You'd think after AMD shipped Xbox One, XboxX, PS4, and PS5 with decent iGPUs and a decent memory bandwidth they might bless one of their APUs with > 128 bit wide memory. Seems an obvious bottleneck that would improve GPU performance.
It's coming with the Halo Strix with it's 256 bit wide memory, but should it really have taken over a decade?
Jumping to a wider memory bus is most feasible in product segments that already use soldered CPUs and RAM. PCs have been reluctant to drop the sockets/slots until recently, and only then in thin and light laptops. In systems that don't have discrete GPUs, Intel hasn't been applying competitive pressure against AMD's integrated GPUs.
So it's really been a question of whether AMD could spare the cash to tape out a larger laptop chip with the goal of cutting off the low end of NVIDIA's dGPU sales for laptops.
This is exactly the reason - and perhaps they finally can do it because "Apple showed the way" with the M-series.
They obviously could have done it from a technical standpoint, as they had done it for consoles, but those are not general purpose from the viewpoint of who is laying out the motherboard design.
Once a competitor gets away with a specific controversial or tricky to explain move, the others can pass through the same hole. Apple dropping the headphone jack or glueing the batteries were the same kind of moves, it expanded to other makers because they got away with it.
Or we could look at Twitter charging ridiculous prices for the API and Reddit jumping through the same loop.
In the CPU I think it's more nuanced, but there's definitely a tradeoff that many makers wouldn't do on their own.
When alternatives are Windows laptops that barely work - it doesn’t seem so expensive.
Case in point. Back in 2018, me and my gf bought new laptops. I’ve opted for MBP and she bought Asus Rog Strix. Both were over 2k euro.
Two years after her laptop has constant issues with Windows updating all the time, WiFi modem dies until you power cycle whole machine, it randomly turns on at night or doesn’t go to sleep at all, fans that sound like jet engines even on lowest of speeds.
My MBP 2018 is still a champ after all these years, I had literally zero issues with it and the only reason why I’m contemplating upgrading is how amazing M processors are.
And seriously? More expensive? MacBook Air with 16GB of RAM and M processor starts at 1500 euro. It has far superior build quality, software and battery life than any Linux/Windows machine in any category. And on top of that it runs completely silently. Apple=expensive meme needs to die.
People who have $1500+ to spend on laptops forget that the vast majority of people don't spend (and can't spend) that much on a laptop. So much of the market is in that $500-$600 price range. You wouldn't be wrong to point out that an M1 or M2 Macbook is a far superior device, but it really doesn't matter.
And in terms of specs, Apple is still not an incredible value for the money. The laptop I'm using now is a last-years model Thinkpad. I bought it refurbished, with a two year warranty included, for $250. It's got a 12 core (edit: 6 core, 12 threads) Ryzen 5000 series, and for less than $150 I've added 32 GB of RAM (total RAM 40 GB), and a 1 TB NVMe disk (the computer has slots for two). That's an absolutely crazy deal for $400, and the build quality on these Thinkpads is miles above cheap Dell Inspirons or whatever.
For most people, this is a much more reasonable choice than spending $1500 on a laptop. My last laptop (Dell XPS) lasted over 12 years; I think the "laptop dies in two years" stories are the exception, not the rule, and Apple's laptops are certainly not lemon-proof, in that regard.
AMD's mobile 5000 series topped out at 8 CPU cores and 8 GPU compute units. I'm not aware of any laptops that used AMD's desktop CPUs prior to the 7000 series; there may have been some boutique OEM doing that but almost certainly not Thinkpads.
Apple=expensive meme will only die if it stops being true. For example, I can get a Thinkpad T14 Gen 4 with the best processor AMD has to offer, 32 GB RAM and a 512 GB SSD for less than the base M2 Air (€1250 vs €1300), which has a pitiful 8 GB/256 GB RAM.
If that's too expensive, an E14 Gen 5 has a previous-generation processor but can be had in a reasonable configuration (Ryzen 5, 16GB RAM, 512GB SSD) for around €700.
You can knock a fair chunk (€60) off of the price by getting them without Windows as well, which is a nice option when you're gonna install Linux anyway. But the prices I mentioned is before you even do that!
