AMD Details the Radeon 890M RDNA 3.5 iGPU of "Strix Point" a bit More

[btarunr]

AMD presented a closer look at the Radeon 890M iGPU powering the Ryzen AI 300 series "Strix Point" mobile processor. The iGPU introduces the new RDNA 3.5 graphics architecture, with several architecture-level improvements built around the existing RDNA 3 SIMD, to yield performance/Watt improvements that AMD could trade in to increase the SIMD muscle for its processors, and proportionately increase performance. The iGPU features one Shader Engine with 8 workgroup processors (WGPs), which amount to 16 CU (compute units), for a total of 1,024 stream processors, 32 AI accelerators, and 16 Ray accelerators. The iGPU also has 4 render backends+, for 16 ROPs. It is specced with a maximum engine clock of 2.90 GHz, which yields over 11 TFLOP/s of FP32 throughput, which is around 30% higher than the iGPU of "Phoenix" (12 CU, RDNA 3), at comparable power.

AMD goes into the finer points of how it yielded the performance/Watt gains. The company worked on the texture subsystem to double the texture sampler rate, and introduced point-sampling acceleration. The shader sub-system features interpolation and comparison rate doubling. The raster sub-system introduces sub-batching of batch raster operations, with a programmable bin order, for the hardware to be more efficient. Lastly, AMD worked on the iGPU's memory-management, to be more aware of LPDDR5 (which has a different physical layer or way of writing/fetching than GDDR6). The company worked on improving the memory compression technologies, to improve performance, and reduce the iGPU's memory footprint.



View at TechPowerUp Main Site

 

[john_]

I hope they fixed the power consumption problem with RDNA3 when playing multimedia files. And they should focus there because when tech sites test battery life, one of the metrics is video playback time on battery.

 

[Apocalypsee]

If this available on desktop it would be the thing I buy next

 

[Daven]

I wonder if the Hawk Point v. Strix Point comparison have the same number of CUs. If its 6 v. 8 then it's less interesting.

 

[persondb]

I wonder if the Hawk Point v. Strix Point comparison have the same number of CUs. If its 6 v. 8 then it's less interesting.

I would think it is 12 CUs(i.e. 6 WGP) vs 16 CUs(8 WGPs) otherwise they would have noted it down.

A considerably part of the improvement will just be from the increased sized.

 

[remekra]

Yes, but if it's at the same power then it's an improvement. But yeah there is a reason it's called 3.5 and not 4, just minor improvements.

 

[A&P211]

It should be available on desktop APUs. The rumored performance is around a desktop 2050.

If this available on desktop it would be the thing I buy next

 

[Minus Infinity]

33% more CU's, higher clocks and only 30% stronger performance. Colour me underwhelmed. Isn't this technically regression?

 

[watzupken]

33% more CU's, higher clocks and only 30% stronger performance. Colour me underwhelmed. Isn't this technically regression?

The fact they called it RDNA 3.5 suggests that this is probably just a refined version of RDNA 3. Since performance don't scale linearly with increased cores and clockspeed, therefore, I feel a 30% increase in performance is reasonable.

 

[Sound_Card]

33% more CU's, higher clocks and only 30% stronger performance. Colour me underwhelmed. Isn't this technically regression?

That's significant.

 

[Minus Infinity]

That's significant.

No it's not. What's the point of RDNA3.5. Youi telling a 16CU RDNA3 iGPU with higher clocks isn't going to be nearly 30% stronger. Maybe they lowered power usage 30%.

What we need is the 12CU RDNA3.5 vs 12CU RDNA3 igpu test at same clocks. It doesn't seem IPC has improved much at all.

 

[Apocalypsee]

No it's not. What's the point of RDNA3.5. Youi telling a 16CU RDNA3 iGPU with higher clocks isn't going to be nearly 30% stronger. Maybe they lowered power usage 30%.

What we need is the 12CU RDNA3.5 vs 12CU RDNA3 igpu test at same clocks. It doesn't seem IPC has improved much at all.

The .5 is there to signify it's not a significant leap over previous chip. How much gain do you expect to have for 33% more CU? Mind you this is an APU with limited memory bandwidth even with LPDDR5.