AMD's Phoenix 1 and Phoenix 2 APUs Differ in PCIe Lane Count, Affects NVMe Drive Performance and GPU PCIe Lane Count

[TheLostSwede]

At CES, AMD didn't give away too many technical details of its upcoming Ryzen 8000G-series APUs, but details are starting to trickle out and it's not all good news. As has been known for some time, AMD is using two different chips to make the Ryzen 8000G APUs and they're known as the Phoenix 1 and Phoenix 2, where the Phoenix 2 parts feature Zen 4c cores, which are not present in the Phoenix 1 APUs. This in and of itself shouldn't be a huge issue, although the Zen 4c CPU cores can be slightly slower in some tasks based on testing of AMD's EPYC server parts.

However, PCGamesN noticed that Gigabyte has posted the full specs for the B650E Aorus Elite X AX Ice motherboard and it looks like there's a much bigger difference between the Phoenix 1 and Phoenix 2 based APUs. Namely, the Phoenix 2 APUs have fewer PCIe lanes and as such are limited to two PCIe 4.0 lanes for the secondary NVMe slot. As if this wasn't bad enough, the Phoenix 2 APUs only have four PCIe 4.0 lanes for add-in GPUs, whereas the Phoenix 1 APUs have eight. This is very likely to lead to reduced performance if a higher-end GPU is used with such an APU. Note that this will vary depending on the motherboard design, but many B650/B650E boards feature a similar design with regards to the PCIe lanes coming from the CPU socket. Luckily, it's easy to avoid this issue, as the Ryzen 5 8600G and the Ryzen 7 8700G are both Phoenix 1 designs, whereas the Ryzen 5 8500G is the only Phoenix 2 design available in retail, as the Ryzen 3 8300G is an OEM only part.



View at TechPowerUp Main Site | Source

 

[R0H1T]

Which of these are Phoenix 2 btw

Also this is a braindead move to get zen4 desktop chips on 8xxx numbering scheme, I couldn't give 2 effs about OEM's but they should've maintained the relatively good status quo like they did with zen3

 

[mosin40]

Which of these are Phoenix 2 btw

Also this is a braindead move to get zen4 desktop chips on 8xxx numbering scheme, I couldn't give 2 effs about OEM's but they should've maintained the relatively good status quo like they did with zen3

8300G and 8500G. First one only has Zen4c cores, while the latter has 2 Zen4 cores + 4 Zen4c cores. Also, 8300G is OEM only. Those are more catered to office use imo.

 

[TheLostSwede]

8300G and 8500G. First one only has Zen4c cores, while the latter has 2 Zen4 cores + 4 Zen4c cores. Also, 8300G is OEM only. Those are more catered to office use imo.

Both of those are the ones with Zen 4c cores.

 

[AnotherReader]

This would be ok if they had restricted the 8500G to OEMs as well. Four lanes for GPUs would limit the use of Phoenix 2. I really don't understand AMD's penny pinching sometimes. Sixteen lanes of PCIe 5 require 2.4 mm^2 on the Alder Lake-S die.

 

[Zubasa]

This would be ok if they had restricted the 8500G to OEMs as well. Four lanes for GPUs would limit the use of Phoenix 2. I really don't understand AMD's penny pinching sometimes. Sixteen lanes of PCIe 5 require 2.4 mm^2 on the Alder Lake-S die.

It has more to do with how the lanes are routed by the mother board then the actual lane count.
Phoenix2 has 14 Gen4 lanes with 10 being usable on AM5 (4 lanes went to the chipset). In theory the board can do an x8 slot and a x2 nvme m.2.
But since the board are releaseds with the 24+4 lanes Raphael in mind, many board have an 2nd pcie x4 / nvme slot wired into the cpu.
So they either ditch those and risk customers complaing why certain slots don't work. Or they do this janky thing where every slot ended up gimped.

 

[Assimilator]

So they either ditch those and risk customers complaing why certain slots don't work. Or they do this janky thing where every slot ended up gimped.

Or... they could not release such an obviously intentionally castrated product. Even the latest Intel dual-core, which is rubbish, has 20 lanes of PCIe 5.0! @AnotherReader is completely correct, this is the most despicable type of nickel-and-diming from AMD in order to force people who just want more IO, to buy higher-priced products. I'm starting to hope that Arrow Lake gives AMD the kick up the arse that it deserves.

