Samsung's “new” August zHBM was unveiled in February, with a verified 4x claim

An AI accelerator can sit idle when data does not arrive fast enough, even if its compute chip is fast. Samsung's zHBM and the HBF project led by SK hynix and Sandisk both attack that delay, but they are not direct substitutes. One shortens the path to fast memory; the other adds a much larger nearby storage layer.

zHBM was already public in February

Samsung first described zHBM at Semicon Korea on February 11, 2026. Conventional HBM sits beside an accelerator, while zHBM would stack it vertically above the logic chip. Reports carrying CTO Song Jai-hyuk's remarks cited potential fourfold gains in bandwidth and power efficiency versus HBM4. Samsung gave no commercialization date.

The claim that zHBM debuted in August, with 8x speed and 3x performance per watt versus HBM5, is not supported by the public material reviewed here. The reported targets are not measurements from a shipping product. Customer, yield, thermal, and production details remain missing.

HBF gained the new specification

HBF stacks NAND flash and places it between HBM and storage. It keeps frequently used model weights and temporary data closer to the accelerator, reducing trips to slower drives. It does not match HBM latency, but it can hold much more data in one package.

The draft specification describes 8-high or 16-high stacks, up to 512GB per package, and 0.4-3.0TB/s of bandwidth. SK hynix and Sandisk are standardizing it under the Open Compute Project. Google and Tenstorrent have joined, while Nvidia, AMD, and Samsung are not confirmed participants. Accelerator adoption matters more than the headline capacity.

The HBF draft specifies up to 512GB and 0.4-3.0TB/s, while major accelerator participation remains limited (Tom's Hardware, August 5, 2026)

출처: tomshardware.com

The timetables are different

SK hynix and Sandisk are standardizing HBF as a new layer between HBM and SSDs (SK hynix, February 25, 2026)

출처: news.skhynix.com

What decides the next memory contest

Samsung can combine memory, foundry, and advanced packaging inside one company. That advantage also means zHBM must solve heat, yield, and customer-accelerator integration together. The HBF group has a nearer timetable and an open standard, but large capacity has little value without accelerator adoption. Samples now matter more than multipliers. If Sandisk delivers HBF samples in 2026 and an accelerator customer announces device testing in early 2027, HBF moves first from specification to product. A named zHBM customer or production date would reopen the speed contest. For now, HBF leads on schedule while zHBM remains a performance concept.