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!

 

[DisEnchantment]

That's the public excuse, much more likely is that 3D-cache uses HD cells and they don't really like going over 1.3V, 3Dcache and base die share the same voltage rail in Milan-X and 5800XD so the die itself isn't allowed to go above that limiting Fmax

It is probably simpler than that, too much cache increases latency which degrade performance outside of BW bound loads.
Also the V-Cache dies have separate DLDOs delivering power and can be turned off, but they have not found a way to make this software transparent at run time. Code is compiled statically after all, and most software assume CPUID Fn8000_0006 bits to be non dynamic.
L3 and the core are not in the same clock domain either. Other thing is thermal constraint.

 

[naad]

It is probably simpler than that, too much cache increases latency which degrade performance outside of BW bound loads.
Also the V-Cache dies have separate DLDOs delivering power and can be turned off, but they have not found a way to make this software transparent at run time. Code is compiled statically after all, and most software assume CPUID Fn8000_0006 bits to be non dynamic.
L3 and the core are not in the same clock domain either. Other thing is thermal constraint.

Cache latency is _primarily_ determined by wire distance, which V-cache doesn't have a problem with, 2D caches can only go wider, increasing distance from the core V-cache sits near the core, which is why V-cache is only 4 cycles higher than the regular Veremeer L3 for 3x the size.

If not the V-cache not liking voltage, maybe the TSV's themselves don't like too high voltage and 1.3V is the most they can take? These are physical design details I doubt AMD will ever come clean on and TSMC will improve each year, hopefully they talk more of this when Raphael-X comes out

 

[DisEnchantment]

Cache latency is _primarily_ determined by wire distance, which V-cache doesn't have a problem with, 2D caches can only go wider, increasing distance from the core V-cache sits near the core, which is why V-cache is only 4 cycles higher than the regular Veremeer L3 for 3x the size.

Cache latency is determined by associativity and tag/directory design. More tags added = more area, more power but faster access time. Wire delay is femtoseconds on a vcache to base die.
Same amount of tags but bigger size = higher latency.
Adding more tags means increasing the base L3, because that is where are the access and control logic exists. Either increase tags at the cost of area and power or take the latency hit. There is a tradeoff somewhere
If tag area and latency due to increased size are not in the picture everybody would just go big on L2 why bother with L3.
There is a reason Zen4 L2 has higher latency. They had to balance cache size, tag area, power and latency of L2.

 

[Kocicak]

... This is where HEDT first came out, and had defined walls.

HEDT and Normal segement never shared the same platform.
HEDT always had its own board and own pinset. ...

You must remember also, what did the most expensive CPUs for the consumer platform cost. That Intel BACK IN THE DAY had two different platforms means nothing for today.

I repeat again, if you look at the cost and performance of 12 and 16 core CPUs for AM5, this is HIGH END PC territory.

Even the 8 core is novadays pretty expensive, and the 3D variant will be even more. One could argue if even these do not fall into the HEDT category.

 

[MadRat]

With 3D printing, it should be trivial to design a microchannel cooling system using an air compressor. Pumping air under pressure and letting it expand directly over the processor would be a big boost over fan cooling. If designed correctly you could do both. Instead of running the compressor full-time, set it up to kick in at specific temperatures.

 

[eek2121]

Considering AMD seens even more hush-hush than usual in regards to Raphael-X and we're not that far off from launch and we don't know anywere as much potential hard data (such as SkyJuice's early 5,85GHz fmax leak) as we did about Raphael 5 months ago, there's certainly some legroom for AMD to surprise us, beyond just what 5800X3D was to the rest of the Zen3 lineup.

I wonder if they're confident enough to push a "Dragon Range-X" SKU out next year as well. It would certainly get some headlines/buzz and grab them some mindshare on laptop.

I suspect we won’t see any X3D parts until Q3 of next year. AMD doesn’t need them currently unless Intel surprises them.

I am not sure that a mobile variant would help much, but it would be neat to see.

Cache latency is _primarily_ determined by wire distance, which V-cache doesn't have a problem with, 2D caches can only go wider, increasing distance from the core V-cache sits near the core, which is why V-cache is only 4 cycles higher than the regular Veremeer L3 for 3x the size.

