AMD Strix Halo Radeon 8050S and 8060S iGPU Performance Look Promising - And Confusing

[GGforever]

AMD fans are undoubtedly on their toes to witness the performance improvements that Strix Halo is ready to bring forth. Unlike Strix Point, which utilizes a combination of Zen 5c and full-fat Zen 5 cores, Strix Halo will do away with the small cores for a Zen 5 "only" setup, allowing for substantially better multicore performance. Moreover, it is also widely expected that Strix Halo will boast chunky iGPUs that will bring the heat to entry-level and even some mid-range mobile GPUs, allowing Strix Halo systems to not require discrete graphics at all, with a prime example being the upcoming ROG Flow Z13 tablet.

As per recent reports, the upcoming Ryzen AI Max+ Pro 395 APU will sport an RDNA 3.5-based iGPU with a whopping 40 CUs, and will likely be branded as the Radeon 8060S. In a leaked Geekbench Vulkan benchmark, the Radeon 8060S managed to outpace the RTX 4060 Laptop dGPU in performance. However, according to yet another leaked benchmark, Passmark, the Radeon 8060S and the 32-CU 8050S scored 16,454 and 16,663 respectively - and no, that is not a typo. The 8060S with 40 CUs is marginally slower than the 8050S with 32 CUs, clearly indicating that the numbers are far from final. That said, performance in this range puts the Strix Halo APUs well below the RTX 4070 laptop GPU, and roughly the same as the RTX 3080 Laptop. Not bad for an iGPU, although it is almost certain that actual performance of the retail units will be higher, judging by the abnormally small delta between the 8050S and the 8060S.



View at TechPowerUp Main Site | Source

 

[TheinsanegamerN]

I cant wait to get one of these in the z13 flow. Finally, a powerful windows laptop with USB A ports.

Not bad for an iGPU, although it is almost certain that actual performance of the retail units will be higher, judging by the abnormally small delta between the 8050S and the 8060S.

Hmmm....there's a lot that could be happening there. The 40 CU chip could be running into TDP issues, or heat problems, or bandwidth limitations, which is the most likely. The current 23CU radeon 780m gets hit with bandwidth limitations on DDR5 systems, where DDR5X 7XXX mhz systems are significantly faster, something like 20-30% over DDR5 5200 setups. the big strix halo is more then 3x bigger at 40CU, so yeah its got double the bus width and faster DDR5X 8xxx memory, but triple the GPU demand.

Man if there was ever a chip begging for x3D cache.....

 

[Neo_Morpheus]

If these leaks keep going, the keynote will be irrelevant.

I wonder if this is what ValvE will use on the upcoming Steam Machines II.

 

[A&P211]

I cant wait to get one of these in the z13 flow. Finally, a powerful windows laptop with USB A ports.


Hmmm....there's a lot that could be happening there. The 40 CU chip could be running into TDP issues, or heat problems, or bandwidth limitations, which is the most likely. The current 23CU radeon 780m gets hit with bandwidth limitations on DDR5 systems, where DDR5X 7XXX mhz systems are significantly faster, something like 20-30% over DDR5 5200 setups. the big strix halo is more then 3x bigger at 40CU, so yeah its got double the bus width and faster DDR5X 8xxx memory, but triple the GPU demand.

Man if there was ever a chip begging for x3D cache.....

The 780m has 12CU, not 23CU.

 

[Dr. Dro]

They finally done lost it... marketing and SKU segmentation spam have reached levels not even Intel dared dream

The 780m has 12CU, not 23CU.

Seems to be a typo, 23 is next to 12

 

[Asni]

However, according to yet another leaked benchmark, Passmark, the Radeon 8060S and the 32-CU 8050S scored 16,454 and 16,663 respectively - and no, that is not a typo. The 8060S with 40 CUs is marginally slower than the 8050S with 32 CUs, clearly indicating that the numbers are far from final.

Honestly it's pretty clear to me: the cTDP goes from 55 to 130w, the difference between 8050s and 8060s depends on the power target and how linear the voltage/frequency curve is.

 

[dj-electric]

So.... what are the chances of us getting an 8C16T 20CU SoC for a significantly cheaper price, so we could create a very interesting category of thin and lights.

 

[Wirko]

They finally done lost it... marketing and SKU segmentation spam have reached levels not even Intel dared dream

You are supposed to stop reading before you get to the model number.

 

[Darmok N Jalad]

Ryzen AI Max+ Pro 395

Gosh AMD, you forgot the “XT,” or is that for later?

 

[Dr. Dro]

Ryzen AI Max+ Pro 395

Gosh AMD, you forgot the “XT,” or is that for later?

