Very simple: thermal budget. Chip performance is limited by thermal budget. You just can't spend more than roughly 100W in a single package, without going into very expensive and esoteric cooling mechanisms.

this is mostly wrong. The real issue has always been memory bandwidth. the highest end consumer x86 CPU has about the same memory bandwidth as a dGPU from 10 years ago. The M1 is extremely competitive with modern dGPUs, only a bit behind a 6900 XT.

If Intel/AMD are seriosus about iGPU, they implement solution for memory bandwidth (they did, Intel Iris Pro eDRAM, AMD gaming consoles using GDDR, upcoming AMD stacking SRAM). So I believe the core problem is that market didn't seriously want great iGPU but just fine with poorer iGPU or rich dGPU by Nvidia.

>The M1 is extremely competitive with modern dGPUs, only a bit behind a 6900 XT.

Do you have a source for this?

Apple compares the M1 Max as having similar performance to a nVidia's 3080 Laptop GPU, which scores around 16,500 on passmark. For comparison, the AMD 6900 XT desktop CPU scores 27,000, while the nVidia 3080 desktop GPU scores 24,500.

So the M1 Max is not as fast as a high end desktop GPU. Still, it is incredible that you are getting a GPU that performs slightly less than a last generation 2080 desktop GPU at just 50-60 watts.

Yes, apple's marketing materials claim 400 GB/s, while the 6900 XT is 512 GB/s. This is very easily Googled. While memory bandwidth isn't everything, it is the major bottleneck in most graphics pipelines. An x86 cpu with 3200 MHz memory has about 40 GB/s of bandwidth, which more or less makes high end integrated graphics impossible.

Ah, I misunderstood your comment. When you said it was competitive with the 6900 XT I thought you were talking about GPU performance in general, not just in terms of memory bandwidth.

According to the numbers Apple is touting, the M1 Max is competitive with modern GPUs in general, being on par with—roughly—a 3070 (laptop version) or a 2080 (desktop version). They've still got a ways to go but this is shockingly close, particularly given their power envelope.

> this is mostly wrong. The real issue has always been memory bandwidth.

Not really wrong. Memory bandwidth is only a limitation for a very narrow subset of problems.

I've gone back and forth between server-grade AMD hardware with 4-channel and 8-channel DDR4 and consumer-grade hardware with 2-channel DDR4. For most of my work (compiling, mostly) the extra memory bandwidth didn't make any difference. The consumer parts are actually faster for compilation because they have a higher turbo speed, despite having only a fraction of the memory bandwidth.

Memory bandwidth does limit certain classes of problems, but we mostly run those on GPUs anyway. Remember, the M1 Max memory bandwidth isn't just for the CPU. It's combined bandwidth for the GPU and CPU.

It will be interesting to see how much of that memory can be allocated to a M1 Max. It might be the most accessible way to get a lot of high-bandwidth RAM attached to a GPU for a while.

GP is talking specifically about GPUs. iGPUs are 100% bottlenecked by memory bandwidth; specifically it is the biggeset bottleneck for every single purchasable iGPU on the market (excluding M1 Pro/Max).

Your compute anecdotes have no bearing on (i)GPU bottlenecks.

They talking specifically about GPUs.

As the charts Apple shared in the event showed, you hit diminishing returns in performance/watt pretty quickly.

Sure. It'd be tough to be the top performing chip in the market, but you can get pretty close.

I dunno. Setting power limits on 3090 at >50% has nearly linear effect on performance.

> I never could see any fundamental reason why "integrated" should mean "underpowered."

There was always one reason: limited memory bandwidth. You simply couldn't cram enough pins and traces for all the processor io plus a memory bus wide enough to feed a powerful GPU. (at least not in a reasonable price)

We solved that almost a decade ago now with HBM. Sure, the latencies aren't amazing, but the power consumption numbers are and large caches can hide the higher access latencies pretty well in almost all cases.

PS4 / PS5 / XBox One / XBox Series X are all iGPU but with good memory bandwidths.

Only time that I can remember HBM being used with some kind of integrated graphics was strange Intel NUC with Vega GPU and IIRC correctly they were on the same die.

That product had an Intel CPU and AMD GPU connected via PCIe on the same package, not the same die. It was a neat experiment, but it was really just a packaging trick.