If not the V-cache not liking voltage, maybe the TSV's themselves don't like too high voltage and 1.3V is the most they can take? These are physical design details I doubt AMD will ever come clean on and TSMC will improve each year, hopefully they talk more of this when Raphael-X comes out

Yep, I believe the power argument over anything else.

 

[Hans Gruber]

I suspect we won’t see any X3D parts until Q3 of next year. AMD doesn’t need them currently unless Intel surprises them.

I am not sure that a mobile variant would help much, but it would be neat to see.


Yep, I believe the power argument over anything else.

Well said. AMD is going to need those 3D V-cache parts.

 

[igor_kavinski]

With 3D printing, it should be trivial to design a microchannel cooling system using an air compressor. Pumping air under pressure and letting it expand directly over the processor would be a big boost over fan cooling. If designed correctly you could do both. Instead of running the compressor full-time, set it up to kick in at specific temperatures.

Please do a Kickstarter campaign. I'll chip in with $5

 

[MadRat]

X3D is a niche, low volume product line.

It was never meant to be mainstream.

 

[MadRat]

Please do a Kickstarter campaign. I'll chip in with $5

Much easier for me to retrofit/fabricate something from old parts than to 3D print. I have no Zen4 to work with. But I'm all in if someone wants to donate me one.

 

[biostud]

X3D is a niche, low volume product line.

It was never meant to be mainstream.

If you have a product that sells like hotcakes and you make a profit from it, wouldn't you move it to the mainstream segment?

Dragon Range 7800X3D in gaming laptops, would be quite popular too I would think.

 

[Atari2600]

With 3D printing, it should be trivial to design a microchannel cooling system using an air compressor. Pumping air under pressure and letting it expand directly over the processor would be a big boost over fan cooling. If designed correctly you could do both. Instead of running the compressor full-time, set it up to kick in at specific temperatures.

You'd still need a precooler to cool the compressed air before letting it into the cpu chamber.

That precooler would need to dump heat somewhere.

It would have more cooling capacity than current approach of heat pipe to blown air. Not sure it'd be more energy efficient though. It's definitely much more expensive.

 

[Kocicak]

I have no idea why would anybody talk about EXPANDING air, you would need the chip to be 1000°C so that the air expands nicely. (Why would you need the air to expand, do you want to run some engine with it???)

Cooling with air works only with large surface area, because the air has very low density and heat capacity. You need large volume of it blowing over a large surface area to allow it to carry away enough heat.

The problem is not cooling, the problem is the source of heat. You do not crank up the volume and then complain about noise and search for ear muffs. You just turn the volume down a little bit.

 

[eek2121]

X3D is a niche, low volume product line.

It was never meant to be mainstream.

Not only does AMD disagree with you, but they have internally considered breaking out that particular business segment in earnings reports.

Note that X3D != gaming, INTEL uses AMD chips with 3D V-Cache chips in house for simulations (allegedly).

Also, if you want to attempt to narrow the focus to gaming, it is the top selling gaming chip, and the TAM for PC gamers is anything but small so…

 

[HurleyBird]

Note that X3D != gaming, INTEL uses AMD chips with 3D V-Cache chips in house for simulations (allegedly).

Yeah, the extra cache either does very little or it makes a massive difference. When you have a specific kind of problem that exploit a competitor's architecture, at a certain point it becomes pig-headed to not make use of your competitors products for it. Wouldn't surprise me in the slightest me if AMD had used Xeons for AVX-512 workloads prior to Zen 4.

 

[MadRat]

I have no idea why would anybody talk about EXPANDING air, you would need the chip to be 1000°C so that the air expands nicely. (Why would you need the air to expand, do you want to run some engine with it???)

Cooling with air works only with large surface area, because the air has very low density and heat capacity. You need large volume of it blowing over a large surface area to allow it to carry away enough heat.

The problem is not cooling, the problem is the source of heat. You do not crank up the volume and then complain about noise and search for ear muffs. You just turn the volume down a little bit.

