Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)

[DisEnchantment]

Speculate at will

 

[JustViewing]

Were you this pushy in a entire decade of Intel giving us 4 core after 4 core, with 3-5% improvement? Did you say the same in the latest 14xxx thread when Intel didn't bring any more cores to (most of) their line up? I bet you didn't, and I bet we know why...

I am sure you know why, for one I was not member of this forum at that time, secondly why should I care about Intel? Out of 25 years of my PC use I only bought Intel CPU (e7200) once and that was nearly 15 years ago.

If you want more cores buy the 14900K, It won't perform better than the 16c 7950X in productivity but it does have the cores you crave.

Why should I? I think you guys are missing the point.

 

[yottabit]

I really think the “more coars” on Desktop people are underestimating the memory bandwidth angle. The list of tasks that are very parrellelizable are small enough, then the ones that are completely independent of mem bandwidth are basically Cinebench or other CPU based offline rendering.

Even for large project code compilation, I’m pretty sure mem bandwidth will be an issue. Personally I upgraded from a 5900x to a 5950x when it was cheap to for fun, and saw zero performance improvement in any of the apps I care about, many of which are known for very good multicore scaling. In probably half the cases it’s because I was memory bandwidth bound and the other half the scaling had diminishing returns. I’ve considered picking up a second hand Epyc setup just for the memory bandwidth.

Also, ST performance gain is MT performance gain! (At least when it’s coming from IPC improvements and power limits are removed) I’d much rather have a 30-40% increase in ST than a 50% increase in cores. Honestly I think the more cores people are essentially clueless (maybe that’s a step too far to claim). I’d be really surprises if your workload scaled as well as you think it will. If you strictly just need threads for virtualization then I guess (maybe?) that makes sense. I might step down to a 8 or 12 core for Zen5 but then again looking at all those cores in Task Manager is fun

 

[Timmah!]

If perf/core was stagnating as well like it was during the Sandy -> Skylake ++++ era then I could understand the issue but we are getting significant perf/core increases gen on gen so it is not like AMD are sitting still.

There were rumors, that Zen5 perf jump will be just around 15 percent over Zen4. As clocks are not going higher, process is the same and core count is not getting increased either (to bump at least MT).
I for one dont find 15 percent perf increase worth of 700 EUROs AMD is gonna ask, if i want to replace my 7950x. Its anything but "pretty significant".
They are not sitting still, but they are giving absolute minimum to extract as much money as possible, they think they can get away with. Obviously, other companies do it as well, and i will happily criticize them for it.

 

[coercitiv]

I can't believe people here hellbent on justifying AMD's decision to limit desktop core count to 16. I can understand AMD is a business and want to maximize profit and also I understand there could be technical issues as well. But why on earth, we as end users need to defend that?

I'm not defending AMD product decisions, for all I care there's always room for bigger and/or cheaper. Let's go, let's have 32+ cores in the mainstream platform. I'm just telling you most consumers will still be buying the fastest cores, not the more cores.

Duh, what is the definition of the client PC space? For example, is a i7-based Lenovo ThinkStation a client PC?

A Thinkstation can be anything from a 1 liter SFF PC to a Threadripper/Xeon Workstation. Everything until the TRs and the Xeons is derived from the mainstream platform and the subject of this discussion.

 

[yuri69]

A Thinkstation can be anything from a 1 liter SFF PC to a Threadripper/Xeon Workstation. Everything until the TRs and the Xeons is derived from the mainstream platform and the subject of this discussion.

I've mentioned a "i7-based ThinkStation" specifically which seems to qualify as the subject of this "client PC" discussion.

So this proves my point at #5,937 . Client PCs are not DIY exclusive.

OEMs supply them in a large volume to the non-gamer users. Workloads ran at those OEM-supplied PCs has had the tendency to utilize multiple cores. Thus OEM-supplied client PCs would benefit from the increased core count available.

 

[StefanR5R]

Our concern should be value for money

Analyze the software that you are running and then figure out what kind of hardware offers good value to run this software.

Client PCs are not DIY exclusive.

OEMs supply them in a large volume to the non-gamer users. Workloads ran at those OEM-supplied PCs has had the tendency to utilize multiple cores.