 

[TheLostSwede]

It has more to do with how the lanes are routed by the mother board then the actual lane count.
Phoenix2 has 14 Gen4 lanes with 10 being usable on AM5 (4 lanes went to the chipset). In theory the board can do an x8 slot and a x2 nvme m.2.
But since the board are releaseds with the 24+4 lanes Raphael in mind, many board have an 2nd pcie x4 / nvme slot wired into the cpu.
So they either ditch those and risk customers complaing why certain slots don't work. Or they do this janky thing where every slot ended up gimped.

Updated the article to make it a bit more clear that it does at least to some degree depend on the motherboard layout.

Or... they could not release such an obviously intentionally castrated product. Even the latest Intel dual-core, which is rubbish, has 20 lanes of PCIe 5.0! @AnotherReader is completely correct, this is the most despicable type of nickel-and-diming from AMD in order to force people who just want more IO, to buy higher-priced products. I'm starting to hope that Arrow Lake gives AMD the kick up the arse that it deserves.

Four PCIe lanes are reserved for the internal USB4 host controller, which is part of the problem.

 

[_Flare]

Phoenix are ment to be used as the APUs they are.
Nobody should use any Phoenix where a 7500F upwards paired with a dGPU is of more desire, obviously.
And with $176 SEP the 8500G seems not cheaper than the mentioned Raphael 7500F.

 

[Assimilator]

Four PCIe lanes are reserved for the internal USB4 host controller, which is part of the problem.

See, AMD, the way this works is that you add features. You do not add features and take others away at the same time.

 

[TheLostSwede]

Phoenix are ment to be used as the APUs they are.
Nobody should use any Phoenix where a 7500F upwards paired with a dGPU is of more desire, obviously.
And with $176 SEP the 8500G seems not cheaper than the mentioned Raphael 7500F.

AMD disagrees with you.

See, AMD, the way this works is that you add features. You do not add features and take others away at the same time.

I don't disagree at all with you here.

See, AMD, the way this works is that you add features. You do not add features and take others away at the same time.

I don't disagree at all with you here.

 

[Tek-Check]

Or... they could not release such an obviously intentionally castrated product. Even the latest Intel dual-core, which is rubbish, has 20 lanes of PCIe 5.0! @AnotherReader is completely correct, this is the most despicable type of nickel-and-diming from AMD in order to force people who just want more IO, to buy higher-priced products. I'm starting to hope that Arrow Lake gives AMD the kick up the arse that it deserves.

- false analogy and conspiratorial nonsense there
- Intel released 300 on salvaged dies that did not qualify for other i3 CPUs. Even 50 Gen5 lanes would not make it any better. Its connectivity is a total overkill for the CPU itself
- no one in their right mind would ever buy 300 CPU and pair it with speedy and expensive NVMe drives or GPUs. So, yes, it has more lanes, but no one really cares.

- let's clarify: Phoenix 2 is not a new chip. It has been around as 7440U and 7540U for entry school laptops and we knew last year already its spec with 10+4 PCIe lanes.
- now, Phoenix 2 is 8440U, 8540U, 8500G and 8300G
- Phoenix 2 is compact, very cheap to produce, with high yields, it has better iGPU than Intel 300 and enough connectivity for entry systems, ideally with A620 boards
- when paired with better B650 or X670 boards, Phoenix 2 chip will look like gimped APU due to board design decisions for more capable CPUs/APUs.
- it's up to motherboard vendor to expose what a SKU has to offer and Gigabyte did not maximize APU's connectivity here, to remain more compatible with other CPUs/APUs
- so, all your complaints to Gigabyte, please, and not to AMD

Gigabyte could have exposed more connectivity baked into APU + what is on B650 chipset:
- two USB4 ports, each at 40 Gbps - Intel 300, on the other hand, does not have integrated Thunderbolt 4
- DP 2.1 and HDMI 2.1 FRL ports for monitors - Intel 300 has older video standards in UHD710 graphics
- reconfigure PCIe 4.0 as x8 for GPU instead of offering two x4 (this was a trade-off in order for the board to be compatible with other CPUs/APUs)
- on the chipset, Gigabyte could have wired WiFi, LAN and two SATA ports on one PCIe x4 PHY, which would enable another Gen4 SSD (again, trade-off)

- personally, I'd never pair 8500G/8300G with such expensive ATX B650 board. Why would anyone do this? They can, and it will work, but it's pointless.
- 8600G, 8700G and other CPUs make much better match
- 8500G and 8300G are more suitable for A620 chipsets and entry B650 ITX boards in small factor entry systems. Pretty simple stuff. AMD hopes that DIY buyers can make reasonable matches of APUs and boards.

