LPDDR, GDDR, HBM: Three Ways to Feed a Processor

"Every memory technology is a different answer to one question: how many pins can you afford, and how far away can they be?"

Three Answers

LPDDR — low-power double data rate — is the phone's memory: narrow channels, low voltage, chips soldered close to the processor on a package or board, tuned for energy per bit. It reaches large capacities cheaply because its dies are ordinary DRAM and many can hang off each channel. Apple's Macs, AMD's Ryzen AI Max, Qualcomm's laptop chips and NVIDIA's DGX Spark all use it. GDDR — graphics DDR — is the discrete card's memory: chips arranged around a GPU on a circuit board, each on a wide, fast, hot channel, giving high bandwidth at moderate capacity; the RTX 5090's 32 GB at 1,792 GB/s is the state of that art. HBM — high-bandwidth memory — stacks DRAM dies vertically and sets the stack on a silicon interposer beside the processor, with an interface so wide it could not run across a board at all; that is the data-centre answer, with bandwidths in the tens of terabytes per second and prices to match.

The Table, With Its Labels

Part	Technology	Capacity	Bandwidth	GB/s per GB	Evidence
Apple M3 Ultra (office)	LPDDR5, on package	512 GB	819 GB/s	1.6	vendor
Apple M5 Ultra	LPDDR5X, on package	512 GB	"1.2TB/s"	2.4	vendor (2026-08-25)
Apple M5 Max	LPDDR5X, on package	128 GB	614 GB/s	4.8	vendor
NVIDIA DGX Spark (GB10)	LPDDR5X-8533, 256-bit	128 GB	273 GB/s	2.1	vendor
NVIDIA Vera CPU	LPDDR5X SOCAMM, detachable	"up to 1.5 TB"	"up to 1.2 terabytes per second"	0.8	vendor (Micron says up to 2 TB — conflict)
GeForce RTX 4090	GDDR6X, 384-bit	24 GB	~1,008 GB/s	42	physics from 21 Gbps × 384 ÷ 8; NVIDIA prints no figure
GeForce RTX 5090	GDDR7, 512-bit	32 GB	"1792 GB/sec"	56	vendor
RTX PRO 6000 Blackwell	GDDR7 with ECC, 512-bit	96 GB	"1792 GB/s"	18.7	vendor
NVIDIA H100	HBM3	80 / 94 GB	"3.35TB/s \| 3.9TB/s"	42	vendor
NVIDIA H200	HBM3e	141 GB	"4.8 terabytes per second"	34	vendor
NVIDIA Rubin GPU	HBM4	"Up to 288 GB"	"up to 22 TB/s"	76	vendor (2026-01-05)
AMD Instinct MI455X	HBM, 12 stacks	432 GB	23.3 TB/s	54	vendor

The last numeric column is the one to stare at. Bandwidth per gigabyte is how fast a byte can be re-read relative to how many bytes there are, and it sorts the technologies more sharply than either raw number: LPDDR at 1–5, GDDR at 18.7–56, HBM at 34–76. Since decode re-reads every byte of the weights per token, GB/s per GB is very nearly tokens per second per model that fills the memory. A card that fills its 32 GB decodes that model at up to 56 tokens per second; a Mac that fills its 512 GB decodes that model at 1.6. Same formula, and the reason the two machines are not rivals so much as neighbours on different streets.

What Each One Costs

LPDDR buys capacity and energy with narrow channels, so it is the only one of the three that reaches half a terabyte on a desk, and the only one whose bandwidth cannot be raised without a wider package. GDDR buys bandwidth with wide hot channels on a board, and pays in capacity — the chips are small and a card has room for a ring of them and no more. HBM buys both, on an interposer, at a manufacturing cost Micron's chief executive described in December 2025 as a "three-to-one trade ratio with DDR5": every gigabyte of HBM made is three gigabytes of ordinary DRAM not made, which is where its price and its scarcity come from. Apple's homework track asks whether a Mac could ever move up this table; the honest starting point is that the three technologies are three different machines' worth of trade-offs, not three settings of one dial.

Code

bandwidth_per_gb.py — the column that sorts memory technologies·python

#!/usr/bin/env python3
"""GB/s per GB = how many times per second the whole memory can be re-read.
For decode, that is roughly tokens/s for a model that fills the memory.
Vendor figures except where marked DERIVED."""

PARTS = [
    # part,                    tech,      GB,   GB/s,  evidence
    ("Apple M3 Ultra",         "LPDDR5",  512,   819,  "vendor"),
    ("Apple M5 Ultra",         "LPDDR5X", 512,  1229,  "vendor '1.2TB/s'"),
    ("Apple M5 Max",           "LPDDR5X", 128,   614,  "vendor"),
    ("NVIDIA DGX Spark GB10",  "LPDDR5X", 128,   273,  "vendor"),
    ("NVIDIA Vera (SOCAMM)",   "LPDDR5X", 1536, 1200,  "vendor 'up to'"),
    ("RTX 4090",               "GDDR6X",   24,  1008,  "DERIVED 21 Gbps x 384b"),
    ("RTX 5090",               "GDDR7",    32,  1792,  "vendor"),
    ("RTX PRO 6000 Blackwell", "GDDR7",    96,  1792,  "vendor"),
    ("H100 SXM",               "HBM3",     80,  3350,  "vendor"),
    ("H200",                   "HBM3e",   141,  4800,  "vendor"),
    ("Rubin GPU",              "HBM4",    288, 22000,  "vendor 'up to'"),
    ("AMD MI455X",             "HBM",     432, 23300,  "vendor"),
]

print(f"{'part':24} {'tech':8} {'GB':>5} {'GB/s':>6} {'GB/s per GB':>12}   evidence")
for name, tech, gb, gbs, ev in sorted(PARTS, key=lambda p: p[3] / p[2]):
    print(f"{name:24} {tech:8} {gb:5d} {gbs:6d} {gbs / gb:12.1f}   {ev}")

print("\nA model that fills the memory decodes at roughly GB/s-per-GB tokens per second.")

Exercise

Run bandwidth_per_gb.py and add your Mac as a row. Then find the largest 4-bit model (from the physics track's bytes-per-token accounting, roughly 0.6 GB per billion parameters) that fills your Mac's GPU working set, and predict its decode ceiling two ways: from GB/s per GB, and from achieved bandwidth ÷ bytes per token. Explain in one sentence why the two agree only when the model fills the memory.

Hint

GB/s per GB assumes the model is as large as the memory. A 9B model on a 512 GB Mac reads 4.5 GB per token, not 512, so its ceiling is ~180 tokens per second, not 1.6. The column ranks technologies; the formula ranks models.