Video Summary - Why PC Memory is Stuck in 1995 (Unified Memory Explained)

Key technological concepts and why PC memory feels “stuck in 1995”

Memory demand surged due to AI, making RAM/“memory” feel more scarce and valuable (described as “more valuable than gold” in the last year or so).
Even when people upgrade system RAM, they may still need a separate GPU with enough VRAM to run:
- games
- video editing
- local AI This creates the feeling of “buy memory twice” (CPU RAM + GPU VRAM).
A core issue is that CPU and GPU don’t share memory efficiently:
- Data transfers between them go through PCIe, creating performance and power bottlenecks.
- This can also cause synchronization/duplication overhead by keeping separate copies of data.

Unified memory aims to let CPU and GPU access a single shared memory pool, reducing or eliminating frequent copies across PCIe.
The idea isn’t brand new:
- Apple didn’t invent unified memory.
- Earlier approaches like Intel integrated graphics attempted similar concepts but struggled due to challenges when sharing across processors with different access patterns.

Optimized more for bandwidth (moving large amounts of data quickly) than raw latency.
GPUs run massively parallel workloads, requiring fast transfers of large blocks of data.

Even with separate CPU and GPU memory, the real limiter is transporting data between them via the bridge called PCIe.
Bandwidth comparison mentioned:
- VRAM bandwidth ~ 1.8 TB/s
- PCIe Gen5 bandwidth ~ 64 GB/s
When transfers are needed, the system funneling through PCIe can become the bottleneck.

Integrated GPUs (iGPUs) and mobile SoCs place RAM closer to the CPU, but often through narrow buses (e.g., 64-bit / 128-bit):
- Works for CPU-like workloads and even some gaming
- But isn’t ideal for bandwidth-hungry GPU workloads
Consoles reduce some of these issues by using one high-bandwidth memory pool for the system:
- Example: 256-bit on PS5, 320-bit on Xbox Series X
This works well because consoles largely focus on GPU-heavy gaming.

Apple uses LPDDR (lower power) instead of console-like GDDR, compensated with very wide memory interfaces:
- 128-bit on base M1
- Up to 1,024-bit on M1 Ultra (described as ~16 memory channels feeding one chip)
Claimed result:
- Up to ~800 GB/s bandwidth using a single shared memory pool
Apple also improves dynamic allocation vs older shared-memory designs:
- Instead of fixed “slices” (CPU vs GPU getting reserved portions),
- Either side can draw from the whole pool on demand, reducing idle/reserved memory.

A key motivation is AI model size:
- The video claims Macs with unified memory can run very large models (e.g., ~70B LLaMA-class models).
- With discrete GPUs, such a model may not fit even on a 32GB VRAM card, requiring multiple GPUs and much higher costs.
Similar benefits apply beyond AI:
- Rendering large video timelines
- Large 3D scenes
- Other workloads that previously demanded lots of VRAM

Memory can’t easily be upgraded
- Traditional PCs let you add RAM later.
- With unified memory (as described), configurations may be effectively fixed and could require soldering changes (and may involve warranty risk).
Resource starvation risk
- Since CPU and GPU share one pool, a GPU-heavy workload can consume bandwidth and starve the CPU, and vice versa.

The video argues the industry is moving toward unified/shared memory on PCs too, citing:
- AMD Strix Halo
- Nvidia RTX Spark
The push is framed as efficiency-focused, not just speed:
- Includes energy considerations like energy per bit and the cost of moving data (picojoules/bit).

For years, people prioritized CPU/GPU compute and treated memory as an afterthought.
The video’s takeaway:
- Memory is probably the most important part now—sometimes you can gain more real performance by choosing better memory, even if compute is slightly lower.

Prashant Nair, Associate Professor of Electrical and Computer Engineering, University of British Columbia (discusses bandwidth/bottlenecks).