AMD Strix Halo lineup led by Ryzen AI Max+ 395 zealously aims to overtake Apple M4 Pro, Core Ultra 9 288V; 2x faster than RTX 4090 in AI with 87% less power

[Redaktion]

AMD has unveiled its most powerful Copilot+ PC APU lineup. Dubbed the Ryzen AI Max series, the Strix Halo series led by the Ryzen AI Max+ 395 offers up to 16C/32T configurations with a 50 TOPS NPU and RDNA 3.5 GPU with up to 40 CUs. AMD is aiming to take on the M4 Pro Macs, the Core Ultra 9 288V, and even the RTX 4090 with these new APUs.

https://www.notebookcheck.net/AMD-Strix-Halo-lineup-led-by-Ryzen-AI-Max-395-zealously-aims-to-overtake-Apple-M4-Pro-Core-Ultra-9-288V-2x-faster-than-RTX-4090-in-AI-with-87-less-power.941958.0.html

[akaprog]

Sounds snake oil to me, for reference GTX 4090 could offer up to 1200 TOPS.

[Mr Majestyk]

Sounds snake oil to me, for reference GTX 4090 could offer up to 1200 TOPS.

I'm sure at 100W Halo is better than a gimped 4090 at 100W, but the way the headline is phrased is clickbait.

[ULTRA AI PRO MAX+ XTX]

Sounds snake oil to me, for reference GTX 4090 could offer up to 1200 TOPS.

I'm sure at 100W Halo is better than a gimped 4090 at 100W, but the way the headline is phrased is clickbait.

Their slide states that they were comparing their 55W TDP 395 APU to a 450W 4090. Regardless, the data is virtually meaningless since the model is larger than 70 GB; there is no way all of it is fitting into the 4090's 24 GB VRAM buffer without extreme IQ2 quantitation, which makes the model worthless. In other words, the majority of the processing was performed by whatever CPU they paired the 4090 with.

[Redflag]

I wonder which person at AMDs marketing department thought it was a bright idea to compare a lunar lake (a SoC literally designed for 15w ideally) with strix halo which is like an 80w part and not to mention just a much bigger chip in general.

Like no sheet an airplane moves faster than a bicycle.

Rather they compared z2 extreme with lunar lake, as we know these are broadly equivalent GPU wise, would be interesting to see how they compare in power efficiency (when you've some zen 5 cores disabled, how much it actually helps over standard strix point part)

Also, why are they comparing it to M4 pro? If they're gonna copy apples naming scheme shouldn't they compare it to the M4 max? Seems a bit of a red flag as it might possibly indicate that it doesn't compete with that in M4 max at all in anyway.

Unless they're basically saying that the reason it was compared to that is because it won't be priced any higher than M4 pro hardware. Hope it's the case because apple max hardware is ridiculously priced.

[dada_dave]

I wonder which person at AMDs marketing department thought it was a bright idea to compare a lunar lake (a SoC literally designed for 15w ideally) with strix halo which is like an 80w part and not to mention just a much bigger chip in general.

Like no sheet an airplane moves faster than a bicycle.

Rather they compared z2 extreme with lunar lake, as we know these are broadly equivalent GPU wise, would be interesting to see how they compare in power efficiency (when you've some zen 5 cores disabled, how much it actually helps over standard strix point part)

Also, why are they comparing it to M4 pro? If they're gonna copy apples naming scheme shouldn't they compare it to the M4 max? Seems a bit of a red flag as it might possibly indicate that it doesn't compete with that in M4 max at all in anyway.

Unless they're basically saying that the reason it was compared to that is because it won't be priced any higher than M4 pro hardware. Hope it's the case because apple max hardware is ridiculously priced.

Agreed but in fairness it isn't like the Arrow Lake processors were available yet so they would've had to compare to older Intel hardware. Probably would've been better. Apple had the same issue when they released the M4 Max. They also compared themselves to Lunar Lake in their presentation, which was similarly nuts.

In addition to not comparing against the M4 Max, the comparisons with the M4 Pro are also CPU only and there the Halo doesn't fare so well in Cinebench 2024. (Also not clear if Corona/V-Ray are even AS native)

Sounds snake oil to me, for reference GTX 4090 could offer up to 1200 TOPS.

I'm sure at 100W Halo is better than a gimped 4090 at 100W, but the way the headline is phrased is clickbait.

Their slide states that they were comparing their 55W TDP 395 APU to a 450W 4090. Regardless, the data is virtually meaningless since the model is larger than 70 GB; there is no way all of it is fitting into the 4090's 24 GB VRAM buffer without extreme IQ2 quantitation, which makes the model worthless. In other words, the majority of the processing was performed by whatever CPU they paired the 4090 with.

If processor A can fit the model in memory and B can't and therefore A does better, that's a fair point to make.

Also in inference bandwidth matters as much as TOPs, but even then the 4090 will have more bandwidth so yes this is a case of the 4090 running out of memory.

[dada_dave]

(Also not clear if Corona/V-Ray are even AS native)

They are, not sure how well optimized they are.