Ditto.

Edit:
Here is my personal experience and perspective:
— I work in engineering. Whenever I have time-consuming computations to be done, they are time-consuming because huge amounts of simulation data are to be processed. The solvers which are involved are reasonably well parallelized, but as I said, they access large datasets. (Vice versa, if I've got simulations with smaller datasets, then these simulations are quickly solved on a small CPU. That is, there are no simulations or any other kinds of engineering software with small data but high CPU loads/ high CPU parallelism.) Postprocessing, which is CPU intensive and time-consuming too, is basically single-threaded! That's across a certain variety of simulation types and implementations.
— At home, I run lots and lots of scientific work as a volunteer. Most of these applications are single-threaded, but you are running several instances of such applications at once. Some of these applications are multi-threaded, but scale only to a moderate thread count per application instance. Often, the point after which thread count scaling gets notably worse is at an order of magnitude of eight threads, but it really differs between applications, and sometimes between different input data for a given application.

 

[soresu]

I'm not defending AMD product decisions, for all I care there's always room for bigger and/or cheaper. Let's go, let's have 32+ cores in the mainstream platform. I'm just telling you most consumers will still be buying the fastest cores, not the more cores.

Especially if more cores is leaning out significantly over £1,000.

I'd maybe splurge £1k for 24C+, but anything beyond £1k and I'm out.

 

[carrotmania]

There were rumors, that Zen5 perf jump will be just around 15 percent over Zen4.

The rumours are AMD going from 4 wide to 6 wide decode.

At 4 wide decode AMD gave us ~+30% perf every zen generation.

Intel gave us +5% every gen, except the move from tiger to alder, where they went from 4 to 6 wide decode, and we got a ~30% perf jump. After that rocket and raptor have been ~+5% again.

So let's recap: without going wider AMD has given the same perf jump as intel when they went 50% wider. But for Zen5, AMD will only give half the usual bump, while "improving" the core, going down a node and going as wide as intel ?

AMDead... as they say...

 

[Timorous]

There were rumors, that Zen5 perf jump will be just around 15 percent over Zen4. As clocks are not going higher, process is the same and core count is not getting increased either (to bump at least MT).
I for one dont find 15 percent perf increase worth of 700 EUROs AMD is gonna ask, if i want to replace my 7950x. Its anything but "pretty significant".
They are not sitting still, but they are giving absolute minimum to extract as much money as possible, they think they can get away with. Obviously, other companies do it as well, and i will happily criticize them for it.

I remember this.

I also remember the proclamations that it would be slower than the 5800X3D in gaming. As it turns out it was typically faster than the 5800X3D outside of the instances where cache really is king.

Any IPC uplift rumour or even announcement needs to be couched in the fact that AMD with their zen products are consistently under promising. I mean to a stupid degree. So even if 15% is a real number from a real internal document we have no idea on the context of how that number was generated so we have zero idea if that number is applicable in the general case.

The only thing leaks are good for is playing with numbers to see what something might look like and to give an idea of if there is going to be competition or not. It is basically just theorycrafting which can be fun but there is zero reason to use it to base any kind of buying decision off of.

 

[coercitiv]

Thus OEM-supplied client PCs would benefit from the increased core count available.

OEM-supplied PCs already have higher core count options, you already know this. Let me ask you this in return: instead of advocating for 32c+ in the mainstream platform, wouldn't you really prefer considerably cheaper HEDT CPUs? Let's say the 7970X were 40% cheaper, would that change your mainstream platform needs?

Like I said in a previous post, I'm all for cheaper CPUs, but the second AMD or Intel start selling modern 8P cores with R5 / i5 stickers... I'm not buying R7 / i7 anymore. (Intel is already in this situation btw, 6P+8E is plenty for me and Average Joe)

 

[yuri69]

OEM-supplied PCs already have higher core count options, you already know this. Let me ask you this in return: instead of advocating for 32c+ in the mainstream platform, wouldn't you really prefer considerably cheaper HEDT CPUs? Let's say the 7970X were 40% cheaper, would that change your mainstream platform needs?

I agree that the HEDT platform prices are quite ridiculous. It seems this holds true even from PoV of the average software house procurement.

