AMD Ryzen 8000G Socket AM5 Desktop APU Lineup Detailed

[btarunr]

Here is our first look at the higher end of AMD's Ryzen 8000G series Socket AM5 desktop APU lineup. The company is planning to bring its 4 nm "Phoenix" and "Phoenix 2" monolithic silicon to the socketed desktop platform, to cover two distinct markets. Models based on the larger "Phoenix" silicon cater to the market that wants a sufficiently powerful CPU, but with a powerful iGPU that's fit for entry-level gaming, or graphics-intensive productivity tasks; whereas the smaller "Phoenix 2" silicon ties up the lower end of AMD's AM5 desktop processor stack, as it probably has a lower bill of materials than a "Raphael" multi-chip module.

The lineup is led by the Ryzen 7 8700G, a direct successor to the Ryzen 7 5700G "Cezanne." This chip gets the full 8-core/16-thread "Zen 4" CPU, along with its 16 MB shared L3 cache; and the full featured Radeon 780M iGPU with its 12 compute units worth 768 stream processors. The CPU features a maximum boost frequency of 4.20 GHz. This is followed by the Ryzen 5 8600G, which is based on the same "Phoenix" silicon as the 8700G, but with 6 out of 8 "Zen 4" cores enabled, and a maximum CPU boost frequency of 4.35 GHz, and the 16 MB L3 cache left untouched. It's likely that the Radeon 780M is unchanged from the 8700G.



Update 13:59 UTC: A CPU-Z screenshot of the Ryzen 7 8700G surfaced, which confirms that it features the maxed out Radeon 780M iGPU

Things get interesting with the Ryzen 5 8500G. This chip is rumored to be based on the smaller "Phoenix 2" silicon. While its CPU is 6-core/12-thread, two of these are "Zen 4," and can sustain higher boost frequencies of up to 3.35 GHz, while four of them are smaller "Zen 4c" cores that run at a lower maximum boost frequency. Both CPU core types feature an identical IPC, ISA, as well as SMT; and AMD's software based OS scheduler optimizations will simply mark the two "Zen 4" cores as UEFI CPPC "preferred cores," so they get priority in processing workloads. This chip gets the full 16 MB of L3 cache present on the silicon.

At the entry level is the Ryzen 3 8300G. This is a quad-core chip based on "Phoenix 2," in that two out of four "Zen 4c" cores are disabled, leaving it with two "Zen 4" cores, and two "Zen 4c." Just like the 8500G, the OS scheduler is made to prefer the two "Zen 4" cores. AMD has also reduced the L3 cache size to 8 MB. Both the 8500G and 8300G feature a physically smaller iGPU that's branded as the Radeon 740M. It only gets 4 compute units (256 stream processors). All four chips feature a TDP of 65 W, and a possible 90 W PPT, which should give them plenty of boost residency compared to their mobile-segment siblings.

In addition to these four chips, AMD is preparing the Ryzen 5 PRO 8500G, which is likely based on the "Phoenix" silicon, with 6 "Zen 4" CPU cores, 16 MB of L3 cache, and a Radeon 780M iGPU. This chip gets the full AMD PRO feature-set, and is designed for commercial desktops.

We still don't see any concrete evidence about AMD enabling the on-chip XDNA Ryzen AI NPU for at least the 8700G, 8600G, and PRO 8500G. "Phoenix" has it, while "Phoenix 2" physically lacks it.

View at TechPowerUp Main Site | Source

 

[Arco]

My dad needs this for work for sure!

 

[ymdhis]

I thought the 8700G would get 8CUs and the 8600G 7CUs, same as Cezanne. Or will they have the full 780M with all cores even on the 8600G?

 

[_JP_]

From the xweet:

• Ryzen 7 8700G (4.2GHz, 65W, 16MB, 8, B2)
• Ryzen 5 8600G (4.35GHz, 65W, 16MB, 8, B2)
• Ryzen 5 8500G (3.35GHz, 65W, 16MB, 6, B2)
  • Ryzen 5 PRO 8500G (3.55GHz, 65W, 16MB, 6, B2)
• Ryzen 3 8300G (3.45GHz, 65W, 8MB, 4, B2)

Fine, but at what cost in this economy?
If this pairing beats the value/perf. ratio of the regular Ryzen 7 with a discreet RX6400 (or RX6500, I'm not sure there the 780M sits ATM), it's a win for integration and businesses, that are already considering ignoring the scalability/upgradability.
That was the whole point of the APUs, be a decent x86-64 offer with a competing iGPU for low demand 2D/3D, with a decrease in cost compared to the discreet option. The pricing of the 5700G alone made me think twice and I just ended-up recommending a regular CPU+GPU combo to people.

