Question AMD Phoenix/Zen 4 APU Speculation and Discussion

[uzzi38]

I can finally make this thread.

https://twitter.com/x/status/1519669375283957760

Phoenix is indeed RDNA3. My advice to everyone: treat the old APU rumours as being out of date.

 

[DrMrLordX]

Maybe not the original owner but when it gets sold to someone else for a fraction of its original price, the new owner may want to extend its life a bit more with a RAM upgrade.

OEMs are not concerned with that.

Soldering means more complexity and more work.

Wait, what? How?

 

[moinmoin]

Wait, what? How?

For a let's say MB manufacturer it's cheaper, less complex and less work to just solder some sockets and tell the customer to buy the chips and shove them in than buy the chips itself and solder them directly. 🤡

Actually I'm not even sure this claim isn't hitting a tipping point in the coming gens with how PCIe 5 and DDR5 are raising the complexity necessary to ensure signal integrity.

 

[Heartbreaker]

For a let's say MB manufacturer it's cheaper, less complex and less work to just solder some sockets and tell the customer to buy the chips and shove them in than buy the chips itself and solder them directly. 🤡

Maybe, If you are only selling MBs.

But if you are selling the whole product like an SFF computer, you will get a lot less hassle with customers screwing things up, if you just solder as much in as you can. That's just less for monkeys to mess up.

IMO there is 99.9%+ chance that any kind of big APU with more than 2 memory channels, will be solder in only.

This would be like Intel Kaby-G. You could only get it it soldered in. There was no socket option. Remember what a "big hit" that was.

 

[moinmoin]

IMO there is 99.9%+ chance that any kind of big APU with more than 2 memory channels, will be solder in only.

Indeed, that's what I'd expect as well. In mobile space powerful APUs can demand a premium (or can they? Strix Halo will show either way). In desktop space APUs will always play the second fiddle to more dedicated CPUs + dGPUs, so there is no chance a separate more expensive socketed platform will be created just to offer such APUs to the DIY market.

 

[FlameTail]

What are you calling a "memory bank"? I have never heard anyone refer to two modules as "two memory banks" before.

Yes I want to know too.

 

[DrMrLordX]

But if you are selling the whole product like an SFF computer, you will get a lot less hassle with customers screwing things up, if you just solder as much in as you can. That's just less for monkeys to mess up.

It's not just that, but if you are an OEM providing the entire package (whether it's a laptop or an SFF), in no way; shape; or form is it technically easier and cheaper to solder on sockets and then install SIMMs or DIMMs when you can just solder the RAM straight onto the board.

 

[Shivansps]

It's not just that, but if you are an OEM providing the entire package (whether it's a laptop or an SFF), in no way; shape; or form is it technically easier and cheaper to solder on sockets and then install SIMMs or DIMMs when you can just solder the RAM straight onto the board.

That is true for 4 and 8GB products low end, when we start talking of 16, 32 and 64GB products costs go up fast.
Specially when you need to build 3 diferent boards for the task, and even more if you also soldering an emmc.

 

[Thunder 57]

It's not just that, but if you are an OEM providing the entire package (whether it's a laptop or an SFF), in no way; shape; or form is it technically easier and cheaper to solder on sockets and then install SIMMs or DIMMs when you can just solder the RAM straight onto the board.

That's twice you've mentioned SIMM's in the past few hours. I feel like I've been yanked back into the 90's .

 

[FlameTail]

It's not just that, but if you are an OEM providing the entire package (whether it's a laptop or an SFF), in no way; shape; or form is it technically easier and cheaper to solder on sockets and then install SIMMs or DIMMs when you can just solder the RAM straight onto the board.

If Intel/AMD/Qualcomm is the one providing the on-package memory, this is not a problem.

 

[MadRat]

Too bad they cannot do opposing sockets or stack the CPU on a dedicated RAM-module board. The former could be on the opposite side of the maiboard as your CPU. The latter probably is more of a notecard shape than a square. Both would cut your trace lengths down considerably, while still supporting traditional RAM sockets for a buffer between the RAM-module and storage, where you could use smart cache technology to preload/mirror your high usage applications. Because your RAM-module would incorporate your CPU, no need for socket. Nothing would stop use of a socket, but its a way to eliminate it for OEM makers. The interface for the RAM-module could be PCIe-based, only laying flat to allow space for cooling.

Mount your fan before installing. Slide your RAM-board into your dual-PCIe x16 interface. Screw down the RAM board to the riser nuts (that also operate as grounding) coming up from the mainboard. Voila! No more mental and physical gymnastics trying to mount the CPU and fan in the confined spaces.

 

[Shivansps]

LPCAMM2 allows you to have quad channel memory chips with two memory banks.

I believe Dell designed this this way, for very good reason of future products, that involve brand that has letters "A", "M", "D" in their name.