Still confused how 32 core M1 Max competes with Nvidia's thousands-of-cores GPUs. Certainly there are some things that are nearly linear with core count, or otherwise they wouldn't keep adding cores, right?

Edit: Found answer here. GPU core is not the same thing as a CUDA core. https://www.reddit.com/r/hardware/comments/73i3ne/why_do_app...

The desktop RTX 3070 has 46 SMs, which are the most comparable thing to Apple's cores.

NVIDIA defines any SIMD lane to be a core. They recently have gotten more creative with definition, they were able to double FP32 executions per unit (versus previous gen) and hence in marketing materials, doubled the number of "CUDA cores".

Funny, all these years I've been wondering how they possibly packed so many "cores" into those things.

The apples to apples comparison would be CUDA cores to execution units. Basically how many units which can perform a math operation. Apple's architecture has 128 units per core, so a 32 core M1 Max has the same theoretical compute power as 4096 CUDA cores. This of course doesn't take into consideration clock speed or architectural differences.

CPU core == Cuda SM

Perhaps with Vista? "Integrated" graphics meant something like Intel 915 which couldn't run "Aero". Even if you had the Intel 945, if you had low bandwidth RAM graphics performance still stuttered. Good article: https://arstechnica.com/gadgets/2008/03/the-vista-capable-de...

Video game consoles have been using integrated graphics for at least 15 years now, since Playstation 3 and Xbox 360.

You are mistaken. On both PS3 and Xbox 360 CPU and GPU is on different chips and made by different vendors(CPU made by IBM and GPU by Nvidia in case of PS3 and CPU by IBM and GPU by ATI for Xbox 360). Nonetheless in PS4/XOne generation they both use single die with unified memory for everything and their GPU could be called integrated.

For the 360, from 2010 production (when they introduced the 45nm shrink), the CPU and GPU was merged into a single chip.

When they did that they had to deliberately hamstring the SOC in order to ensure it didn’t outperform the earlier models. From a consistency of experience perspective I understand why, but it makes me somewhat sad that the system never truly got the performance uplift that would have come from such a move. That said there were significant efficiency gains from that if I recall.

Yup. Prior they were absolutely different dies just on the same package

If you mean including PS3 and X360, these two consoles had discrete GPUs. The move to AMD APUs was on the Xbox One and PS4 generation

Longer since integrated graphics used to mean integrated onto the north bridge and it's main memory controller. nForce integrated chipsets with GPUs in fact started from the machinations of the original Xbox switching to Intel from AMD at the last second.

In that case, it's more like discrete graphics with integrated CPU :)

Yeah, and vendors like Bungie are forced to cap their framerates at 30fps (Destiny 2).

They capped PC as well.

If they did, it definitely wasn't at 30. I was getting 90+ on my budget rig.

But no, I don't think they did.

Destiny 2 is capped on PC? The cutscenes are but the actual game is not

It used to have a bug that randomly cap the fps at 30. Only toggle vsync on and off again can fix it. It have no idea whether that has been fixed.

The software side hasn't been there on x86 GP platforms, even though AMD tried. It's worked out better on consoles.

What software is missing? I figured the AMD G-series CPUs used the same graphics drivers and same codepaths in those drivers for the same (Vega) architecture.

My impression was that it was still the hardware holding things back: Everything but the latest desktop CPUs still using the older Vega architecture. And even those latest desktop CPUs are essentially PS5 chips that got binned out.

Deep OS support for unified memory architectures for one. Things they tried to do with HSA etc. Also NVidia winning so much gpu programming mindshare with Cuda, and OpenCL failing to take off on mobile, dooming followon opencl development plans, didn't help.

In the wider picture, gpu compute in general on PC also failed to become mainstream enough to sway consumer choices. Development experience for GPUs is still crap vs the cpu, the languages are mostly bad, there's massive sw platform fragmentation among os vendors and gpu vendors, driver bugs causing OS crashes left and right, etc.

Re your impression, yes, AMD shifted focus more toward cpu from gpu in their SoCs after a while when their initiatives failed to take off outside consoles. But it's been an ok place to be, just keeping the gpu somewhat ahead of Intel competition and getting some good successes in the cpu side.

iGPU Vega is actually really, really good esp when it comes to perf/watt. It is bottlenecked by the slow memory bandwidth. DDR5 will more or less double iGPU performance.