I picked the upgrade to the low-power screen, which should look much better than the base screen, although probably not as good as a Mac one. OLED would cost an extra €100 compared to that, although it does have some compromises such as higher power draw. Also, current Thinkpads have 1920x1200 screens on most models now, not 1920x1080. And the OLED ones are 2880x1800, which sounds nice but honestly 1200p sounds nicer as it'd be usable without scaling.
Those prices in the French store look absolutely wild to me. In the Dutch store, the base model T14 Gen 4 AMD starts at 999, without any discounts visible!
PC laptops with similar build quality cost the same or more.
x86 machines only compete when you don’t consider power consumption (which really matters for laptops). It’s so bad that most top end x86 machines won’t even try to hit those peak numbers unless the machine is plugged in and then only for short periods of time.
My M1 beats most x86 machines even after it throttles and it’s over three years old now. My M3 Max machine blows away the x86 machines I run into and it does it while still having good battery life.
1. Most x86 machines aren't running the fastest CPUs. They are running U-series CPUs and usually aren't even running the best U-series CPUs. H-series laptops are relatively rare.
2. Even U-series machine have a peak power consumption north of 50 watts. H-series CPUs peak out at over 3x that (157w for Intel 13/14th gen HX CPUs). There's simply no way to disperse that much energy with the small heatsinks available (even in a gaming laptop).
3. Most of these machines downclock massively when not plugged in.
4. The laptop size required for an H-series CPU is massive.
Putting all of that together, an M1 machine is simply faster than most x86 laptops because they contain low-binned, U-series CPUs. For the small fraction that contain H-series CPUs, My M1 is still generally faster if both machines are being used as laptops rather than bad desktops. This performance delta gets even bigger if the workload sustains enough that the x86 machines begin to throttle because they can't sustain the insane peak power draw.
You are comparing the SW, not the HW. Windows 10, by default, sucks on any HW. You can make it to behave a little but you need special tools and knowledge (Ooshutup10 and advanced power modes).
>When alternatives are Windows laptops that barely work - it doesn’t seem so expensive.
Case in point.
Yeah there are some junky laptops out there, but case in point I've only had sub-1000 Euro Windows notebooks from Acer, HP and Lenovo, and all worked fine and lasted no matter how hard I abused them. Maybe I'm just super lucky. Or maybe I was an informed consumer with a trained eye to spot the good from the junk. Or both.
>And seriously? More expensive? MacBook Air with 16GB of RAM and M processor starts at 1500 euro
16GB ram and how much storage?
Yeah, that's pretty expensive for the specs and EU purchasing power. Case in point, for Christmas I bought myself a new Lenovo with the new 4nm Ryzen 7840HS, 32GB RAM and 1TB NVME, for ~800 Euros sticker price, no special promotion or sale. That seemed like a good deal to me.
Now how much does Apple charge for 32GB RAM and 1TB storage? 2500+ Euros for 24GB RAM. Damn son! Apple are only price competitive at the bare minimum base models(8GB RAM on a "Pro" device?! Give me break Apple). When you start adding cores, RAM and storage, the price skyrockets.
Each to his own if the price is worth it, I don't judge. My friend who does a lot of video editing gets his money's worth, but for my uses cases(Windows, Linux, gaming, coding), that 800 Euro laptop will be more than enough, no need to spend way more on Apple for less functionality. Battery life on those Macs is tight tough. If you're on the road a lot, and don't need windows/linux stuff, I'd definitely get one.
Mine doesn't have any of those issues. I think some people are stuck with the impression that all Windows laptops are stuck in 2006 and haven't kept up in terms of quality. Mine is slim, full metal build, 2.5k 16:10 IPS screen.
So what am I compromising on compared to spending over 3x more on a Macbook? Is the Mac display, trackpad and build quality a bit better? Most likely. Is it over 3x better? Definitely not.
> I’ve opted for MBP and she bought Asus Rog Strix. Both were over 2k euro.
She bought a gaming brand. There's no such a thing as a good gaming brand and the issue isn't with PCs but buying something targeted at people who don't even use laptops as /laptops/ so the build quality is horrendous (particularly on the hinges, it's common for gaming laptops to have failed hinges but the consumer target doesn't care because the PC never leaves their desk).
