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 My advice to everyone: treat the old APU rumours as being out of date.

 

[jpiniero]

I really want to see the 8500G compared to a 2400G/3400G to see what the 5/6-years worth of improvements looks like.

At least from a quick glance in Geekbench, probably 20% faster than the 5700G in ST and maybe a push in MT.

 

[igor_kavinski]

At least from a quick glance in Geekbench, probably 20% faster than the 5700G in ST and maybe a push in MT.

I think he's more interested in the GPU perf improvement.

 

[Thunder 57]

Yes, as i said, Mendocino is tailored for diferent market, the same as ADL-N, and as a result they cant launch it on AM5 due to the cutdowns. Raven2 was on AM4, budget mobile devices and the embedded market.


I wonder if AMD is going to allow reviewers to compare it with older apu from 2200G to a 5700G. I really want to see the 8500G compared to a 2400G/3400G to see what the 5/6-years worth of improvements looks like.

Probably not great at all, but why should it look good? You are comparing 4 CU's to 8 or 11. You should be comparing those to an 8600G or 8700G. An 8700G should be much better than a 5700G.

 

[Abwx]

Probably not great at all, but why should it look good? You are comparing 4 CU's to 8 or 11. You should be comparing those to an 8600G or 8700G. An 8700G should be much better than a 5700G.

Compared to a 8700G you can increase a 8500G s 4CUs GPU frequency by roughly 50% at same power, add that frequency scale better than CU count and that at 4 units it wont be much RAM bandwith starved.

At the end of the day this could get you in the 60% of the 8700G gaming perf and should do well in respect of older siblings like the 5700G.

 

[TESKATLIPOKA]

Another handheld with 7840U -> AYANEO Slide (Videocardz)

I like that It has a small keyboard. Track point or touchpad would have been great and also back stand. The 46Wh battery is mediocre.

 

[jpiniero]

I think he's more interested in the GPU perf improvement.

Just sayin, but the 6700 XT is on sale at Newegg for $299 right now. I imagine it would be a tad faster than the Phoenix IGP.

 

[Thunder 57]

Another handheld with 7840U -> AYANEO Slide (Videocardz)

I like that It has a small keyboard. Track point or touchpad would have been great and also back stand. The 46Wh battery is mediocre.

Those prices are insane.

 

[eek2121]

Is anything known about Hawk Point? (Other than its existence on roadmaps showing 1Q2024 availability and a bump in XDNA TOPS.)

There's a Geekbench entry from yesterday that's making the rounds with what appears to be 2024 version of the 7940HS. What's notable is that it scores quite a bit higher in Geekbench 6.2.1 than the 7940HS (+100 points in ST, +1400 MT) despite apparently having same base/boost frequency and same power consumption.

My initial thought was that AMD simply bumped the TDP or that the processor was a nicely binned selection like the Lenovo S models, but the AuthenticAMD CPU identifier is different, so there *might* be something more in it. The 7940HS (and other 7040 family "big" Phoenix) features family 25, model 116, stepping 1. This new chip is model 117, stepping 2.

Standard Standard - Geekbench

Benchmark results for a Standard Standard with an AMD Eng Sample: 100-000001319-50_Y processor.

browser.geekbench.com

I haven’t been following things lately, but if it is respin, it may be on an improved process. Peak clocks won’t change much, but multicore workloads could improve substantially with any power savings at all.

Probably not great at all, but why should it look good? You are comparing 4 CU's to 8 or 11. You should be comparing those to an 8600G or 8700G. An 8700G should be much better than a 5700G.

It is worth pointing out that RDNA1 and RDNA2 is not as fast as RDNA3/3.5/4. I bring this up because at equal clocks and a similar power budget, 8 RDNA3 CUs should outperform 8 RDNA 2 CUs quite easily. They should even beat 11-12 RDNA2 CUs depending on the task, especially if code is optimized for the improvements RDNA 3 brought.

The issue with RDNA3 was obviously efficiency. It is currently unknown if upcoming APUs have similar efficiency problems.

 

[SteinFG]

I wonder if AMD is going to allow reviewers to compare it with older apu from 2200G to a 5700G.

Those APUs are not as important as you think. Reviewers can compare anything with anything when it comes to pc parts

 

[TESKATLIPOKA]

It is worth pointing out that RDNA1 and RDNA2 is not as fast as RDNA3/3.5/4. I bring this up because at equal clocks and a similar power budget, 8 RDNA3 CUs should outperform 8 RDNA 2 CUs quite easily. They should even beat 11-12 RDNA2 CUs depending on the task, especially if code is optimized for the improvements RDNA 3 brought.

RDNA3 is faster than RDNA2 in some games, but expecting 8CU RDNA3 to beat 11-12CU RDNA2 is pretty unrealistic. Shaders have higher throughput, but there is still Texel, Pixel fillrate and BW, which will become a bottleneck.

Best results: RX 7600 vs RX 6650XT
Control: +17.2% at 1080p, but at higher resolution significantly slower than N23 for some reason. (TPU)
Cyberpunk : +17% at 1440p (TPU)
Callisto Protocol: +16% at 1080p (TPU)
Warhammer III: +24% at 1080p (TPU)
In the other tested games the difference was lower or even slower.

 

[TESKATLIPOKA]

Those prices are insane.

Price is high, but at least you can configure It up to 64GB + 4TB.
In Legion GO or Asus Ally at most you can change SSD, but only by opening the handheld. RAM is soldered, so 16GB and nothing else.

