Discussion Speculation: Zen 4 (EPYC 4 "Genoa", Ryzen 7000, etc.)

[Vattila]

Except for the details about the improvements in the microarchitecture, we now know pretty well what to expect with Zen 3.

The leaked presentation by AMD Senior Manager Martin Hilgeman shows that EPYC 3 "Milan" will, as promised and expected, reuse the current platform (SP3), and the system architecture and packaging looks to be the same, with the same 9-die chiplet design and the same maximum core and thread-count (no SMT-4, contrary to rumour). The biggest change revealed so far is the enlargement of the compute complex from 4 cores to 8 cores, all sharing a larger L3 cache ("32+ MB", likely to double to 64 MB, I think).

Hilgeman's slides did also show that EPYC 4 "Genoa" is in the definition phase (or was at the time of the presentation in September, at least), and will come with a new platform (SP5), with new memory support (likely DDR5).

What else do you think we will see with Zen 4? PCI-Express 5 support? Increased core-count? 4-way SMT? New packaging (interposer, 2.5D, 3D)? Integrated memory on package (HBM)?

Vote in the poll and share your thoughts!

 

[CouncilorIrissa]

I am aware of that, and i agree, but i have seen people waxing lyrically about Zen5, maybe even before Zen4 was released, or shortly after, how Zen5 will be that "supercycle", thats gonna improve things drastically, after all the chief architect was excited about it...
and now it seems to turn out that it might bring less performance uplift that Zen4 brought, so maybe not so special after all? Or is it strictly because people are specifically enthusiastic about possible ST performance gains? Which, ultimately, may not be as big as they initially hyped to be.

Extracting performance from IPC improvements alone is much more impressive than from raised power limits and a new node.

 

[Timmah!]

Extracting performance from IPC improvements alone is much more impressive than from raised power limits and a new node.

So you care more for how the performance is gained, than its amount? If one CPU runs something 30 percent faster as result of combined 10 percent IPC and 20 percent clockspeed increase, and another one 15 percent faster as result of just IPC increase, are you more excited for the latter?

 

[Bigos]

This does not look like the Zen 5 thread.

Speaking of the new Threadrippers, I was thinking about them but predictably the cost is through the roof. 7970X will have 2x as many cores and 2x as many memory channels (and 2x? maybe more PCIe lines) as 7950x but it will cost 4-4.5x as much. I could rationalize the purchase if it was 3x, but 4x+ is insane. I guess this is what lack of competition does.

I could just buy two 7950x PCs and distribute workloads on them and it would be cheaper. AMD HEDT might as well not exist to me.

 

[StefanR5R]

Disabling SMT on Cineben R23 on these beasts(that put more than 256 Threads) is the only way to increase the score.

If neither the OS nor the application itself knows how to schedule the application threads in the optimum way, you still don't need to outright disable SMT. You simply can correct the scheduling of the application by assigning processor affinities to its threads.

 

[Ajay]

This does not look like the Zen 5 thread.

Speaking of the new Threadrippers, I was thinking about them but predictably the cost is through the roof. 7970X will have 2x as many cores and 2x as many memory channels (and 2x? maybe more PCIe lines) as 7950x but it will cost 4-4.5x as much. I could rationalize the purchase if it was 3x, but 4x+ is insane. I guess this is what lack of competition does.

I could just buy two 7950x PCs and distribute workloads on them and it would be cheaper. AMD HEDT might as well not exist to me.

Then you don't really need Threadripper. If you did, you'd pay up. The prices are not unusual for HEDT, if you correct for inflation.

 

[StefanR5R]

7970X will have 2x as many cores and 2x as many memory channels (and 2x? maybe more PCIe lines) as 7950x but it will cost 4-4.5x as much. I could rationalize the purchase if it was 3x, but 4x+ is insane. I guess this is what lack of competition does.

I could just buy two 7950x PCs and distribute workloads on them and it would be cheaper. AMD HEDT might as well not exist to me.

