HBM Better GDDR Cheaper for Consumers | Generated by AI

Home 2026.07

Question: Why did NVIDIA move back from HBM2 to GDDR6/GDDR6X in consumer GPUs if HBM2 is better?

Answer:

The short answer: HBM is better technically, but GDDR is much cheaper and easier to scale for consumer GPUs.

NVIDIA did not “downgrade” from HBM2 to GDDR6X. They split the market:

Data Center:
V100 -> A100 -> H100 -> H200
        HBM2 -> HBM2e -> HBM3/HBM3e

Consumer:
GTX/RTX Pascal -> Turing -> Ampere -> Ada
                GDDR5 -> GDDR6 -> GDDR6X

Different optimization targets.

1. HBM is expensive

HBM requires:

GPU die
  |
silicon interposer
  |
HBM stacks

Example:

        HBM
       ||||||
    ===========
    Interposer
    ===========
       GPU

The GPU and memory become one advanced package.

Problems:

For a $599 RTX card, adding HBM could destroy the economics.

GDDR:

GPU ---- PCB ---- GDDR chip
GPU ---- PCB ---- GDDR chip
GPU ---- PCB ---- GDDR chip

Much cheaper.

2. Consumer workloads don’t need extreme bandwidth as much

A V100:

32GB HBM2
900 GB/s

is designed for:

matrix multiply
large batch training
scientific computing

A gaming GPU:

RTX 4070:
12GB GDDR6X
504 GB/s

also has:

huge L2 cache
compression
rendering optimizations

Games do not continuously stream giant tensors like training does.

3. NVIDIA still uses HBM where it matters

Look at H100:

H100 SXM:
80GB HBM3
3.35 TB/s bandwidth

Because AI training:

Transformer:

Attention:
Q @ K^T

MLP:
X @ W

is basically:

large matrix multiplication
+
memory movement

Bandwidth is everything.

4. Why GDDR6X became good enough

Micron’s GDDR6X uses PAM4 signaling:

Traditional:

0 -> 1
two voltage levels

GDDR6X:

00
01
10
11

four voltage levels

More bits per clock.

So NVIDIA got:

GDDR6:
14-16 Gbps

GDDR6X:
19-24 Gbps

without HBM cost.

5. The interesting AI angle

Your MI50/V100 observation is actually correct:

Old data center cards have a hidden advantage:

MI50:
16GB HBM2

V100:
16/32GB HBM2

RTX 4070:
12GB GDDR6X

For some AI workloads:

VRAM capacity
+
memory bandwidth

matter more than gaming-era TFLOPS.

That is why used V100 32GB, MI50, P100, etc. still have a niche.

But the reason NVIDIA doesn’t put HBM on RTX cards is simple:

RTX customer:
"I want $599 GPU"

AI customer:
"I pay $30,000 for H100"

Different markets.

The future is probably a split architecture:

HBM is not replacing GDDR; they are optimized for different economics.

Back Donate