Next year Strix Point will be out, and I expect we will finally see 32GB RAM + 2TB in non-Chinese handhelds.

 

[Thunder 57]

Price is high, but at least you can configure It up to 64GB + 4TB.
In Legion GO or Asus Ally at most you can change SSD, but only by opening the handheld. RAM is soldered, so 16GB and nothing else.

Next year Strix Point will be out, and I expect we will finally see 32GB RAM + 2TB in non-Chinese handhelds.

64GB is completely unnecessary. And for $1600? It's a joke.

 

[FlameTail]

I haven’t been following things lately, but if it is respin, it may be on an improved process. Peak clocks won’t change much, but multicore workloads could improve substantially with any power savings at all.

Is Phoenix on N4?

Then Hawk Point may be on N4P

 

[Joe NYC]

I haven’t been following things lately, but if it is respin, it may be on an improved process. Peak clocks won’t change much, but multicore workloads could improve substantially with any power savings at all.

It could be a re-spin with new, larger AI unit, to comply with Windows 12 requirements.

 

[TESKATLIPOKA]

64GB is completely unnecessary. And for $1600? It's a joke.

I like that such option exist even If It's an overkill, much better than being limited to 16GB.

Price is certainly high, but that was expected. They don't have the volume of Asus or Lenovo.
Asus or Lenovo don't give you such an option so It's hard to tell what would have been their price.

 

[FlameTail]

It could be a re-spin with new, larger AI unit, to comply with Windows 12 requirements.

MLID theorised they simply boosted the clock speed to meet those requirements.

Apparently Win12 requires minimum 45 TOPS of AI compute power

Phoenix has a 10 TOPS AI engine.

MLID said Hawk Point will have a 16 TOPS AI engine. He is theorising AMD achieved that by increasing the clock speed by 60%

Supposedly the 16 TOPS of XDNA and the rest of the TOPS from RDNA combined is sufficient to get them above the 45 TOPS requirement.

 

[Joe NYC]

MLID theorised they simply boosted the clock speed to meet those requirements.

Apparently Win12 requires minimum 45 TOPS of AI compute power

Phoenix has a 10 TOPS AI engine.

MLID said Hawk Point will have a 16 TOPS AI engine. He is theorising AMD achieved that by increasing the clock speed by 60%

Supposedly the 16 TOPS of XDNA and the rest of the TOPS from RDNA combined is sufficient to get them above the 45 TOPS requirement.

I came across that.

Since AMD is planning to hijack part of the Mi300 announcement to announce Hawk Point, I think it is going to be a bigger deal than just + 60% increase of clock speed on the AI / iGPU units (which may also be easier said than done).

Which is why I think it AMD may have increased the size of these units.

 

[DrMrLordX]

Apparently Win12 requires minimum 45 TOPS of AI compute power

That's going to be a dealbreaker for a lot of people. As in anyone not using modern hardware like Hawk Point or an overclocked Phoenix or whatever.

 

[Kepler_L2]

Is Phoenix on N4?

Then Hawk Point may be on N4P

PHX1 was already on N4P.

 

[TESKATLIPOKA]

I came across that.

Since AMD is planning to hijack part of the Mi300 announcement to announce Hawk Point, I think it is going to be a bigger deal than just + 60% increase of clock speed on the AI / iGPU units (which may also be easier said than done).

Which is why I think it AMD may have increased the size of these units.

You can forget 60% increase in IGP's frequency.
Boost is already at 2.8GHz, for which you need >=45W.
Another 60% on top of that would mean 4.5GHz boost! Totally unrealistic.

 

[TESKATLIPOKA]

PHX1 was already on N4P.

Are you sure?
Then those leaks about supposed performance of next gen are unrealistic.
Strix Point at 35W would never manage 2000pts in CB2024. The same 1419pts as R9 7900 would be already great at 1/2 TDP.

 

[qmech]

You can forget 60% increase in IGP's frequency.
Boost is already at 2.8GHz, for which you need >=45W.
Another 60% on top of that would mean 4.5GHz boost! Totally unrealistic.

From the leaked benchmark scores, there's no reason to expect any change in GPU frequency.

The XDNA throughput is 60% higher according to the roadmap, but the unit is still labeled XDNA 1, leading to speculation that the clock for that domain has simply been increased 60%. Considering the ridiculous size of the unit, increasing throughput by increasing the width would seem less likely.

 

[FlameTail]

Considering the ridiculous size of the unit, increasing throughput by increasing the width would seem less likely.

You are calling that 10 TOPS unit's size 'ridiculous' ?

What then, have you to say about the 45 TOPS NPU in the Snapdragon X Elite?

 

[qmech]

You are calling that 10 TOPS unit's size 'ridiculous' ?

What then, have you to say about the 45 TOPS NPU in the Snapdragon X Elite?

The *size* of the XDNA unit in Phoenix absolutely is ridiculous. I said nothing about the *performance*. There is a reason it was dropped in little Phoenix / Phoenix2.

That said, you may have been misled about the performance of the NPU in the X Elite. Some of the reporting on the X Elite announcement stated that the 45 TOPS throughput is for INT4, not INT8. None of the reporting that I have seen specifically claims INT8 conditions.

 

[Fjodor2001]

What's everyone's expectations with regards to 7800X3D vs 8700G?

1. ST performance difference?
2. MT performance difference?
3. Power consumption & perf/watt difference?
4. iGPU performance difference?
5. Idle power consumption difference?
6. Price difference?
7. Other differences?

Please share your thoughts.