Samsung’s GDDR7 and Shinebolt HBM3E Memory Announcement
During its annual Memory Tech Day in San Jose, Samsung shared updates on its next-generation memory products, including the development of GDDR7 memory for future graphics cards. This new memory type is anticipated to be a part of the upcoming Blackwell RTX 50-series and RDNA 4 RX 8000-series graphics cards. A significant highlight of the announcement was the target for GDDR7 memory to achieve 50% less stand-by power compared to its predecessors, which is expected to contribute to reduced idle power consumption, benefiting scenarios such as multi-monitor setups and video playback.
Specifications of GDDR7 Memory
The upcoming GDDR7 memory is set to operate at a lower voltage of 1.2V, compared to the 1.35V of its predecessors, GDDR6 and GDDR6X. This reduction in operating voltage is a step towards power efficiency. With anticipated speeds reaching up to 32Gbps, GDDR7 could potentially deliver a total bandwidth of 1TB/s over a 256-bit bus. If implemented with a 384-bit bus, the bandwidth could see an increase of 50%, further enhancing the memory’s performance capabilities.
| Memory Type | Operating Voltage |
|---|---|
| GDDR7 | 1.2V |
| GDDR6/GDDR6X | 1.35V |
PAM3 Signaling in GDDR7
Samsung’s GDDR7 memory will utilize PAM3 (pulse-amplitude modulation) signaling, which is a shift from the PAM4 used in GDDR6X. PAM3 signaling is favored for its ease of implementation, which makes it a more viable option for consumer-grade graphics cards. This choice is particularly relevant for mid-range GPUs, which could benefit from faster memory but have been limited by cost considerations.
- Lower average bit-error rates compared to PAM4
- Easier to implement in consumer products
- More cost-effective for mid-range GPUs
The potential for a future GDDR7X standard remains open, which could further enhance the performance for high-end graphics cards. Such a standard could theoretically enable a hypothetical RTX 5090 Ti with a 512-bit bus to achieve bandwidths exceeding 2TB/s, although the cost of such cards would likely be substantial.
Market Availability and Expectations for GDDR7
As the industry anticipates the arrival of GDDR7 equipped cards, the expectation is that they are at least a year away from being launched. Initially, it is predicted that GDDR7 will be featured predominantly in high-end graphics cards due to its newness and the associated production costs. Over time, as manufacturing scales up and costs decrease, it may become more prevalent across a wider range of graphics cards.
In the meantime, next-gen entry-level cards are likely to continue using GDDR6 memory, which remains a cost-effective solution. The transition to GDDR7 for these more affordable options will probably occur once the technology becomes more accessible and economically feasible for a broader market segment.
Next-Generation HBM3E Memory Discussion by Samsung
Samsung also discussed its next-generation HBM3E memory, known as Shinebolt, which boasts impressive performance capabilities. Shinebolt memory can achieve up to 9.8Gb/s per pin, which translates to a bandwidth of 1.225GB/s for a single HBM3E stack. This is a significant increase from the 819Gb/s offered by HBM3.
When considering multiple stacks, the bandwidth potential is even more remarkable. For example, a flagship GPU equipped with multiple HBM3E stacks could theoretically reach an astonishing 7.35TB/s of bandwidth. To put this into perspective, a reference RTX 4090 currently delivers 1,008GB/s, which is already considered quite fast.
Despite these impressive performance figures, the inclusion of HBM3E memory in gaming graphics cards is unlikely due to its high cost and complexity. Therefore, while HBM3E represents a significant advancement in memory technology, it may be reserved for applications beyond the consumer gaming market.
