Hell, SGI O2s from 1996 had this. For all of the hype the performance gains were pretty modest.

FWIW, the O2's UMA let it handle far more textures than almost any other contemporary system with reasonable performance.

Most other SGIs had single or low double-digit megabytes of texture memory, whereas the O2 could host one gigabyte of unified memory and use a huge chunk of that for textures.

UMA was never about performance and it still isn't. Spark is slower than a 5090.

did they learn why? were there other gains?

O2 GPU was slower than other SGI options at the time, however it could use hilariously larger pool of memory without copying, which meant that O2 could use approaches that were punishingly hard (very tight transfer loops) or impossible (huge textures that couldn't be virtualized due to needing whole texture).

That was because unlike other GPUs at the time, O2's didn't have dedicated memory but shared the memory with CPU - way slower, but zero copies and bigger.

Arguably early home computers and workstations also used "unified memory" :D

FYI it existed long before that. Shared memory between CPU and iGPU has been a thing for a long time.

Zero-copy shared memory?

yes, here is 2013 AMD presentation of the topic as example: https://events.csdn.net/AMD/GPUSat%20-%20hUMA_june-public.pd... see slide 14 especially

Ah. Well, what kind of consumer hardware/software combo could I purchase to use this? outside of perhaps the... PS4?

Everything that doesn't have a discrete GPU has unified memory these days. If you're asking for something closer to the RTX Spark or Apple Silicon then look at AMD's Strix Halo systems.

> Everything that doesn't have a discrete GPU has unified memory these days.

Sorry, I meant before the M1 came out. And you and I both know that "unified memory" doesn't refer to allocating ram to the gpu for zero-swap sharing.

Also, if you'd like to know of an earlier example, the very first Raspberry Pi has fully unified memory. Released 2012, SoC from 2011. It's not exposed in the usual APIs which is why you have a configurable RAM allocation for the GPU, so you need to write special code targeting the GPU cores. You can pass a pointer from the CPU and have the GPU read data directly. Results are written back to RAM which can be read by the CPU.

> Unified memory is supported on Linux by all modern AMD GPUs from the Vega series onward

Vega series is 2017.

https://rocm.docs.amd.com/projects/HIP/en/docs-6.3.0/how-to/...

Every AMD APU since introduction of HSA did it, which is how AMD ended up doing SoCs for PS4, PS5, and Xbox

Ok, so which one of these contemporary or previous chipsets could compete with the M1 for inference? Perhaps I'm missing some major detail.

The M1 isnt particularly good at inference, so pretty much every major current competitor with a 256+ bit unified memory system is better: AMD Strix Halo, NVIDIA DGX Spark, possibly Intel Panther Lake

Sure, but none of these shipped before the M1. That was the first chip I encountered that managed to do something useful without a discrete GPU.

Even in Apple land M1 isn't the first with unified memory - pretty much all intel on-chip GPU (Sandy Bridge and newer) count - it was even a reason for driver issues early in intel's new dedicated GPU lineup, as the drivers expected unified memory - but M1 is essentially modification of an iPad chip, and you can see "unified memory" there going all the way back to first chip Apple bought from Samsung to power iPhone 2G

You were likely only using Intel systems then. AMD systems have had iGPUs capable of light gaming for a very long time. It took Intel a long time to get to that level, after M1 (2020).

You are missing a major detail: integrated GPUs are crap. They win on efficiency but not on raw compute. Before AI (and crypto too, I guess) people bought GPUs to render graphics and that was their main consumer. People built more and more demanding games that required increasingly powerful GPUs to render well. Gaming systems always had a discrete GPU so there was no reason to scale up integrated GPUs because they wouldn't sell, or they would be a waste of die space.

I don't think the M1 specifically focused on inference. Their goal was to replace Intel/AMD/Nvidia with their own chips, and since the previous Macs shipped discrete GPUs, they had to match or beat those so they don't ship something slower.

I want unified but not uniform - everything can address anything, but you can add slower RAM to the system without requiring an entirely new chip. NUMA is cool.

AMD Fusion knocking from 2010.