There's no issue with Lenovo ThinkPad or Dell Latitude. There's a wide range of PC manufacturers out there, many of which make classes of computers Apple will never make, like the Panasonic Toughbook which are water, dust and shock resistant to the point where you can use them as weapons to bash someone's head and the computer will still run fine.
By the way, if you're unlucky enough to get a lemon, Dell and Lenovo offer /on site/ warranty service on their business class hardware (another perk of buying latitude rather than consumer targeted garbage like inspiron). You don't have to ship your computer back or go to a specific store like with Apple. They come to fix it.
> And on top of that it runs completely silently
Most PC laptops are quite silent, you can't get silence from a device that was made to push dedicated GPUs to their limits like a gaming laptop.
As for Apple's legendary build quality, the last macbook I've owned before ditching the Apple ecosystem entirely was one that was affected by both the keyboard dust issue and the short display cable that bends too much whenever you open the monitor.
Yeah, great build quality. I think it's the first time in my entire life I saw a keyboard fail so quickly and hard. And that display cable.. who designs things like this? It should have been obvious to anyone who designs computers that putting that much tension on the cable was going to make it rip. I can never rid myself of the suspicion Apple makes hardware designed to fail after warranty/Apple care period after this.
Or they have enough leverage for good contracts, and a lot of money on hand to be able to make semi-risky bets that others would not be willing to. Combined with good marketing and reputation.
That enables them to take bigger bets than most, and their reputation means customers are more likely to buy into their big bets.
Quality products which can be bricked on system update and have so many issues that everyone says to not upgrade for at least 6 months after major release!
Also their software is garbage if you need to connect to or use something outside of apple ecosystem.
MBP user at home for 6 years and switched at least 3 MBPs at work (switching companies and generations of MBPs).
iPhone user for 3 years.
My gf has been using iPhones and iPads since who knows when. A couple of friends who are Apple users. I have heard zero complaints about bricked device.
I routinely heard and experienced issues with Windows and Android, though.
> This is exactly the reason - and perhaps they finally can do it because "Apple showed the way" with the M-series.
This isn't generally what you want as a customer though. Most current games will need something in the range of 8-16GB of VRAM, but will need that in addition to whatever you need in terms of main system memory. If you also want 16GB of system memory, you need 32GB in total but only benefit from 16GB being fast, and you might often want more than 16GB of system memory.
Apple is fine with this because then if you want a lot of RAM they can charge you an arm and two legs for that much of the fast stuff, and solder it all so you can't upgrade it. But what you really want is to solder a modest amount of the fast stuff to the CPU, put it in the CPU socket (so even it can still be upgraded by upgrading the CPU, and then the socket doesn't need all of those pins) and still have the same DDR slots as ever to add gobs of the cheap stuff to suit.
I don't think LPDDR5 is that much more expensive than DDR5 on a per-GB basis. The performance advantage is mostly from a wider memory bus, and a bit from being soldered.
In other words, the CPU having access to huge amounts of memory bandwidth is not really costing the users much, after they've paid for the GPU and its memory system. The cost is really just the incremental cost of enlarging a GPU die by strapping some CPU cores on the side.
It may not be precisely most cost and performance optimized configuration for gaming, but on the other hand it provides huge benefits for anything that doesn't fit in a cheap GPU's VRAM (eg. LLMs).
> I don't think LPDDR5 is that much more expensive than DDR5 on a per-GB basis.
It is by the time the OEMs are through with it, and once you're soldering it you have to buy it from them.
> The performance advantage is mostly from a wider memory bus, and a bit from being soldered.
Even the little bit from being soldered should go away when CAMMs enter the scene. But the wider memory bus is part of what makes it expensive. It's not just the input pins on the APU, it's routing all of those pins through the socket and the system board and the memory DIMMs. Which you can avoid by putting some amount of HBM on the APU package without losing the ability to add memory via the narrower bus in the traditional way.
> It may not be precisely most cost and performance optimized configuration for gaming, but on the other hand it provides huge benefits for anything that doesn't fit in a cheap GPU's VRAM (eg. LLMs).