Living in the real world a few decades has thought me that selling stuff with high margin is the thing you aim for. Therefore the pricey HEDT makes perfect sense.

Anyways, HEDT platforms are kinda overblown for an average developer. ~100 PCIe lines? A fast M.2 slot or two is usually all the average dev needs. Thermal/power design accomodating ~400W TDP is rather wasteful. Having to deal with 4-8ch of memory also brings questions.

It still feels like a naturally progressing mainstream PC platform could cover the future needs of the average developer better than a HEDT. It's just me.

 

[StefanR5R]

HEDT platforms are kinda overblown for an average developer. ~100 PCIe lines? A fast M.2 slot or two is usually all the average dev needs. Thermal/power design accomodating ~400W TDP is rather wasteful. Having to deal with 4-8ch of memory also brings questions.

It still feels like a naturally progressing mainstream PC platform could cover the future needs of the average developer better than a HEDT. It's just me.

Unfortunately for prospective HEDT or WS users, most likely there will never be cost-cut specialized HEDT or WS platforms which are segregated for the high-CPU or high-memory or high-I/O markets (and further special markets which want combinations of two out of these three). So indeed, if cost-cutting is a user's topmost requirement, he might have to make do with the upper end of the offerings on the low[ish]-CPU + low-memory + low-I/O PC market (which, at least, is much larger and partly more cost-driven).

Personally, I have been visiting the server market for my own special high-CPU/medium-memory/low-I/O computing desires, and the experience so far hasn't been all bad, despite attached compromises.

 

[Joe NYC]

There were rumors, that Zen5 perf jump will be just around 15 percent over Zen4. As clocks are not going higher, process is the same and core count is not getting increased either (to bump at least MT).
I for one dont find 15 percent perf increase worth of 700 EUROs AMD is gonna ask, if i want to replace my 7950x. Its anything but "pretty significant".
They are not sitting still, but they are giving absolute minimum to extract as much money as possible, they think they can get away with. Obviously, other companies do it as well, and i will happily criticize them for it.

I used to upgrade for double the performance. It just happened that it took ~13 years to get ~doubling of CPU performance.

Office PC i7-2400K -> 5800x
Home PC: I7-5930K -> 7800x3d

There wasn't really any reasonable upgrade, just swapping the CPU, because Intel went through like 6 sockets without delivering any material performance improvement. Well short of 2x

I might be persuaded to get Zen 5 V-Cache for the existing mobo, if it is just a CPU swap. If the new Zen 5 V-Cache is ~ $400. I may have a use for the 7800x3d too.

 

[Ajay]

Especially if more cores is leaning out significantly over £1,000.

I'd maybe splurge £1k for 24C+, but anything beyond £1k and I'm out.

Good point. Even if, beyond all reason, AMD decided to make a 32 core Zen client - it would cost more than double that of the 16 core (halo pricing, lower yields, etc.). So, IDK, so maybe $2K US. I think that would make some of those here think again. Ugh, this topic just needs to go away and I just added to it

 

[Tuna-Fish]

It's just DIMM sideways (for the most part).

That's really not a fair summary of CAMM. Especially for LPCAMM2, the signal path is much, much nicer than for any DIMM. It's not just that it's shorter, the lenght-matching is done better and the impedance of the connector is much lower.

 

[Abwx]

I used to upgrade for double the performance. It just happened that it took ~13 years to get ~doubling of CPU performance.

Office PC i7-2400K -> 5800x

5800X has 4.19x the perfs of a i7 2600K, the 1800X is at 2.55x.

AMD Ryzen 7000 im Test: So schnell sind 7950X und 7700X: Benchmarks in Apps und Games

Ryzen 9 7950X & Ryzen 7 7700X im Test: Benchmarks in Apps und Games / Leistung in Single-Core-Lasten (klassisch)

www.computerbase.de

 

[Fjodor2001]

Good point. Even if, beyond all reason, AMD decided to make a 32 core Zen client - it would cost more than double that of the 16 core (halo pricing, lower yields, etc.). So, IDK, so maybe $2K US. I think that would make some of those here think again. Ugh, this topic just needs to go away and I just added to it

Price per core is lower for 12/16C vs 8C on Zen4.