You mistaken expansion of the compressed air for thermal expansion. You seem to underestimate cooling using compressed air. Pnuematics make excellent cooling devices for industrial purposes. They can be extremely loud if you compress air next to where you are cooling, and it is higher cost. But they invented these things called hoses and storage tanks that would help mitigate some of the sound generation. Some expansion devices are less loud than others, so that's not an issue by default. What it allows you to do is concentrate a lot of cooling capacity with pinpoint accuracy. It should quickly drop a hot heatsink down in temperature nearly instantly while the CPU is at full or near full load. As CPUs get more compact they concentrate heat generation into smaller, finer areas. At some point a heatsink with a fan will not keep up.

 

[Det0x]

You mistaken expansion of the compressed air for thermal expansion. You seem to underestimate cooling using compressed air. Pnuematics make excellent cooling devices for industrial purposes. They can be extremely loud if you compress air next to where you are cooling, and it is higher cost. But they invented these things called hoses and storage tanks that would help mitigate some of the sound generation. Some expansion devices are less loud than others, so that's not an issue by default. What it allows you to do is concentrate a lot of cooling capacity with pinpoint accuracy. It should quickly drop a hot heatsink down in temperature nearly instantly while the CPU is at full or near full load. As CPUs get more compact they concentrate heat generation into smaller, finer areas. At some point a heatsink with a fan will not keep up.

Linus Tech Tips of all people had a video about this a while back, ill have a look if i can find it

*edit*

 

[Kocicak]

Oh, cooling with expanding previously pressurised air - that could theoretically work, but seems extremely impractical. Compare its energy efficiency to a liquid cooling loop, which has the liquid cooled by a heat pump. It consumes MULTIPLE times more energy.

 

[DrMrLordX]

I've heard high-pressure pneumatic hoses leak when seals or valves break on the line. The insane SSSSSSSSSSSSS noise is profoundly unpleasant. I can't see how rapid expansion of stored, pressurized air would be much better.

 

[igor_kavinski]

Imagine a pump operated by gravity. At the start, the piston is at the top and it is relatively light. Fill it with water and it starts pushing down on the gas beneath it. When it goes down as much as it can, multiple valves in the piston open at the same time to let the water out instantly, releasing pressure on the gas which will then boost the piston back up and cause a cooling effect. I know it will be large and kinda unwieldy but if the cooling effect is good, maybe it can work? If anyone manages to make this work, do let me in on a few cents of the royalties

 

[repoman0]

Why waste the time, effort and money to try to force air to behave more like a liquid when we have the technology to cool things with an actual liquid?

... not that Zen 4 seriously needs a ton of cooling capacity to perform near its peak, especially with some minor tweaking.

 

[nicalandia]

Why waste the time, effort and money to try to force air to behave more like a liquid when we have the technology to cool things with an actual liquid?

I have mentioned before. Instead of Making Larger and Larger Water Cooling contraptions(Like Car Radiator Fan Set Up) lets use Chemistry And Science to increase the Thermal Conductivity of Water. at a volume fraction of 0.1 vol % Copper Nanoparticles on Water increases the thermal conductivity of water by 24% and 90% by 1%. Volume

You could do this on your Kitchen, there is a video on Youtube on how to do it on a household blender


Also at 5.5 Ghz

 

[igor_kavinski]

You could do this on your Kitchen, there is a video on Youtube on how to do it on a household blender

But has anyone tried to use those copper nanoparticles in a cooling contraption?

 

[nicalandia]

But has anyone tried to use those copper nanoparticles in a cooling contraption?

For CPU Cooling? No that I know of. But it's used on other water cooling systems and CPU cooling is not that complex.

Who here is brave enough to try it?

Perhaps on old AMD Bulldozer FX processors

 

[Det0x]

I have mentioned before. Instead of Making Larger and Larger Water Cooling contraptions(Like Car Radiator Fan Set Up) lets use Chemistry And Science to increase the Thermal Conductivity of Water. at a volume fraction of 0.1 vol % Copper Nanoparticles on Water increases the thermal conductivity of water by 24% and 90% by 1%. Volume

You could do this on your Kitchen, there is a video on Youtube on how to do it on a household blender


Also at 5.5 Ghz

View attachment 68759

Here you go @nicalandia, i know you like GB5 runs

7950x PBO CO tuned

ST= 2407
MT = 26241


I also have higher MT scores, but lower ST then..