See, AMD, the way this works is that you add features. You do not add features and take others away at the same time.

And what is it that they "took away" on Phoenix 2 chip?
Have a look how packed Phoenix 2 chip is with 137 mm2 size.

 

[AusWolf]

No one is going to pair a high-end GPU with a Phoenix 2 APU, so I don't see this being an issue. Interesting to note, though.

 

[sLowEnd]

8300G and 8500G. First one only has Zen4c cores, while the latter has 2 Zen4 cores + 4 Zen4c cores. Also, 8300G is OEM only. Those are more catered to office use imo.

The regular 7000 series Zen 4 chips would do fine in a office PC too, though.

 

[john_]

This is the case for ages. Even on AM4, Athlon APUs had fewer lanes than Ryzen APUs who had fewer lanes than CPUs.

X470 GAMING PLUS MAX | Motherboard | MSI Global

Intel is selling hybrid CPUs with now 3 types of different cores. But, yeah, let's talk about the possibility of someone buying an 8500G to use it with an RTX 4090 and two hi end PCIe 4.0 NVMe SSDs instead.
I mean, obviously someone will buy an APU to use with a hi end GPU. Right?

 

[KellyNyanbinary]

At least it’s PCIe 4.0 and not 3.0.

 

[thestryker6]

As far as I'm aware the 4 lanes for NVMe aren't optional no matter what chipset so they're going to end up with 4 lanes to that primary slot unless someone designs a board specifically around these which seems unlikely. It's unfortunate that this negatively impacts upgradability for the 8500G, but given that they're just the mobile die there isn't any option other than just not releasing this outside of OEM which I think would have been the better choice.

 

[AusWolf]

As far as I'm aware the 4 lanes for NVMe aren't optional no matter what chipset so they're going to end up with 4 lanes to that primary slot unless someone designs a board specifically around these which seems unlikely. It's unfortunate that this negatively impacts upgradability for the 8500G, but given that they're just the mobile die there isn't any option other than just not releasing this outside of OEM which I think would have been the better choice.

Personally, I think the measly 4 CUs in the iGPU is a much bigger turnoff than the 2 lanes for the secondary m.2 slot that most of the potential buyers aren't gonna use anyway.

 

[thestryker6]

Personally, I think the measly 4 CUs in the iGPU is a much bigger turnoff than the 2 lanes for the secondary m.2 slot that most of the potential buyers aren't gonna use anyway.

The 4 lanes for the primary PCIe slot is the problem not anything with the NVMe.

Agreed on the 4 CU IGP being problematic despite being double that of regular Zen 4 desktop parts.

 

[AusWolf]

The 4 lanes for the primary PCIe slot is the problem not anything with the NVMe.

Agreed.

Agreed on the 4 CU IGP being problematic despite being double that of regular Zen 4 desktop parts.

Double of nothing is still nothing, unfortunately. If Phoenix 1 has 12 CUs, then Phoenix 2 should have 8 to be usable as more than a display adapter / video accelerator, in my opinion.

 

[trsttte]

Are we really pretending this matters? It's a laptop cpu, a lower end one at that, that amd is now also selling as a desktop cpu because we're at an age where that's more than enough for the thin clients office drones use.

Even the latest Intel dual-core, which is rubbish, has 20 lanes of PCIe 5.0! @AnotherReader is completely correct, this is the most despicable type of nickel-and-diming from AMD in order to force people who just want more IO, to buy higher-priced products.

It's a laptop part designed for laptops. Despicable nickel and diming would be if they chose to cut something that is there, i.e. if Intel had cut the lanes to fit the lower end market when they didn't need to or how Intel does lock ECC support and other features for workstation priced products. More PCIe lanes do not exist on this AMD chip, if you need them buy something better - you don't even need to spend that much more, a 7600 can be had for 210$ with almost double the cache.