News AMD boasts its Ryzen AI Max+ 395 is up to 12.2x faster than Lunar Lake in AI workloads

[Admin]

AMD benchmarked its Ryzen AI Max+ 395 flagship against Intel's Core Ultra 7 258V in a plethora of AI benchmarks. The AMD chip was up to 12x faster than the Intel counterpart.

AMD boasts its Ryzen AI Max+ 395 is up to 12.2x faster than Lunar Lake in AI workloads : Read more

 

[igor002020]

That comparison is rather misleading. LM Studio doesn't support Arc GPUs natively and uses a Vulkan backend instead. Nor does it use NPU. This pretty much tanks AI performance on Intel Arc.

 

[HideOut]

That comparison is rather misleading. LM Studio doesn't support Arc GPUs natively and uses a Vulkan backend instead. Nor does it use NPU. This pretty much tanks AI performance on Intel Arc.

Theres lots wrong with this comparison
For once, AMD's MOST powerful ai trash CPU vs a mid range intel one? that makes sense.

 

[Cooe]

"significantly higher configurable TDP (rated up to 120W)"

🤦... The ROG Flow Z13 absolutely does NOT run at 120W! Due to its 2-in-1 tablet form factor its max TDP is only 70W, meaning there's still a TOOOOON of unused performance headroom we haven't yet seen with Strix Halo! The max TDP for the chip is literally almost DOUBLE the Z13's!

A 120W Strix Halo in a tiny desktop ala Framework or in a well-cooled gaming laptop is gonna absolutely DOMINATE!!! Think R9 9950X CPU performance + clear cut RTX 5070 Mobile beating GPU performance (aka potentially desktop RTX 4070 beating) with GOBS of unified memory!

 

[igor002020]

Theres lots wrong with this comparison
For once, AMD's MOST powerful ai trash CPU vs a mid range intel one? that makes sense.

Agree

 

[Giroro]

What is the significance and scale of "time to first token"?
Is it something like a multi-second workload takes 1us to get started instead of 12uS? As in a workload that takes 2,000,012 uS would now take 2,000,001 uS? Because that kind of "12x" improvement would be insignificant, uninteresting, and within the margin of error.
That's what it sounds like, or does it mean something else? It definitely doesn't mean the entire job gets done 12x faster, because AMD is only claiming ~2.1x performance in the overall throughput.
But they got their "12x more betterer" Headlines I guess, which is all they wanted.

Plus its still annoying to see companies' marketing department say things like "12 times faster" to mean "the task gets done in one twelfth the time", because that's a pretty backwards way of using multiplication. Granted, people working in marketing probably flunked math.

 

[igor002020]

DOMINATE

That thing makes little sense in desktop IMO. It will be crashed by any middle end cpu+egpu combo for 2x less price (unless you really need to use super large models that doesn't fit 16GB).

 

[Cooe]

Theres lots wrong with this comparison
For once, AMD's MOST powerful ai trash CPU vs a mid range intel one? that makes sense.

🤦 Strix Halo is literally the coolest, most exciting & innovative x86 chip we have had since probably Matisse (Zen 2 desktop) way back in 2019... Calling it "trash" is absolutely freaking absurd and tells me that you are tech illiterate. 🤷

 

[Cooe]

That thing makes little sense in desktop IMO. It will be crashed by any middle end cpu+egpu combo for 2x less price (unless you really need to use super large models that doesn't fit 16GB).

Show me a system as small as the Framework desktop (4.5L) with more power.... I can wait. 🤷

 

[igor002020]

Show me a system as small as the Framework desktop (4.5L) with more power.... I can wait. 🤷

I can show you 5L system, plenty of 6-7L systems ... Is there any sacramental meaning for 4.5L number?

 

[rluker5]

Lunar lake was meant for Ultrabooks and business laptops with great battery life.
How is the 395 for battery life running office/streaming video? Because LLs target market couldn't care less about the AI workloads tested. They probably care more about fan noise levels.

As long as we are comparing apples to oranges by comparing a 17w tablet chip to a 120w gaming console chip, I heard a rumor that the CPU in the moto g play obliterates the 395 in valued phone metrics when used in a phone. Such a shame that AMD is such a failure in that giant market.

 

[anoldnewb]

What is the significance and scale of "time to first token"?
Is it something like a multi-second workload takes 1us to get started instead of 12uS? As in a workload that takes 2,000,012 uS would now take 2,000,001 uS? Because that kind of "12x" improvement would be insignificant, uninteresting, and within the margin of error.
That's what it sounds like, or does it mean something else? It definitely doesn't mean the entire job gets done 12x faster, because AMD is only claiming ~2.1x performance in the overall throughput.
But they got their "12x more betterer" Headlines I guess, which is all they wanted.

