AMD’s latest anti-Nvidia ‘AI’ focused PR blitz is falling flat once again, they just don’t learn. At this point, SemiAccurate is wondering why the person pulling the strings just can’t be force fed a clue.
Over the last few days, AMD put out five ‘AI’ focused PR pieces, all of which dug their hole deeper. Since we know the front line PR people and unless they all suffered closed head wounds at the same time, there is someone above who doesn’t like AMD having a positive public image. And is acting on it. In case we are wrong, please heal up people, rest is best for such head wounds.
Back to the story at hand, the five pieces were a pathetic ‘analyst’ focused AI PC piece, two on the Ryzen 395 with a mewling ‘AI’ moniker, an Epyc puff piece that avoids comparisons to actual competition, and a Versal piece that, err, doesn’t suck. Until they went down the AI path for no good reason. All of this stupid would be ignorable if it wasn’t done in lieu of real information.
Lets start out with the most noxious, AMD’s paid for ‘analyst’ piece about why you should buy an AI PC. It is wrong on just about every point and embarrassing for both the authors and AMD, so lets dig into why it exists. AI PCs are turds, everyone knows it. Microsoft is DESPERATE to push them because it pawns off their ballooning data center costs. At the time there were no real apps to justify an ‘AI PC/Copilot’ device and nearly a year later that app count has multiplied by 10. For the math averse you end up at the same number, zero.
AMD The analysts go on to define an AI PC as one with an NPU and lesser specs than Microsoft’s 40+ TOPS number because AMD has a large number of SKUs that don’t meet that bar. All this said the Microsoft spec is BS because the things MS claims need and NPU run perfectly without one, Microsoft just blocks them. So at the moment there is no point to an NPU at all, other than marketing.
Then they go on to define an NPU in a way that, well, they just don’t get it. “The NPU is built specifically to run AI workloads as efficiently as possible. This frees up the CPU and GPU to run other system tasks, while the NPU focuses on advancing AI workloads with minimal impact on battery life. Given the NPU efficiencies, the operating system and AI apps can run constantly, providing persistent and pervasive capabilities such as real-time translations or background blur during meetings.” Let us point out that EVERY PC CPU from Intel, AMD, and Qualcomm for the past many years has had a hardware block that meets this definition. Remember the low power DSP/AI island for doing, wait for it, AI tasks such as voice recognition in hardware? Come on AMD, pick your idiotic talking point for the paid analysts to parrot better. Also note that the ‘AI’ tasks listed can, are, and have been done for years without an NPU, albeit at a negligibly higher power use.
Then comes the real point to all of this, selling. The message is the same tired but dead wrong one that sounds good unless you think about it just a little. They say with AI apps coming, there is no reason to delay buying one of these new ‘AI’ PCs. Sure there are no apps now, but don’t you want to be ready when they do come? The problem is this logic doesn’t parse for several reasons.
First up is that all the use cases run just fine on current PCs with a slightly higher power draw. SemiAccurate says it is negligible because every time we ask AMD, or Intel and Qualcomm, to quantify the added draw, they refuse. Any guesses why? On top of that there are literally no real new killer apps for the NPUs yet, everything falls under the ‘use once’ or tech demo categories. How many times do you need to go over all your stored images to tag again? Once? The way AMD and others are pimping it, people need to do it several times a day or civilization will collapse.
Next is their promise that AI apps will come, soon is intoned but knowingly unrealistic, and then you will need the TOPS an NPU can deliver. This is wrong because, wait for it, hardware gets faster with each generation. If TOPS do matter, holding off until the apps arrive will get you the best device for this yet not-on-the-roadmap killer app. The key argument for buying now is actually just the opposite, if you wait you will get a device that is much better suited to task. If there is ever a task. Should you need to buy now, we feel we need to point out that Intel’s Arrow Lake-H delivers ~2x the platform TOPS of AMD’s best NPU and Lunar Lake is close to 3x.
So in the end this ‘analyst’ piece is paid embarrassment. Their main argument is actually an argument to avoid upgrading until a killer app emerges, not one to buy now. There is no killer app here, or even a marginally useful one, nor is there one impending, just ask Microsoft’s Satya Nadella. That said, Nvidia has an AI conference going on as we speak so time to ride their coattails.
Moving on to the next bit, AMD is trying to point out that their ‘AI’ branding of the new Ryzen 395 is more than a marketing gimmick. It actually is, if you buy one, keep the box, it ages better than a bottle of syrup of ipecac in the medicine cabinet, and works faster too. Seriously though, AMD’s two presentations on the topic are well beneath them for disingenuous messaging.
This graph is NOT about AI performance
If you look at the pic above, it purportedly measures AI performance. If you understand the tech, you quickly realize it does nothing of the sort, it measures memory performance under a very skewed set of circumstances. It measures performance vs an Nvdia discrete desktop card, and points out the power use of it too. The problem is that the Nvidia card beats AMD silly until the model size exceeds the 16GB memory limit of the GPU and has to swap. If you exceeded the AMD’s memory size, the Nvidia card would still crush AMD, but they will never point this out. It is borderline dishonest.
Pick a single test you win on….
The one above looks a little more like a real test but it only measures the time to first token. We couldn’t dig into this one because AMD dumped these presentation on a Friday afternoon with an embargo lift of Monday, so no time to get numbers and statements from competitors. Think there is a reason for this? Given that the listed Intel system has the same NPU TOPS, vastly more full system TOPS, and software that doesn’t suck, heck it actually works in the real world, you have to wonder. This is cherry picking for data skewing in the extreme. Come on AMD, you don’t need the footgun this badly.
So for their two ‘AI PC’ Ryzen 395 puff pieces, AMD can’t seem to actually win on head to head performances vs Intel or Nvidia. If you pick your benchmarks very carefully, especially ones that don’t represent what the average reader thinks they do, you can win. AMD took the wrong path.
Then we come to the server side, an obvious attempt to deflect from Nvidia’s PR messaging at their GTC conference. AMD claims that their Epyc CPUs beat the pants off Nvidia’s Grace in energy efficiency for AI and business apps. No one, let us repeat, NO ONE is dumb enough to buy Grace for business apps because it isn’t meant for that purpose. AMD knows this but still lowers the messaging bar. It may be more energy efficient but, well, speed matters too, and that topic is avoided.
More importantly, AMD beats the pants off Intel’s best in generalized server applications but doesn’t bring this up either. The one place where Intel shines in server, and has actual wins on merit, is AI. Why? They have a dedicated on-die AI coprocessor/matrix unit. If Intel was included on AI, guess who would win here? Other than that the presentation is fairly innocuous, AMD does really win at most every server task now.
Last up was a really cool one until AMD’s palpable desperation dragged AI into it. AMD announced the Xilinx Versal XQRVE2302 has now achieved Class B qualification and is approved for spaceflight. If you care you can read up on the MIL-PRF-38535 spec here, it is a gripping read for those into rad-hardening minutia.
The Versal in question is their AI version but AMD is pimping that over the rest of the chip. It would have been interesting if AMD had actually briefed on this rather than intentionally dumping it on a Friday afternoon to preclude real disclosure for another own goal.
All in all over the past few days, AMD continues their messaging downfall. In a desperate attempt to capitalize on Nvidia’s GTC PR blitz, they put out a lot of crap in one of the worst ways possible. The messages ranged from wrong to disingenuous with the Versal bits only being slightly noxious. AMD is better than this, whoever is making the PR team do this crap, stop it, you are losing the plot in spite of your (at times) hardware advantages.S|A