Mainboard cost will possibly be >2x too. On the other hand, you do get the benefit of consolidated local mass storage (edit: also, depending on the application, perhaps consolidated RAM) and form factor, and of a single operating system instance. The latter is priceless for some usages (though apparently not yours), which is one reason for the markup.

Another downside of the consolidated system: One 7970X compatible air cooler (out of the existing choices) will not be twice as performant as an 7950X compatible air cooler from the upper tier. This can only be compensated by high air speed or by resorting to watercooling.

 

[nicalandia]

Threadripper and Bergamo AIDA64

ThreadRipper 7995WX



Bergamo 9754

Also this.

AMD Ryzen Threadripper PRO 7995WX scores 100K Cinebench R23 points at stock, claims 148K record with OC - VideoCardz.com

AMD Threadripper breaks 148K Cinebench R23 score, a new world record The company invited members of the media to its headquarters for the first public demo of the new Threadripper platform. One of the sessions included an extreme overclocking attempt, a successful one, one may add. The review...

videocardz.com

 

[Timmah!]

This does not look like the Zen 5 thread.

Speaking of the new Threadrippers, I was thinking about them but predictably the cost is through the roof. 7970X will have 2x as many cores and 2x as many memory channels (and 2x? maybe more PCIe lines) as 7950x but it will cost 4-4.5x as much. I could rationalize the purchase if it was 3x, but 4x+ is insane. I guess this is what lack of competition does.

I could just buy two 7950x PCs and distribute workloads on them and it would be cheaper. AMD HEDT might as well not exist to me.

Look at it from the bright side. AMD being greedy saves you money :-D If it was god forbid reasonably priced to be actual choice, you would be already threading plans to get one, hehe.

 

[Abwx]

7970X will have 2x as many cores and 2x as many memory channels (and 2x? maybe more PCIe lines) as 7950x

8 RAM channels is 4x the 7950X dual channel.

 

[Joe NYC]

So you care more for how the performance is gained, than its amount? If one CPU runs something 30 percent faster as result of combined 10 percent IPC and 20 percent clockspeed increase, and another one 15 percent faster as result of just IPC increase, are you more excited for the latter?

I think @CouncilorIrissa is talking about skills required. It takes higher level of skills to improve IPC than to increase power limit or to move to new process node.

​

 

[Bigos]

8 RAM channels is 4x the 7950X dual channel.

Non-pro Threadripper only has 4 channels. It's the Threadripper Pro that has 8. And it will be even more expensive.

 

[Abwx]

Non-pro Threadripper only has 4 channels. It's the Threadripper Pro that has 8. And it will be even more expensive.

Right but 32C is 2500$, two 7950X PCs would cost this price, so it would be more economical but not that much.

Since it s linearly priced the 64C is 5000$, that s about 4 7950X set ups, so that s not much more expensive even if we include the rest of the components, not relevant for the amateurish user but quite competitive for pro usages.

 

[Mopetar]

So you care more for how the performance is gained, than its amount? If one CPU runs something 30 percent faster as result of combined 10 percent IPC and 20 percent clockspeed increase, and another one 15 percent faster as result of just IPC increase, are you more excited for the latter?

In general the better IPC means that moving to a new node or increasing the clock speed means even more room for performance.

Obviously not every node brings massive upgrades, and designs which enable higher IPC may make decisions that limit clock speeds, but typically high clock speeds also mean high power consumption so getting the performance through IPC can also mean a more energy efficient or easier to cool chip even if the performance numbers are essentially the same.

 

[Timmah!]

I think @CouncilorIrissa is talking about skills required. It takes higher level of skills to improve IPC than to increase power limit or to move to new process node.

​

I do understand that, though not sure i quite agree. I mean, seeing Intels troubles with their manufacturing, new and smaller process node seems to be as much feat as new chip design with higher IPC. Though, granted, its more of a TSMCs feat than AMDs.

 

[Joe NYC]

I do understand that, though not sure i quite agree. I mean, seeing Intels troubles with their manufacturing, new and smaller process node seems to be as much feat as new chip design with higher IPC. Though, granted, its more of a TSMCs feat than AMDs.