I mean, why else AMD would use a 256bits LPDDR5X for Strix Halo when you could archive the same with 128bits GDDR6 or LPDDR6. In fact if Strix Halo really is 2025 and with LPDDR5X that would mean it is arriving a bit too late. There has to be another reason behing this decision and i think it is because they want to align all Strix behind LPCAMM2, but if it is only AMD i dont know how successfull this will be, they need Intel too.

 

[Shivansps]

Maybe, If you are only selling MBs.

But if you are selling the whole product like an SFF computer, you will get a lot less hassle with customers screwing things up, if you just solder as much in as you can. That's just less for monkeys to mess up.

IMO there is 99.9%+ chance that any kind of big APU with more than 2 memory channels, will be solder in only.

This would be like Intel Kaby-G. You could only get it it soldered in. There was no socket option. Remember what a "big hit" that was.

It is not so easy to just solder things as people may expect, YES, it is cheaper and simplier than putting a slot and then adding a memory to it, in fact this is a lot cheaper and simplier.

But there are hidding costs. First off you need to decide how many boards to produce of each configuration AND DO NOT GET IT WRONG, if you get this wrong you lose a lot of money. And when you start adding variations the posible combination really start to get out of hand. Lets say, you have 3 possible socs, then you have 8, 16 and 32 gb of ram variations of each one, and finally storage, and you can see why storage is the last thing it gets soldered on, otherwise you would se EMMC used a lot more.

It was decided long ago that soldering the cpu to the boards was acceptable as long as memory and storage were modular, because guessing what class of cpu would sell more was easy.

So you can see why this is generally used for low end products, so they can produce one or two boards and thats it. A tactic used in laptops lately was to solder 4GB-8GB DDR4 and offer one extra slot. Not sure if the same can be done with LPCAMM2.

Also for things like Strix Halo we are talking about 32 and 64GB of LPDDR5X, who is going to solder all of that? it is far too risky.

 

[Heartbreaker]

It is not so easy to just solder things as people may expect, YES, it is cheaper and simplier than putting a slot and then adding a memory to it, in fact this is a lot cheaper and simplier.

So you agree it's a lot cheaper and simpler to just solder it in.

You are off to a good start. It's also lower customer service load, because it's one less thing for monkeys to mess up and engage customer service.

But there are hidding costs. First off you need to decide how many boards to produce of each configuration AND DO NOT GET IT WRONG, if you get this wrong you lose a lot of money. And when you start adding variations the posible combination really start to get out of hand. Lets say, you have 3 possible socs, then you have 8, 16 and 32 gb of ram variations of each one, and finally storage, and you can see why storage is the last thing it gets soldered on, otherwise you would se EMMC used a lot more.

It was decided long ago that soldering the cpu to the boards was acceptable as long as memory and storage were modular, because guessing what class of cpu would sell more was easy.

So you can see why this is generally used for low end products, so they can produce one or two boards and thats it. A tactic used in laptops lately was to solder 4GB-8GB DDR4 and offer one extra slot. Not sure if the same can be done with LPCAMM2.

Also for things like Strix Halo we are talking about 32 and 64GB of LPDDR5X, who is going to solder all of that? it is far too risky.

You have a wall a text here with only one argument. That it's some kind of major problem predicting how many machines with each config.

That's absurd.

Companies have been making machines for years, they have VERY good idea on the percentage of takers for each RAM config.

Even easier if it's for a Halo part with a 256 bit bus. You could just make them all 32GB if multiple configs were a serious concern. Or you predict and have a sale on the machines that are slow moving. It's a non issue.

It's also completely irrelevant to the APU discussion. It doesn't matter if the memory is soldered or on LPCAMM. It's still LPDDR either way, and performs the same either way (that is kind of the point).

It really seems that if people want to have a discussion about the "wonders" of LPCAMM, that should be it's own separate thing, as it really has no impact on the success or failure of the APU. There will probably be designs that are soldered and designs that use LPCAMM.

 

[FlameTail]

I mean, why else AMD would use a 256bits LPDDR5X for Strix Halo when you could archive the same with 128bits GDDR6 or LPDDR6

LPDDR5X = 8533
LPDDR6 = 12 800

So 128 bit LPDDR6 /= 256 bit LPDDR5X

. In fact if Strix Halo really is 2025 and with LPDDR5X that would mean it is arriving a bit too late. There has to be another reason behing this decision and i think it is because they want to align all Strix behind LPCAMM2, but if it is only AMD i dont know how successfull this will be, they need Intel too.

LPDDR6 won't come to market until 2026, I believe.

 

[Tigerick]

LPDDR5X = 8533
LPDDR6 = 12 800

So 128 bit LPDDR6 /= 256 bit LPDDR5X

How about try to calculate with 192-bit LPDDR6?

 

[Abwx]

Updated review of the 8600G/8700G, first reviews were made with a mobile dedicated firmware apparently that reduced the perfs, GPU perfs increase by 13% with the new firmware :

AMD Ryzen 8700G & 8600G im Test: Gaming, Multimedia, Apps

AMD Ryzen 7 8700G und Ryzen 5 8600G sind APUs mit Zen-4-CPU und RDNA-3-iGPU. Der Sprung gegenüber Ryzen 5000G mit Vega ist im Test groß.

www.computerbase.de

GPU perf of the 8600G/8700G are 1.7x and 2x the 5700G perfs respectively, 8600G being 2x faster than the 5600G.

