what do you mean? the next apple X generation will come at least 12 months later or even 18 months if they go ipad path
And what others? i bet for the next 2 years, no one could deliver this kind of cpu+gpu performance/W
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Expected GPU performance of M1X/M2
- Thread starter joelypolly
- Start date
- Sort by reaction score
And what others? i bet for the next 2 years, no one could deliver this kind of cpu+gpu performance/W
My guess is extra battery. If they announced a mobile workstation that with a workstation load drained the battery in a couple of hours, they’d get all the flack that they got from the intel machines throttling. If it can do 8 hours at full tilt, that’s something that would shut naysayers up and get props from people who actually want to get stuff done.
its not legal to go extra battery on the 16" it already has Built‑in 100‑watt‑hour lithium
Again if in fully load my Intel 16" with dGpu could keep around 3-4 hours ...i bet this M1x can keep for double that in fully load...so , no is not the extra battery
Hmm… good point about the 100 Wh aviation limit. Maybe they figured out a clever loophole? That’ll be interesting to find/figure out tomorrow.
Last edited:
For certain tasks iPad performs better than a "full computer" though. Editing photos on an iPad with Apple pencil is easier for me than doing it with a mouse. But the software on iPad is always missing certain pieces compared to the desktop version, but eventually that will disappear.
This is indeed a fascinated topic! I learned most of what I know from this very well written blog series:
A look at the PowerVR graphics architecture: Tile-based rendering - Imagination
Rys looks at our PowerVR graphics architecture and describes how Tile-Based Rendering (TBR) works in practice.
www.imaginationtech.com
A look at the PowerVR graphics architecture: Deferred rendering - Imagination
The deferred rendering technique in PowerVR GPUs takes the information generated by the tiler to defer the rendering of subsequently generated pixels.
www.imaginationtech.com
It is worth noting that Apple likely adapted their TBDR hardware from Imagination directly (Apple's shading backend is entirely different though).
@mr_roboto already summed it up in a very neat way. I would only add that tile based rendering as most modern GPUs do is a caching optimization (by processing triangles in a close proximity to each other it's more likely that you will fetch texture data that is closely laid out in the memory), but triangles are still rasterized and shaded immediately. The big difference to TBDR GPUs (currently only Apple and IMG) is that they will first rasterize every single triangle in the tile and only then do shading. This means that every visible pixel is shaded exactly once. While the result is the same, the underlaying implications are actually very significant: TBDR GPUs can dispatch shading work in regular grids of 32x32 pixels (always doing 1024 pixels per shader invocation), while immediate renderers have to dispatch shaders for each triangle individually (they usually side 8x4 groups of pixels at once, which creates inefficiencies at triangle edges). Also, the TBDR model offers guarantees that the immediate model simply can not. For example:
- with TBDR you know that you "own" the pixel as no other shader invocation be computing the value of that pixel at the same time (in immediate rendering there can be data races, e.g. two overlapping triangles may be shaded simultaneously). This allows you to do things like race-free framebuffer reads that are key to programmable blending and other advanced techniques
- with TBDR you know that all other triangles in the same tile have already been rasterized, this allows you to deterministically track the state of multiple pixels at once. This again allowed Apple to do implement some really cool stuff like persistent cache between shader invocations (multiple shaders can work on the data in sequence, something that no other GPU supports) which is the key to performing some advanced effects without ever leaving the GPU cache
You are right of course. I was stressing the bandwidth issue since Apple GPUs are much more limited in this area. TBDR allows them to fetch only what is needed, saving precious bandwidth.
I don't thin it's weak per se, it just won't be more efficient (it will have to shade transparent triangles immediately, just like an IMR would). There is still some potential for efficiency wins (e.g. if you have multiple non-overlapping transparent triangles in a tile). I suspect that the GPU will flush the tile when a newly rasterized transparent pixel hits a not-yet-shaded transparent pixel. So basically, if you have a lot of transparent pixels with plenty overdraw, TBDR performance will tank due to constant flushing. But so will IMR performance.
Transparency is a big problem anyway, and if I remember correctly the optimization tips for TBDR and IMR are identical: draw transparent objects last (you have to draw them last and sorted anyway if you want correct results). And then again there is order independent transparency which again can be done more efficiently on Apple GPUs.
It is not close, since it's not deferred. Saying that they are close because they both use tiles is like saying that a hybrid is basically the same as a full electric car
Using tiles makes sure that the modern GPU will on average fetch texture data from similar locations, which massively improves caching and is the key for the large performance and efficiency increased of Maxwell and Navi. But true TBDR is much more difficult to achieve because there are just so many edge cases... Imagination was basically the only company to ever do it, and Apple bought the tech from them.
And regarding most mobile GPUs (Mali, Adreno), these are simply atrocious. They cut a lot of corners in order to achieve higher efficiency (e.g. shader splitting) which results in inconsistent programming models. They don't scale with complex geometries since they don't use deferred shading. But it's good enough for lower quality games on mobile that don't try to do anything ambitious.
Last edited:
If benchmark performance could in fact be on par or close to a 3060 or even 3070 I'd be so freaking impressed.
Prepare to be impressed then
With a caveat — they should be on par with the mobile Ampere GPUs (~100W TDP). The desktop class Ampere is something entirely different... can't really expect a 40W Apple GPU to compete with a 200W monster.
When showing off Apple Silicon originally they showed SOTR, then they snuck in some Metro Exodus at the last WWDC. Do we think they will get 4AGames to show Metro Exodus Enhanced Edition this time around?Prepare to be impressed then
if M1 max score 62 Mh/s Eth hashrate at 50 watt, it will be instantly out of stock.
4x would put it in W5700XT territory of the Mac Pro, at what looked like a 55W envelope. Jeeeebus.
I am somewhat disappointed there still was no mention of realtime ray tracing in hardware.
Yeah, I am happy about being able to move to a smaller notebook for my next go around though. It will be interesting to see what comes of having the (potential for) equivalent to PS5 in Teraflops on Apple's most sold systems will do though.
Going by Metal Benchmarks, the M1 Max should land just shy of the 6700, TBH. Peak Flops are under 11TFlops for the M1 Max, which puts it more in the 5700XT category at around 9.7TFlops. Still, the W5700XT (which I've used) is a 200+W card facing off against a 55W GPU design.
If they even get in the ballpark I'll be happy.
Does anyone know if the single core performance of the M1Pro and M1 Max is any different? That is the key factor for me in choosing which chip I need.