 

[Denver]

I would love to build a mini PC with one of these.

 

[usiname]

I thought the 8700G would get 8CUs and the 8600G 7CUs, same as Cezanne. Or will they have the full 780M with all cores even on the 8600G?

8700G - 12CUs RDNA3
8600G - 8CUs RDNA3

 

[Denver]

8700G - 12CUs RDNA3
8600G - 8CUs RDNA3

Yes, and there would be no need to cut down the chip after extending the APU line to the desktop...

 

[JaymondoGB]

Have they finally allowed PCI4 or even 5 to the chipset ? The old "G" were hobbled by PCIE3

 

[damric]

Hopefully they have a strong memory controller for some very fast RAM speeds. That 12cu will really need that bandwidth!

 

[G777]

Those boost clocks seem low; even the 7840U can boost to 5.1GHz. Perhaps those numbers are the base clocks?

 

[Denver]

Those boost clocks seem low; even the 7840U can boost to 5.1GHz. Perhaps those numbers are the base clocks?

Yeap, this information is incorrect.

 

[physicsjunky]

I hope the memory controller is strong enough to handle 8000Mhz ram.

 

[P4-630]

I hope the memory controller is strong enough to handle 8000Mhz ram.

Doubt...

 

[ncrs]

Doubt...

I wouldn't be so hasty. The APU memory controllers are of a different design - Pheonix in laptops supports DDR5-5600 (while current desktop AM5 officially goes up to 5200) and LPDDR5X-7500. Since they are integrated onto one die, they are using a different manufacturing process from previous AM5 designs.
Also after recent AGESA updates Buildzoid has managed to get 8000MHz running on a 7900X.

 

[JohH]

The Phoenix IMC supports LPDDR5x-7500. But I doubt that will translate to regular DDR5.

 

[Lew Zealand]

From the xweet:

• Ryzen 7 8700G (4.2GHz, 65W, 16MB, 8, B2)
• Ryzen 5 8600G (4.35GHz, 65W, 16MB, 8, B2)
• Ryzen 5 8500G (3.35GHz, 65W, 16MB, 6, B2)
  • Ryzen 5 PRO 8500G (3.55GHz, 65W, 16MB, 6, B2)
• Ryzen 3 8300G (3.45GHz, 65W, 8MB, 4, B2)

Fine, but at what cost in this economy?
If this pairing beats the value/perf. ratio of the regular Ryzen 7 with a discreet RX6400 (or RX6500, I'm not sure there the 780M sits ATM), it's a win for integration and businesses, that are already considering ignoring the scalability/upgradability.
That was the whole point of the APUs, be a decent x86-64 offer with a competing iGPU for low demand 2D/3D, with a decrease in cost compared to the discreet option. The pricing of the 5700G alone made me think twice and I just ended-up recommending a regular CPU+GPU combo to people.

The 780M is still behind the RX 6400 as it has the same core count but with slower shared memory. It performs at about 70-80% of the 6400 in most tests. I've gamed with the 6400 and 70-80% of that from an iGPU is reasonable performance.

 

[Selaya]

so their apus following the same naming scheme as cpus to reduce confusion lasted a whole whopping ONE generation.

.

The 780M is still behind the RX 6400 as it has the same core count but with slower shared memory. It performs at about 70-80% of the 6400 in most tests. I've gamed with the 6400 and 70-80% of that from an iGPU is reasonable performance.

im sure on desktop w/ a proper OC 'nd stuff you can make it match the 6400 rather easily

 

[kapone32]

The 780M is still behind the RX 6400 as it has the same core count but with slower shared memory. It performs at about 70-80% of the 6400 in most tests. I've gamed with the 6400 and 70-80% of that from an iGPU is reasonable performance.

I have been watching those Google videos about Gaming on a Chromebook and having 120hz panels makes that claim viable. As long as these APUs support HDMI 2.1 people with 4K 120Hz Freesync TVs are going to be happy switching out their Vega 5700Gs for these. Even with a A620 board you will be good. I expect these to sell well and work well in the right scenarios.

 

[ymdhis]

12CU on the 8700G, holy moly. That's a monster of an APU. Hope it will be affordable too.

so their apus following the same naming scheme as cpus to reduce confusion lasted a whole whopping ONE generation.

One could argue that having 5600 5600G 5600X 5600X3D is also confusing, just in a different way - in fact if you take the G chip out, the rest would make sense, so grouping the G chips under a different naming scheme would in fact reduce confusion (and also help sales because higher number == better).

I wouldn't be so hasty. The APU memory controllers are of a different design - Pheonix in laptops supports DDR5-5600 (while current desktop AM5 officially goes up to 5200) and LPDDR5X-7500. Since they are integrated onto one die, they are using a different manufacturing process from previous AM5 designs.
Also after recent AGESA updates Buildzoid has managed to get 8000MHz running on a 7900X.