Plus its still annoying to see companies' marketing department say things like "12 times faster" to mean "the task gets done in one twelfth the time", because that's a pretty backwards way of using multiplication. Granted, people working in marketing probably flunked math.

the AMD blog post that is the basis for this article https://community.amd.com/t5/ai/amd...or-breakthrough-ai-performance-in/ba-p/752960

has a link to an AI analysis of a MRI image on YouTube:
https://community.amd.com/t5/ai/amd...or-breakthrough-ai-performance-in/ba-p/752960

Look near the bottom of the screen at the end of the video. 5.7 seconds is the time to first token (when the answer begins to appear on screen) and that is followed by 10 tokens per second until the answer (361 token) is complete. I'm not sure where you're getting the 1-12 microseconds till first token/

 

[alceryes]

I would love to see this chip maxed out on a SFF feature-packed MB.
(not a VRM-starved Ali special)

 

[RadicalDann]

The comparison supplied by AMD is like trying to race a sports car against a minivan. Its a really misleading way to present the comparative specs.

Laptop Mag did an article testing it against a chip in the same class and its a much truer representation: https://www.laptopmag.com/ai/copilot-pcs/amd-ryzen-ai-max-395-vs-apple-m4-pro-benchmarks

 

[Notton]

Seeing the 395+ run at 120W, I am somewhat interested in what happens when you fully unleash a lunar lake from thermal and power constraints.
I don't expect LNL to run at 120W, but surely it must push 60~90W?

 

[redgarl]

"Ryzen chip consumes more than 3x the power, but fits in the same thin-and-light form factor as Lunar Lake competition."

You are forgetting to say that its is also having a GPU more than 3 times bigger, and it is hosting a 16 cores CPU. The power draw is not even a matter here, but Intel legitimacy as a leader in mobile is.

This CPU is probably the most disruptive being released since Zen 2 or the 5800x3D.

 

[redgarl]

The comparison supplied by AMD is like trying to race a sports car against a minivan. Its a really misleading way to present the comparative specs.

Laptop Mag did an article testing it against a chip in the same class and its a much truer representation: https://www.laptopmag.com/ai/copilot-pcs/amd-ryzen-ai-max-395-vs-apple-m4-pro-benchmarks

Yes and no... the problem is that the competition cost 4000$ and is not able to run on PC properly without buggy emulation.

Also, as much as Apple want to brag about their ARM CPUs, the intrinsic problem is that they are ARM CPUs.

 

[redgarl]

Agree

ROFL... it is their laptop flagship you genius...

The R7 is only a bit underclock vs the R9, which host the same GPU and the same core configuration.

The AI 395+ is having a GPU able to match the 4070 mobile and it is hosting 16 cores with onboard memory up to 128GB.

Like stated before, this chip host a GPU that is 3 times bigger and a CPU that is, at minimum, 2 times bigger. So the power figures are understandable in compute workloads.

 

[cyrusfox]

🤦 Strix Halo is literally the coolest, most exciting & innovative x86 chip we have had since probably Matisse (Zen 2 desktop) way back in 2019... Calling it "trash" is absolutely freaking absurd and tells me that you are tech illiterate. 🤷

He is calling the AI trash, and it truly is, I have the lunar lake, never even use the npu, all AI bells and whistle have yet to find a true use case. GPU or CPU tops are more than sufficient. Agree that this comparison is absurd.

 

[cyrusfox]

Seeing the 395+ run at 120W, I am somewhat interested in what happens when you fully unleash a lunar lake from thermal and power constraints.
I don't expect LNL to run at 120W, but surely it must push 60~90W?

lunar lake doesn't have legs if you give it juice, it was gimped to be an arm competitor, hence the tight integration with dram on chip. Compare the performance of 288V vs 268V vs 258V, 5% improvement going to 30W vs 17W. Its ats its golden efficiency range at 17W. First battlemage product though, and it is quite decent for Intel. It does everything I need in a laptop, I love it, just wish it was bit more mature.

But these two are different classes of products and really not a valid comparison. Go compare this monster AMD chip to Intels H or HX line (285HX)at least the TDPs would be more comparable then (55W vs 120W) as the thread/core count. Lunar lake with its 4P/4lE cores is more akin to phone/low power mobile then desktop class replacement line.

 

[alceryes]

The comparison supplied by AMD is like trying to race a sports car against a minivan. Its a really misleading way to present the comparative specs.

Laptop Mag did an article testing it against a chip in the same class and its a much truer representation: https://www.laptopmag.com/ai/copilot-pcs/amd-ryzen-ai-max-395-vs-apple-m4-pro-benchmarks

I think we're to the point where wattage during testing needs to be included in all SFF/laptop/ultra-portable test results. Unless you use similar wattage chips the results could really be misleading.