect to see the price for the top end processors go up in the next generation.

Dinner for 4 at a decent Manhattan restaurant is more than 9800x3d, which one can enjoy for 5 years. Inflation is mostly elsewhere, not really affecting CPU prices nearly as much.

It's just a question of how much money people have left over after paying for inflated prices of life necessities...

 

[Thunder 57]

have you seen the cost of a 64 Zen 4 thread ripper it is $4400 even accounting for $1200 next gen motherboard and $1000 NVL-S CPU I would have $2200 to spare for the rest of the build.
7980WX Amazon

Here is the W880 MB

â Supermicro X14SAE W880 ATX Motherboard, LGA1851 Socket V1, DDR5-5600MHz | eBay

Cloud Project is located in the Silicon Valley of California, USA. 2 PCIe 5.0 x16 slots, 1 PCIe 4.0 x4 slot.

www.ebay.com

X14SAE | Motherboards | Products | Supermicro

I am a bit confused here 😅.

So you already know the price of a CPU that may or may not ever be released?

 

[Joe NYC]

And it'll be worse at everything a many-core CPU is supposed to be.
If Intel ships 16+32+4 NVL-S it'll be the most idiotic niche part unless it has phenomenal 1T performance and they make it the only good 1T part in their line up to make people pay for more than they want.

24 Zen 6 cores is enough to compete. It doesn't need to match it in CB nT, only beat it in 1T. Much like the 7700K didn't need to match the 1800X in nT for it to mop the floor with it in sales. Making a product for the one or two lunatics using a consumer CPU to transcode is an interesting business decision if real.

The "deflation" of the 52 core NVL will be instant, after single CCD Zen 6, with V-Cache, beats it by same margin as 9800x3d beats ARL in gaming.

I just hope the same "usual suspects" don't buy a NVL motherboard a month before launch of NVL (like they did with ARL).

 

[511]

The "deflation" of the 52 core NVL will be instant, after single CCD Zen 6, with V-Cache, beats it by same margin as 9800x3d beats ARL in gaming.

I just hope the same "usual suspects" don't buy a NVL motherboard a month before launch of NVL (like they did with ARL).

The 52 Part also has a big LLC Version and the big LLC is only on the 52 core part so suddenly it will become relevant.

 

[Thibsie]

Rather smells like 511 shades of wishful thinking.

The 52 Part also has a big LLC Version and the big LLC is only on the 52 core part so suddenly it will become relevant.

 

[Joe NYC]

The 52 Part also has a big LLC Version and the big LLC is only on the 52 core part so suddenly it will become relevant.

I wonder if 144 MB L3 on TSMC N2 passes the 50% Gross Margin test...

 

[LightningZ71]

So it'll depend on data set size and their L4 cache bandwidth.

 

[OneEng2]

For specific algos with dataset fully fitting in L1 cache I can get about 250-300 GB/s running on single core. When running from memory it drops to 30GB/s at best. So nope, it won't be well fed anytime soon, seeing single CCD has 8 cores and soon 12. But of course every improvement helps.

So the new definition of "well fed" is that main memory has L1 bandwidth?

Products, such as 9800x3d, 9950x3d already have astronomical margins (especially boxed products). $10-$20 cost difference is a margin of error.

Much bigger question is, between Intel and AMD, who will dominate the premium market.

Right now, AMD dominates DIY end of the spectrum, while Intel has a near monopoly on premium CPUs for corporate PCs.

For AMD, the goal is to move the needle on the premium corporate PCs. You don't get there by saving nickels and dimes on cost, but by demonstrating clear product superiority.

X3D was a great idea; however, lets not get to carried away about its contribution to profit margins. Certainly it gave AMD all-out bragging rights in gaming; however, the unit sales is trivial compared to other x86 client segments.

Dinner for 4 at a decent Manhattan restaurant is more than 9800x3d, which one can enjoy for 5 years. Inflation is mostly elsewhere, not really affecting CPU prices nearly as much.

Inflation effects CPU's... and certainly tariffs do.

The "deflation" of the 52 core NVL will be instant, after single CCD Zen 6, with V-Cache, beats it by same margin as 9800x3d beats ARL in gaming.

Hardly. I suspect that there will be plenty of consumers that buy into the "more cores" marketing that Intel is sure to implement. Very very few people care about gaming on a PC.

 

[Joe NYC]

X3D was a great idea; however, lets not get to carried away about its contribution to profit margins. Certainly it gave AMD all-out bragging rights in gaming; however, the unit sales is trivial compared to other x86 client segments.

You should refer to latest (Q1) AMD earnings conference call. Higher ASPs especially in desktop were referred to more than once as a contributor to the strong Q1 results.

Inflation effects CPU's... and certainly tariffs do.

Inflation in CPU prices is way, way below inflation in general.

As far as I know, CPU and semiconductors in general have not been affected by tariffs (yet?)

Hardly. I suspect that there will be plenty of consumers that buy into the "more cores" marketing that Intel is sure to implement.

The "moar cores" marketing hit a bit of a snag with Intel's Atom cores.

Very very few people care about gaming on a PC.

It just happens that the people who care about performance of PCs coincide with people who care about gaming on PC.