So following that trend, 24/32C Zen5 should not be that expensive.

 

[coercitiv]

So following that trend, 24/32C Zen5 should not be that expensive.

The 7950X launched at $699, roughly $44 per core. Assuming a successful arch, a bigger chiplet and/or factoring in some inflation, $50 per core could be a rough estimate for the next launch. (5950X was $50/core, 3950X was ~$47) That would put a 32 core SKU at ~$1440. for reference, the 24c 7960X is $1499.

Since the launch, the 7950X went on to become the cheapest SKU per core in the X lineup. One could argue this shows lower demand relative to the others. The most expensive per core is still the 8-core.

 

[adroc_thurston]

That's really not a fair summary of CAMM

Crude and simplified but it works (just like zen5 elmao).

Especially for LPCAMM2, the signal path is much, much nicer than for any DIMM. It's not just that it's shorter, the lenght-matching is done better and the impedance of the connector is much lower.

Yeah you don't have to tell me about DIMM insertion loss being complete dog.

 

[Joe NYC]

5800X has 4.19x the perfs of a i7 2600K, the 1800X is at 2.55x.

AMD Ryzen 7000 im Test: So schnell sind 7950X und 7700X: Benchmarks in Apps und Games

Ryzen 9 7950X & Ryzen 7 7700X im Test: Benchmarks in Apps und Games / Leistung in Single-Core-Lasten (klassisch)

www.computerbase.de

I am not using the computer for Cinebench. The extra cores from 4 to 8 are really diminishing returns on most client applications.

Single core performance seemed (subjectively) like 2x.

Also, using another eyeball method of clocks being up ~40% with IPC increases accounting for another 60% for 2x. There are instances where 5800x offers a bit of a bonus over 2x, namely in some games.

PS: Thanks for pointing out it was i7-2600K not 2400K. It's been out of sight for a couple of years.

 

[Fjodor2001]

The 7950X launched at $699, roughly $44 per core. Assuming a successful arch, a bigger chiplet and/or factoring in some inflation, $50 per core could be a rough estimate for the next launch. (5950X was $50/core, 3950X was ~$47) That would put a 32 core SKU at ~$1440.

Yes, correct. So $44 per core would mean 32C Zen5 at $1408, and 24C at $1056, at launch. If we toss in a mix of Zen5 + Zen5C cores (e.g. 8xZen5 + 16/24 Zen5C), it would be even lower. So much lower than $2K+ as previously suggested by some.

Also, consider that 7950X now is $550, so $34 per core. Means that after 1-1.5 years, 32C Zen5 would be at $1088, and 24C at $816. Nowhere near $2K+.

And all this is assuming that Moore's law is dead, and that we more or less have fixed cost per transistor count 4ever. No technological/price advancement in that aspect. Otherwise we'd see doubling of transistor count every 18 months at same price. So 16C CPU would cost same as 32C in 18 months, using same core.

 

[adroc_thurston]

and that we more or less have fixed cost per transistor count 4ever.

Oh no you see, cost per xtor is going up.

Yes, correct. So $44 per core would mean 32C Zen5 at $1408, and 24C at $1056, at launch. If we toss in a mix of Zen5 + Zen5C cores (e.g. 8xZen5 + 16/24 Zen5C), it would be even lower. So much lower than $2K+ as previously suggested by some.

That's not happening.
Forget about it.

 

[Glo.]

And all this is assuming that Moore's law is dead, and that we more or less have fixed cost per transistor count 4ever. No technological/price advancement in that aspect. Otherwise we'd see doubling of transistor count every 18 months at same price. So 16C CPU would cost same as 32C in 18 months, using same core.

It slowly enters era of exponential growth of cost/Xtor.

 

[adroc_thurston]

It slowly enters era of exponential growth of cost/Xtor.

It's not exponential but it's not nice.

 

[Fjodor2001]

Oh no you see, cost per xtor is going up.

Reverse Moore's law. No we're not seeing that yet. Stagnated Moore's law is bad enough.

That's not happening.
Forget about it.

It's already happened with previous Zen generations. Deal with it.