When I switched from a 3600 to a 5600G, the memory controller used significantly less power, most likely because it was integrated instead of being on a chiplet. And the one in Phoenix is also on a smaller node compared to the 7xxx CPUs. At the very least it will be interesting to see what these can do and how far they can be pushed.

 

[kapone32]

12CU on the 8700G, holy moly. That's a monster of an APU. Hope it will be affordable too.


One could argue that having 5600 5600G 5600X 5600X3D is also confusing, just in a different way - in fact if you take the G chip out, the rest would make sense, so grouping the G chips under a different naming scheme would in fact reduce confusion (and also help sales because higher number == better).


When I switched from a 3600 to a 5600G, the memory controller used significantly less power, most likely because it was integrated instead of being on a chiplet. And the one in Phoenix is also on a smaller node compared to the 7xxx CPUs. At the very least it will be interesting to see what these can do and how far they can be pushed.

I want to see what kind of GPU clock this can go to.

 

[AnotherReader]

I wouldn't be so hasty. The APU memory controllers are of a different design - Pheonix in laptops supports DDR5-5600 (while current desktop AM5 officially goes up to 5200) and LPDDR5X-7500. Since they are integrated onto one die, they are using a different manufacturing process from previous AM5 designs.
Also after recent AGESA updates Buildzoid has managed to get 8000MHz running on a 7900X.

In addition, the IGP has a very wide link to the memory controller: enough for 11.2 GT/s DDR5 which is enough to ensure that it'll be limited by the memory controller rather than the fabric.

Ryzen 7840HS (Phoenix)	Ryzen 4800H (Renoir)
CPU Load	2 GHz	1.6 GHz
GPU Load	1.4 GHz	1.6 GHz
Mixed CPU/GPU Load	1.6 GHz	1.6 GHz

Infinity Fabric clock when running CPU/GPU bandwidth tests

The GPU has four 32B/cycle ports to fabric, letting it get enough memory bandwidth even at low fabric clock.

In addition, the memory controller seems to be optimized for the IGP. The IGP gets higher bandwidth than the CPU and enjoys more bandwidth from DDR5 5600 than a 7950X from DDR5 6000.

Contrast this with the 7950X below

 

[ymdhis]

I want to see what kind of GPU clock this can go to.

Laptop Phoenix goes up to 2.6-2.8GHz depending on model.

In addition, the IGP has a very wide link to the memory controller: enough for 11.2 GT/s DDR5 which is enough to ensure that it'll be limited by the memory controller rather than the fabric.

In addition, the memory controller seems to be optimized for the IGP. The IGP gets higher bandwidth than the CPU and enjoys more bandwidth from DDR5 5600 than a 7950X from DDR5 6000.

That's still just ~80 GB/s while the RX6400 has 128 GB/s. So the Phoenix IGP has to be more effective and be able to do more with far less bandwidth, but it has less cache on it and has to share the memory bandwidth with the CPU. It is limited by memory speed, the only question is if the APU version is better than the laptop version or not.

There are also comparison videos online, specifically the 780M on a mini pc vs a RX6400:

The performance ranges from them being equal to the RX6400 being 60% faster.

 

[AnotherReader]

Laptop Phoenix goes up to 2.6-2.8GHz depending on model.


That's still just ~80 GB/s while the RX6400 has 128 GB/s. So the Phoenix IGP has to be more effective and be able to do more with far less bandwidth, but it has less cache on it and has to share the memory bandwidth with the CPU. It is limited by memory speed, the only question is if the APU version is better than the laptop version or not.

There are also comparison videos online, specifically the 780M on a mini pc vs a RX6400:

The performance ranges from them being equal to the RX6400 being 60% faster.

It should do better with faster memory, but it's unlikely to match the RX 6400 which has more memory bandwidth and a 16 MB last level cache to augment that bandwidth.

 

[Count von Schwalbe]

It should do better with faster memory, but it's unlikely to match the RX 6400 which has more memory bandwidth and a 16 MB last level cache to augment that bandwidth.

From your link:

Past L2, Phoenix’s iGPU suffers much higher latency because it lacks an Infinity Cache. (LP)DDR5 gives the iGPU a better bandwidth per SIMD operation ratio than that of desktop GPUs, so dedicating area towards another level of cache can’t be justified.

 

[AnotherReader]

From your link:

Given that the RX 6400 has the same number of compute units, I believe that the 780M would do better with a dedicated last level cache. Desktop parts won't be configured with LPDDR5, but that doesn't seem to be a big loss as DDR5 5600 is very close to LPDDR5 6400 in bandwidth.