LLMs are nearly the only thing that currently uses this, and there is nothing stopping a CPU option with a large amount of HBM on the package from existing and fitting into the socket on an ordinary system board for the people who want that.
But CPU makers also don't really want to make a thousand SKUs with every possible combination of cores and memory, so what they're probably going to do is have lower core count CPUs with smaller amounts of HBM and higher core count CPUs with more. In which case you can still buy the $100 APU with a modest amount of HBM and then add as much DDR as you want via the slots on the system board instead of having to buy $1000 worth of CPU cores you don't need, or replace a perfectly good CPU with a different one, just to increase the amount of memory for the majority of applications that aren't bounded by memory bandwidth.
> I don't think LPDDR5 is that much more expensive than DDR5 on a per-GB basis.
It's also worth pointing out that while Apple is using LPDDR5 and getting more bandwidth out of it by just using more pins, that isn't that most GPUs use -- they use some version of GDDR, which is more expensive on a per-GB basis than DDR, because it has more bandwidth per pin.
So putting some "fast" memory on the APU package has the further advantage that you don't have to do what Apple did, you could actually put HBM or GDDR on the CPU package, and then the faster memory would be faster.
Doubling the bus width means more wiring and probably more PCB layers than offering two slots per channel.
Right now, using DDR5 SODIMMs instead of LPDDR5 would throw away half of the benefits of doubling the bus width, so that's a non-starter. Using CAMM style modules may allow for full-speed operation, but those have almost zero adoption so far and I think it's too soon to tell how widespread they'll become.
Whatever option for connecting the DRAM, doubling the bus width doubles the amount of DRAM the OEMs must populate the system with, adding another cost on top of the more expensive SoC and MLB.
No. Desktops make the cost issues even worse, because supporting twice the memory bus width not only means more wires, but also a larger, more expensive CPU socket, and maybe a new heatsink mounting standard.
Current PC desktops have two memory channels and often four slots, because each channel has two slots. You get more bandwidth if you populate both channels, but four isn't better than two. If you wanted four channels the CPU socket would need more pins, because right now the second and third slots use the same pins as the first and second.
What you could do is put some fast/wide DRAM on the CPU package. Then you could still upgrade the fast memory by upgrading the CPU, and the pins don't have to go through the socket. This should also be compatible with existing CPU sockets, and not incompatible with also having ordinary DDR in the existing memory slots. You basically get an L4 cache made out of HBM, which is available to the iGPU.
You get more bandwidth by having all memory channels populated. When you have multiple slots per channel, filling all slots at best gets you a few percent better memory performance due to having more ranks of memory, but for DDR5 platforms the extra load on the memory controller due to multiple DIMMs per channel brings a big hit to the (officially supported) memory clock speeds.
The fastest memory configs for a desktop currently are using one single-rank module per channel. Using one dual-rank module per channel doubles the capacity for typically a slight performance penalty.
If you soldered DDR5 memory, wouldn't you still need to have a lot of precise trace lengths and a lot of wasted space to make that match across banks of chips?
If the solution is to compensate for delays in a smarter way, then I would assume that solution could be applied to sockets too. Which goes back to asking what took them so long.
The footprint of a DIMM slot is not just the area the slot itself takes. DIMMs force you to do a lot of length-matching, which takes a lot of PCB space (and layers). The actual footprint of the DIMMs also includes all the space between the slots and the CPU. Especially LPCAMM2 does this inside the connector, and you can place the connector as close as you can without causing clearance issues.
>> Unfortunately CAMM appears to have a larger footprint than a DIMM slot.
We need a new motherboard form factor that assumes no PCIe graphics card.
I like my ITX box and put the power connector and button where the expansion slot would normally go because I used an APU. But we could do a lot better with the board size by assuming integrated graphics. A lower riser on the back would be nice too - maybe only 2 USB high.
We need a new form factor that allows a small motherboard to be parallel to a back-mounted GPU/PCIe. This is what people are already trying to do with SFF cases and those wonky PCIe cables/boards.
> Nominal module dimensions listed in the standard point to "various" form factors for the modules, with the X-axis measuring 78 mm (3.07 inches) and the Y-axis 29.6–68 mm (1.17–2.68 inches).