 

[SteinFG]

Updated review of the 8600G/8700G, first reviews were made with a mobile dedicated firmware apparently that reduced the perfs, GPU perfs increase by 20% with the new firmware (26% for the 8600G) :

There are still issues after running the game for more than 300 seconds, they're waiting for fixes

 

[Abwx]

There are still issues after running the game for more than 300 seconds, they're waiting for fixes

Guess that they have extracted most of the missing perfs, the FPS at 40 img/s are the same at 400 seconds than before 300s although there s a peak at 46 img/s when they start the run, but that could be just an artifact with low rendering at the start.

2x the perfs in respect of the previous gen seems to be what was expected from PHX2 given that RAM bandwith is only 62% higher while the CU count x frequency uplift is about 2x, look like they gained most of the possible perfs.

 

[Asterox]

All three CPU-s are cooled with the same AMD stock Wraith stealth CPU cooler.

If this is correct data, in the Cinebench test the R5 5600G heats up to 89 degrees. The R5 8500G and R5 8600G do not go more than 76 degrees in the same test.

iGPU gaming R5 8500G vs R5 5600G. As expected, considering the system memory used+old Vega vs "small RDNA3 iGPU with very high iGPU frequency". DDR5 memory has a significantly higher bandwidth. In short, in this comparison R5 8500G usually offers minimum 10 and sometimes 20 more FPS.

Let's say that slower DDR5 memory 5200 or 5600mhz was used. Again 4CU RDNA2 iGPU at 2.8ghz would be faster vs old Vega 7CU iGPU at 1.9ghz.

R5 5600G, 32gb 3200mhz DDR4

R5 8500G/R5 8600G, 32gb 6000mhz DDR5

 

[DrMrLordX]

That's twice you've mentioned SIMM's in the past few hours. I feel like I've been yanked back into the 90's .

Yeah I'm sure my brain is trying to say something different and it keeps spitting out the wrong acronym.

edit: SODIMM hur dur

That is true for 4 and 8GB products low end, when we start talking of 16, 32 and 64GB products costs go up fast.
Specially when you need to build 3 diferent boards for the task, and even more if you also soldering an emmc.

Depends on how many different configs they're trying to sell.

 

[Shivansps]

All three CPU-s are cooled with the same AMD stock Wraith stealth CPU cooler.

If this is correct data, in the Cinebench test the R5 5600G heats up to 89 degrees. The R5 8500G and R5 8600G do not go more than 76 degrees in the same test.

iGPU gaming R5 8500G vs R5 5600G. As expected, considering the system memory used+old Vega vs "small RDNA3 iGPU with very high iGPU frequency". DDR5 memory has a significantly higher bandwidth. In short, in this comparison R5 8500G usually offers minimum 10 and sometimes 20 more FPS.

Let's say that slower DDR5 memory 5200 or 5600mhz was used. Again 4CU RDNA2 iGPU at 2.8ghz would be faster vs old Vega 7CU iGPU at 1.9ghz.

R5 5600G, 32gb 3200mhz DDR4

R5 8500G/R5 8600G, 32gb 6000mhz DDR5

Its not just that, see Alan Wake for example that just dosent run on Vega at all.

Now i really want to see a single channel vs dual channel on the 8500G.

 

[Asterox]

Its not just that, see Alan Wake for example that just dosent run on Vega at all.

Now i really want to see a single channel vs dual channel on the 8500G.

From what we've seen so far, on this Computerbase comparison the R5 8500G iGPU would probably average around 37-38 FPS.



Youtube comparison, if instead of the R5 5600G we insert the R7 5700G into the same system. Estimation, R5 8500G iGPU would probably push 5-10 FPS more vs the R7 5700G.

 

[SteinFG]

From what we've seen so far, on this Computerbase comparison the R5 8500G iGPU would probably average around 37-38 FPS.

Why? It seems like 8500G is ~1.4x faster than 5600G, so around 46fps.

 

[Asterox]

Why? It seems like 8500G is ~1.4x faster than 5600G, so around 46fps.

I estimate a little lower, to feel more comfortable(if I missed)when we soon see the real numbers.

If we look at this comparison, the R5 8500G iGPU will be faster than the R7 5700G. The only question is how much more FPS for the R5 8500G iGPU.

For classic HTPC computer, 2 x 8/16GB DDR5 memory is quite enough.16GB DDR5 6000mhz memory has a quite ok price.

Protis - Kingston DRAM 16GB 6000MT/s DDR5 CL36 DIMM (Kit of 2) FURY Beast Black EXPO EAN: 740617331721 KF560C36BBEK2-16

www.protis.hr

 

[DAPUNISHER]

8500G tuning in some impressive performance per watt -

https://www.phoronix.com/review/amd-ryzen-5-8500g