Going from Zen 3 to Zen 4, AMD was able to get a little bit from all 3 buckets:
- IPC increase
- increased max power
- new node
Sum of all 3 added up to a very nice improvement gen on gen.

Going from Zen 4 to Zen 5, there will be less from the new node (N5P to N4P), and probably nothing from increased power budget. So majority of the gains will have to come from IPC gains.

If AMD delivers on IPC, that's something to be excited about.

 

[coercitiv]

Look at it from the bright side. AMD being greedy saves you money :-D If it was god forbid reasonably priced to be actual choice, you would be already threading plans to get one, hehe.

Remember what I told you a while ago?

Like I said before, we would not be having this conversation if HEDT lineups from both companies were alive and reasonably priced. The reality is you have a specific need, and these two companies want you to show them your wallet. More cores on the mainstream consumer platform is just a band-aid, in the end either you or AMD/Intel will blink first and you will get your true workstation. We'll just have to see at what price.

From a declarative standpoint it looks like you (as in the prosumers in the forum, not you specifically) are not going to blink, not yet anyway.

 

[igor_kavinski]

What's with there being no V-cache TR part?

It can't be that there is no demand for it. What is AMD's excuse? What do they have against making a little more money over the already exorbitant TR pricing?

 

[biostud]

What's with there being no V-cache TR part?

It can't be that there is no demand for it. What is AMD's excuse? What do they have against making a little more money over the already exorbitant TR pricing?

EPYC?

Which HEDT software would benefit from large cache?

 

[igor_kavinski]

Which HEDT software would benefit from large cache?

The main issue with 7950X3D is the dissimilar nature of the CCDs so it gets dragged down by the non-V-cache CCD in the cache sensitive applications. A TR V-cache part with similar V-cache CCDs would be formidable and downright unbeatable in multitasking.

 

[biostud]

View attachment 87577
View attachment 87578
View attachment 87579
The main issue with 7950X3D is the dissimilar nature of the CCDs so it gets dragged down by the non-V-cache CCD in the cache sensitive applications. A TR V-cache part with similar V-cache CCDs would be formidable and downright unbeatable in multitasking.

Then you just have to pay for EPYC

 

[igor_kavinski]

Then you just have to pay for EPYC

Pay more and get lower frequency CCDs? No thanks

 

[nicalandia]

What's with there being no V-cache TR part?

It can't be that there is no demand for it. What is AMD's excuse? What do they have against making a little more money over the already exorbitant TR pricing?

You can build a Genoa-X Workstation

Ultimate AMD Genoa-X Workstation Review

Our ultimate AMD Genoa-X workstation might actually be a server in workstation clothing but it's absolutely amazing.

www.storagereview.com

 

[gdansk]

AMD does not desire to compete with Epyc. 9224 is about $1800 so that set the price floor.

For my purposes Epyc 8124 would work, I mainly need *some* PCIE lanes but not infinite. But my workload is less workstation and more server. With a plan to add a lot of NVMe storage.

But I think I'll repurpose my 7950X when it is replaced instead. Especially if I can find an ECC UDIMM compatible AM5 motherboard for less than a TRX50 motherboard. It can support up to 6 NVME drives which is probably enough.

 

[Ajay]

Right but 32C is 2500$, two 7950X PCs would cost this price, so it would be more economical but not that much.

Since it s linearly priced the 64C is 5000$, that s about 4 7950X set ups, so that s not much more expensive even if we include the rest of the components, not relevant for the amateurish user but quite competitive for pro usages.

Presuming you have software workload which is distributed across PC. Add the cost of at least 10Gb NICs and a 10gb switch for half way decent bandwidth between them. So, not that many workstations/HP PC applications.

 

[Ajay]

What's with there being no V-cache TR part?

It can't be that there is no demand for it. What is AMD's excuse? What do they have against making a little more money over the already exorbitant TR pricing?

Because for the target market the demand isn't high enough to offset the cost/performance benefit. Duh.