Clasically, OEMs like to use whatever crap they have in the warehouse, especially when it comes to SODIMMs etc. They always seem to have something they ordered too much of stashed in the back that they're just dying to throw into their next build for a small markup.It is a somewhat harder sell to OEMs due to lack of supply flexibility.
RDNA3.5 is alright, maybe not so much vs Blackwell but Strix Halo SoC die is N3E, that is a superior node and will help a lot, especially if they take proper advantage of FINFLEX.The advantage that a big APU can be more power efficiency doesn't work in the real world, because Nvidia has a much superior architecture with significantly better perf/watt.
There's no way BoM is lower.APU BoM is lower due to unified memory.
32GB of LPDDR5X is cheaper than 24GB LPDDR5X + 8GB of GDDR6.
It is a somewhat harder sell to OEMs due to lack of supply flexibility.
Silicon cost is a wash, NV charges a lot for their mobile GPUs, good room to undercut with a big APU.
Integrating two complex chips aren't like copy and paste on Paint. What you are saying is in theory, but reality often turns out to be much different.RDNA3.5 is alright, maybe not so much vs Blackwell but Strix Halo SoC die is N3E, that is a superior node and will help a lot, especially if they take proper advantage of FINFLEX.
What would a "reasonable price" be? Especially in desktops it makes no sense because you can't resell it.I don't know where Strix Halo could lead, but I can promise you that if they can somehow get 90-100% of a 7900XTX (for example, and I'm probably stretching things a bit here) performance out of a single APU that they can sell soldered onto a board with a lot of fast memory for a reasonable price,
HSA is not the reason why they did the shared floating point. It was quite literally an area-thing.Wasn't HSA part of the reason why they did the shared floating point with Bulldozer? From what I remember they hoped that most floating point operations would be moved to the GPU. Interesting to see that turned on it's head. AMD's CPUs have the biggest baddest floating point performance, and their support for GPU compute (in client GPUs at least) is bad.
Cooling a single package is easier than two, especially when it could be 120W vs ~150W combined. Just take a traditional SoC only cooling solution and scale it to the required wattage.There's no way BoM is lower.
-Better thermal design because it's all one one package.
Starting to talk about R&D, that is not raw BoM but overall costs, I'm talking raw BoM, and most PCB's are pretty cheap.-No design re-use because you need a dedicated board. That means increased engineering costs. You have to spend time and money on engineers and resources.
Yes, and there you need MDF from the IHV for such laptop designs, that is standard practice.-Ultra high volume versus ultra low volume. Low volume = High proportion is spent on fixed costs, meaning lower margin.
2 sets of memory, GDDR being more expensive than LPDDR per GB. Memory costs more than offset die yield cost deficits.-Single large die versus two smaller ones.
Yes it does add a bit to fixed costs, but GDDR tracing is also expensive.-Dedicated specialized memory or multi-channel memory requiring complex layout and more copper layers versus a board and memory setup that is mass produced.
Without you know, the performance.A regular iGPU avoids all the above pitfalls.
It is a SoC die and 1-2 Z5 dies with an enhanced interconnect, on an InFO substrate.Integrating two complex chips aren't like copy and paste on Paint. What you are saying is in theory, but reality often turns out to be much different.
Mini PC guys might go for some, it is something I guess.What would a "reasonable price" be? Especially in desktops it makes no sense because you can't resell it.
At the expected volumes, Strix Halo will never make ROI, no matter the price, it is a loss leader and effective prototype for a lot of Zen6 parts.The price will reflect the ultra-low volume. Low margin is accepted, as long as the volume is super high. Strix Halo or any other Halo iGPUs have super, super low volume.
In the future, will take a while for it to be amortised.AMD isn't going Strix Halo for no reason. They are looking to profit.
Tech wasn't there, they were largely MDF queens in niche T&Ls.Intel with Iris parts were optional upgrades that added $150-200 on top of the most pricey i7 price.
So? At 120W TDP and ~440mm^2 total Si you cannot challenge high end mobile dGPU parts which are nearly 400mm^2 by themselves.So you have big APU thinking it's creme de la creme versus a dGPU competitor which is a mid range at best.
$2000 for 128GB, yeah nah.They may have an SKU that's fully specced that's pushing $2000 with 128GB RAM and 4070 performance,
Yeah, I'm hoping for $2k USD for 32GB full config.$2000 for 128GB, yeah nah.
It WILL be higher for the 128GB SKU
What would a "reasonable price" be? Especially in desktops it makes no sense because you can't resell it.
The price will reflect the ultra-low volume.
AMD isn't going Strix Halo for no reason. They are looking to profit.
Intel with Iris parts were optional upgrades that added $150-200 on top of the most pricey i7 price.
So you have big APU thinking it's creme de la creme versus a dGPU competitor which is a mid range at best.
I don't understand why anyone is talking about DESKTOP and profit in the same sentence in this thread. Desktop is dwarfed by laptops and other mobile devices right now. Narrowing down to desktops, a supermajority of those are black box OEM generic business machines with a midrange CPU with an iGPU that can barely paint a screen. The DIY section is barely a blip on anyone's radar, and precious few of those boxes ever get opened up and have a part resold.