The GRE, XT, XTX and XTXH, Ultra and Platinum versions will come later

 

[tabascosauz]

Need a iGPU only Flow X13 with these. Will ditch my G14 for it

 

[lilunxm12]

Why do they have a score for titan v ceo edition? I think only a few billionaire ceo should have access to this card?

 

[whaTHEll]

Man if there was ever a chip begging for x3D cache.....

Actually, it does. rx80x0s have 32mb infinity cache, same as rx7600 do.

 

[Punkenjoy]

A great benefits of those GPU would be VRAM. Since they use the main memory, you would have the ability to buy a laptop with a lot of RAM and have plenty for the iGPU.

Of course, you wouldn't be able to run the latest game at ultra for long at 1440p on those, (not sure even at release) but maximising texture quality is still today a good way to optimize visual without impacting too much performance.

Unlike by example a stronger dGPU with 8 GB of ram. You could run more effects but you would have to do a lot of sacrifice on texture and that would have a big impact on visuals.

I still find frustrating that we no longer double vram every gen. (i know, it had to stop at some point but still). The last few games i played were a mix of superb detailed texture on character and main surface while other had very poor texture. This is a way for the game dev to "optimize" but it is becoming distracting.

 

[TheinsanegamerN]

Actually, it does. rx80x0s have 32mb infinity cache, same as rx7600 do.

AFAIK, the infinity cache of this iGPU is NOT an x3d cache, it is the GPU L3 cache with a fancy name. It's not connected to the GPU in any way, and 32MB is too small to make a significant difference.

The x3d cache on the desktop chips could be used by both the CPU and iGPU, and even with only 2 CUs showed major improvements.

View attachment 377048

They finally done lost it... marketing and SKU segmentation spam have reached levels not even Intel dared dream



Seems to be a typo, 23 is next to 12

Yes that was a typo.

So.... what are the chances of us getting an 8C16T 20CU SoC for a significantly cheaper price, so we could create a very interesting category of thin and lights.

Highly unlikely. That would be half the iGPU cut off. If defects were common enough to justify mass production of a chip with half of it disabled, the yields would be bad enough that the big 40 CU chip wouldnt exist int he first place.

Closest you will get is the top Strix Point, which is 8c/16t with 16CUs.

 

[Vya Domus]

On a different note, I can't believe a mobile 3080 is as slow as a 6600XT lol.

 

[Dr. Dro]

On a different note, I can't believe a mobile 3080 is as slow as a 6600XT lol.

It is an RTX 3070 Ti with slower GDDR6 and heavily reduced clock speeds, so not that surprising

 

[TheinsanegamerN]

On a different note, I can't believe a mobile 3080 is as slow as a 6600XT lol.

Imagine paying over $2k for a laptop that gets beaten by a $800 desktop.

 

[ymdhis]

I want this in a desktop.
Or at least a mini-PC that can either take standard AM5 coolers, or has a full sized 120x25 fan on it, not those stupid little laptop blowers.

 

[JustBenching]

Imagine paying over $2k for a laptop that gets beaten by a $800 desktop.

A 4k$ laptop gets beat by a 1200$ desktop. It's just physics.

 

[mikesg]

I want this in a desktop.
Or at least a mini-PC that can either take standard AM5 coolers, or has a full sized 120x25 fan on it, not those stupid little laptop blowers.

Laptop coolers are a little more advanced now. High-end have Vapour Chambers, phase-change thermal paste, fins the length of the back, twin fan, alu/magnesium base.

 

[ymdhis]

Laptop coolers are a little more advanced now. High-end have Vapour Chambers, phase-change thermal paste, fins the length of the back, twin fan, alu/magnesium base.

And they still have only a small tiny thin fan that goes brrrrrr. That's the deal breaker.

 

[trsttte]

Laptop coolers are a little more advanced now

But mini pc's aren't

 

[lhl]

The 2023 Flow Z13 (i9-13900H/RTX 4060 65W) had a dual-fan chamber cooling module that per Notebookcheck's testing could handle up to 90W sustained in gaming/max workloads, so I suspect that if the same cooling system is used, it might actually hit higher TDPs than people might expect from a thin and light.

My understanding is the 32MB of Infinity Cache is MaLL for the GPU (basically required by RDNA3 to get effective performance for 3D rendering - not sure it's so useful for AI inference though.

BTW, this is probably where Strix Halo is most interesting. It supports up to 75% of its unified memory as VRAM, so up to 48GB out of 64GB. That's enough to (slowly) inference a 70B Q4 (slightly below GPT4 class) model locally. You can also run SOTA coding models like Qwen2.5-Coder w/ plenty of context. There's a lot of interest from the LocalLLama crowd on the possibilities of Strix Halo.