Quote[..] It's like the APU only silicon are only relegated to NUCs/handhelds.

Notebooks with dGPU are neither here, not there: way too expensive for the gaming performance one gets vs a desktop build, too heavy to carry every day (2.5kg (for good cooling) vs 1-1.5kg with APU only) and when one wants to play in all quietness, a desktop PC at home is fine; and if you are on the move, enjoy the time/scenery, instead of staring at a small screen, close and tireding to the eyes and game at home on an OLED 4K 120Hz 1500nits HDR TV.

Judging from last year's releases, APU only notebooks should come out Q4 this year, Q1 next year for 0.7k to 1k bucks.

 

I have to say that the memory bandwidth from the GPU-Z screenshot is wrong.
Quad Channel (128 bit) DDR5-5600 should have like 89.6GB/s of memory bandwidth.

Although I agree that we should have somewhat faster memory controller, I don't really see the point of crazy 400GB/s bandwidth like the M1/2 Max on something like Zen4 45HX or Zen4 40HS/U.

I just hope that they (AMD/Intel) can optimize their chip to be able to fully take advantage of LPDDR5X-8533 (136GB/s for 128bit memory bus)

@NikoB:

But why are you comparing 45HX zen 4 series chips to Apple M2 is what I don't get. These chips are marketed to completely different segments. One is a 55w desktop chip packaged into laptop form factor and other isn't. You're also comparing a platform that uses socketed memory modules to one that uses LPDDR5. This isn't a like for like comparison, so obviously they'll perform / behave very differently.

Also, not too sure how accurate this is (as gpu-z can often misreport data) but this screenshot notebookcheck.(dot)net/fileadmin/Notebooks/Beelink/GTR7_7840HS/gpu1(dot)png shows 179 GB/s bandwidth on a zen 4 chip which seems pretty respectable to me.

x86 isn't going away anytime soon. As long as Microsoft keeps refusing to build a performant x86-to-ARM compatibility translation layer, as long as android remains the defacto standard platform for developing arm apps (which is debugging hell for dev's and has horrendous gpu drivers) and as long as modern emu's lack netplay / online multiplayer within their arm/android builds - x86 is here to stay.

Quote from: need PhoenixAPU-only test on July 25, 2023, 08:21:54

QuoteYet the difference in battery life is massive, M2 macbooks are pulling 16-17 hours easily while zen 4 laptops are barely pulling 7-8 hrs

Are there any AMD Phoenix APU-only [without dGPU] notebook reviews? Preferable one battery runtime test with APU limited to 15 Watt [7840U], then to 20 Watt and then to 25-28 Watt (U = 15W to 28W).
There was a recent news how most notebook companies sold 15% less, but APPLE sold 10% more. Maybe the long battery runtime is part of it.

Not that I know of and tbh apart from worse runtimes I prefer dgpu notebooks as they usually come with superior cooling. Even if such reviews existed it'd be pretty pointless as in actual stores the few zen 4 laptops that are available all come with dgpu. It's like the APU only silicon are only relegated to NUCs/handhelds.

Quote from: BokiN on July 24, 2023, 21:45:02Don't understand where this "slow" memory controller is coming from.

The shame of AMD is that, as I wrote about many times, that in the 45 Zen4 series, they have a total required bandwidth for simultaneous service at the peak of 28 pci-e 5.0 lines exceeds 110Gbyte/s, and real RAM works at best at a speed of 70Gbyte/s, and only for the soldered lpddr5 variant, and in the case of memory in slots, everything is much worse.

It's just nonsense - to make an SoC that is not able to serve all the pci-e links at the same time.

Apple's memory controller is literally over 5x faster than the best mobile AMD Zen4s, in the same soldered configuration.

Peripheral requirements have skyrocketed, and x86 memory controllers are growing at a rate of 10-15% in 2 years...

The complete failure of the x86 architecture is obvious to all IT experts. Intel is fussing and has now introduced an urgent APX extension for the x86 architecture, but it's a dead poultice - x86 memory controllers slower than it should be at times for consumers are leading to a gradual global defeat of the x86 architecture.

Nothing is eternal. x86 is dying and that's a fact.

The only way to revive the x86 corpse is a radical transition to 512-bit HBM with a throughput of 400GByte/s.

think my 4800h will serve me well until zen 5 comes, that should be a real upgrade finally.

QuoteYet the difference in battery life is massive, M2 macbooks are pulling 16-17 hours easily while zen 4 laptops are barely pulling 7-8 hrs

Are there any AMD Phoenix APU-only [without dGPU] notebook reviews? Preferable one battery runtime test with APU limited to 15 Watt [7840U], then to 20 Watt and then to 25-28 Watt (U = 15W to 28W).
There was a recent news how most notebook companies sold 15% less, but APPLE sold 10% more. Maybe the long battery runtime is part of it.

The one thing I don't get is, Zen 4 both on server and desktop has shown huge efficiency gains but on mobile there seems barely any, in fact we've seen regressions instead? I get that you can't expect much from handhelds due to smaller batteries and this gen has more power hungry displays (120/144 -> 165 hz, miniLEDs) but even accounting for all that still seems a bit disappointing.

The 7840U, depending on configured TDP (15w - 30w) can perform somewhere between a M2 and M2 Pro depending on benchmark. According to NBC's own tests, AMDs Zen 4 mobile APUs are only slightly Apple's M2 SoC's in terms of efficiency. Yet the difference in battery life is massive, M2 macbooks are pulling 16-17 hours easily while zen 4 laptops are barely pulling 7-8 hrs.

Is windows just not optimized for zen 4 mobile APU's? Or are rdna3 drivers just that buggy?

Maybe the problem lies with buggy Nvidia Optimus implementations not properly powering down and never truly shutting off completely?

Don't understand where this "slow" memory controller is coming from. All AMD does is limit the channels to 2. And this is done by design (Don't think Sony would be very happy with AMD if they built an PC APU more powerful than the ones they supply to consoles). It's not like the competition (x86 or ARM) is any better, besides Apple.

If AMD couldn't supply fast memory controllers in their SoC's, do you honestly believe they'd of been able to take as much of the data center / server market as they've now? Or win console contracts?

Also, I don't agree with your wheels on Ferrari analogy at all. While AMDs PC APU's are still bandwidth limited, they're not that powerful to begin with (only 12 CUs or less).

Remember this, the goal of AMDs PC iGPU's is to be slightly faster than intel's in actual games. With the exception of Sarlak / Strix Halo (which depending on pricing / availability - could be another DOA), this isn't gonna change next year either.

Due to the extremely slow AMD memory controller in all Zen, even against the backdrop of Intel, AMD must either increase the L3 cache to more than 256MB or dramatically speed up the controller or add 2 more channels or switch entirely to HBM memory. Without it, its integrated graphics cards would needlessly choke on the low bandwidth of RAM. As well as many cores.

It's like putting DDR Trabant wheels on a Ferrari. What is the use of power if there is nothing to drive it?