Again, the bulk of the laptop market, where there is ANY focus of GPU performance, is in the $600-$800 range cheapest dGPU I can get in a laptop AND the $~1000 range 4060 with a high end CPU with decent screen range. You venture above that and unit sales plummet.
AMD aiming above that price point with Strix Halo would be suicide. They may have an SKU that's fully specced that's pushing $2000 with 128GB RAM and 4070 performance, but there is going to be a chunk of the market that's in the 8-12 core range that performs like a 4060 with 32GB RAM and sells for $1200 or so. The thing about Strix Halo is that it isn't anything outlandish for the OEMs. They've been selling laptops with LPDDRX modules embedded on the board for years. They've been selling boards with DGPUs with 4, 6, and 8 DDR chips embedded around them forever. This just simplifies and combined what they've been doing all along. They can still control cost by speccing cheaper dram chips or not populating channels or even limiting thermals and power.
And it'll largely be mobile or uSFF micro desktops where nothing is upgradeable already! It's a sound idea, so long as they can keep costs reasonable and drivers don't grenade the whole thing.
We still don't know if AMD has tackled the chiplet and I/O die idle power consumption issue with Strix Halo btw, because if they haven't improved enough on that, it will be a poor product IMO.
Well it's a mobile part.
They aren't.Heck, the laptops with Strix Point (HX 370 at least) are selling for more than that 1000$ range you mention, how the heck could they suddenly sell Strix Halo for the same or less than that?
The challenge is quite tricky here. Apple is seen as a vertically integrated provider where the claim that they choose only the best possible ingredients is believable. That approach is boosted by not even allowing any alternatives.Apple gets away with it, somehow.
HSA is not the reason why they did the shared floating point. It was quite literally an area-thing.
AMD went for the heterogeneous system architecture early on to exploit the fact that many compute intensive tasks can be offloaded to parts of the CPU that are designed to run them faster or at low power. By combining CPU and GPU on a single die, the system should be able to shift work around to complete the process quicker.
One such example that AMD are willing to share today is stock calculation using LibreOffice's Calc application – calculating the BETA (return) of 21 fake stocks and plotting 100 points on a graph of each stock. With HSA acceleration on, the system performed the task in 0.12 seconds, compared to 0.99 seconds when turned off.
The laptops with Strix Point that are selling for notably more than that all have dGPUs. Strix Halo parts won't.Heck, the laptops with Strix Point (HX 370 at least) are selling for more than that 1000$ range you mention, how the heck could they suddenly sell Strix Halo for the same or less than that?
It has dedicated boards because no current boards are designed for 256 bits of DRAM connecting to the CPU. This is a new layout.I highly doubt AMD will sell all of these chips at a loss. When you consider the space the NPU takes up in the design, the extra space from the extra CUs isn't much. OEMs want cheap and there is no way this costs more for OEMs to stick in their systems than a dGPU.
Why does it need a dedicated board? Doesn't this have a new FP11 socket?
The laptops with Strix Point that are selling for notably more than that all have dGPUs. Strix Halo parts won't.
Well it's AMDs fault then, I checked on Amazon and Newegg and there are zero laptops with HX 370 below 1500$. Most of them have dGPU but even ones with a measly 4050 were going for more than 1700$.
I mean, what's the point in creating an APU with the most CU (890M) but still worse than an entry level 4050 if all the laptops using them are going to be so expensive? makes zero sense to me.
AMD makes the SoC, it can't dictate what design the OEMs put it in... Seems that blaming the OEMs for stupid design choices would be a better option. Still, would have loved to see some SKUs without dGPUs like the Asus Zenbook S16/S14. (But with sane pricing)Well it's AMDs fault then, I checked on Amazon and Newegg and there are zero laptops with HX 370 below 1500$. Most of them have dGPU but even ones with a measly 4050 were going for more than 1700$.
I mean, what's the point in creating an APU with the most CU (890M) but still worse than an entry level 4050 if all the laptops using them are going to be so expensive? makes zero sense to me.
HSA is not part of the reasons for the shared FPU. The architectural choice for the shared FPU was done before Fusion/HSA was born.I didn't say it was the main reason, just part of the reason. And going back, AMD definitely advertised HSA as a way to boost performance in compute heavy tasks.
The challenge is quite tricky here. Apple is seen as a vertically integrated provider where the claim that they choose only the best possible ingredients is believable. That approach is boosted by not even allowing any alternatives.
Same is not true with most other manufacturers that are not vertically integrated and as such supply chain dictates much more of the final composition than the quality of individual components. So in such an environment specialized optimized parts like Strix Halo have a harder time getting into products.
Zen6 client will abandon CCDs + IODs if what we're being told is true. It'll be APUs for client desktop and mobile. Not sure what will happen to the Range product line though.That said Strix Halo may well see some X3D style hype on the back of its improved memory layout that unlocks more of Zen 5's performance. The question with that is how AMD would follow up such a product since it exploits a weakness of the base Zen 5 client configuration. Once that's fixed with a better cIOD in Zen 6 a straight Strix Halo successor will no longer be a standout.
AMD makes the SoC, it can't dictate what design the OEMs put it in... Seems that blaming the OEMs for stupid design choices would be a better option. Still, would have loved to see some SKUs without dGPUs like the Asus Zenbook S16/S14. (But with sane pricing)