From Artechnica. Whereas a regular DIMM is 5.5 inches long & narrow on the Y.
Yeah same. Especially that those consoles were launched all the way back in 2020, nearly 4 years ago.
I guess it's difficult to replicate the console performance when those run on GDDR as system memory and PCs on DDR so the bandwidth difference is huge.
But it's not like a system integrator like Asus or Gigabyte couldn't have built a small form factor PC based on a PC variant of the consoles APU and use GDDR instead.
AMD really missed out on this, especially during the pandemic when everyone was stuck indoor trying to game and looking for inexistent GPUs.
GDDR isn't necessarily all that better on use for general purpose CPU.
Most importantly, buyers didn't want machines with soldered memory, and that was main method of going with wider bus without prohibitively expensive boards etc.
Yes, but what's your point with those facts? Those were Bulldozer based APUs, not very good compared to what you could buy for PC back then, in an era where AMD wasn't very competitive and GPUs weren't overpriced unobtanium.
Just pointing out that AMD has been shipping cheap APUs with nice memory systems that increase iGPU performance for over 10 years.
Just mystifies me that AMD hasn't shipped similar in a laptop, SFF, or desktop yet. Especially during the GPU shortage. The AMD 780m iGPU does pretty well, clearly AMD could add more cores and ship a faster memory bus, like the ones on their Threadripper (4 channel), Siena (6 channel), threadripper pro (8 channel), or Epyc (12 channel).
Apple's doing similar with M1/m2/m3 (128 bit), M1/M2/M3 Pro (256 bit), and M1/M2/M3 Max (512 bit wide) available in laptops, tiny PCs (up to 256 bit wide) with the mini, and small desktops (studio).
Jaguar was the codename for the CPU core microarchitecture. AMD had two separate families of microarchitectures, much like Intel's Core and Atom families (since united in one chip in Alder Lake and later).
What's to stop someone from taking a mid-high end dedicated graphics card and soldering on a CPU, RAM, and IO? (Do they do this already?) I know it sounds like sacrilege, but I suspect most people upgrade their mobo/CPU/RAM/graphics all at once.
Soldering doesn't do anything about pin count. An off the shelf GPU chip is still going to support at most a PCIe x16 interface to other chips, which isn't enough to share DRAM with a separate CPU.
Taping out a new chip that combines existing GPU and CPU IP blocks is relatively straightforward, but taping out a large chip is extremely expensive.
Apple does not use HBM, just LPDDR5 with wider interfaces. Xbox X/One, PS4/5, AMD server/workstation chips (threadripper (and pro), siena, and epyc), and apples above the base model m1/m2/m3 all use wider memory interfaces.
AMD agpu (and Intel with iGPUs) do use unified memory. As well as the high end AMD (MI300A) and Nvidia (Grace+hopper) server+iGPU chips.
This REALLY helps GPUs performance which is very bandwidth intensive, but even normal workloads have cache misses. Part of the benefit isn't just more bandwidth, but that you can have 4,8, or 16 cache misses running in parallel which can help keep more of your cores doing useful work instead of waiting on memory. This doesn't decrease memory latency, but can help improve throughput when using multiple cores.
Sad that even a M3 pro (the midrange, 3rd fastest in the apple M series lineup) available in Apple laptops or mac minis has more bandwidth than every AMD Ryzen and every Intel i3/i4/i7/i9. Even Intel's fastest i9-14900KF using a maximum of 250 watts has 1/8th the memory bandwidth of the M2 Ultra, 1/4th the bandwidth of the M3 max, and 1/2th the bandwidth of the M2 Pro. Same (or less actually) then the MBA or base Mac mini.
Ah! I thought 64bit buswidths had been standard in DDR memory for, well, a very long time now? Why did that change? Sounds like nothing much changed actually, in terms of total (bus) bandwidth.
Yeah, up to DDR4 it's standard, or at least common to have 64 bit channels. But with DDR5 it's standard to have 2 32 bit channels per dimm. Unfortunately many confuse number of dimms with number of channels, which is wrong for DDR5.
It's coming with the Halo Strix with it's 256 bit wide memory, but should it really have taken over a decade?