Summary
- Samsung's GDDR7 is set to be showcased in February 2024, with transfer rates up to 37Gbps, while also reducing power consumption.
- Compared to GDDR6 and GDDR6X, GDDR7 is expected to deliver noticeable performance enhancements and use less power.
- GDDR7's widespread availability coincides with the next generation of GPUs from Nvidia and AMD, potentially resolving the issue of limited memory bandwidth in high-performance computing products.
Samsung has been teasing its next-generation Graphics Double Data Rate 7 (GDDR7) memory since July 2023. The highest-end GDDR7 chips are expected to hit transfer rates as high as 37Gbps per pin, which means dramatically increased data throughput. Additionally, GDDR7 should offer these higher speeds at a reduced power consumption per transfer. Now, the company is expected to have the GDDR7 on display at a tech conference in February 2024.
Samsung’s GDDR7 will make an appearance at the 2024 IEEE International Solid-State Circuit Conference, where it is being billed as a “16Gb 37 Gbps GDDR7 DRAM with PAM3-Optimized TRX Equalization and ZQ Calibration” (via Tech Radar). However, few details have been publicized aside from what was already previously released by Samsung.
What we know so far is that GDDR7 will reach its promised 37Gbps bitrate using PAM3 and NRZ signaling rather than PAM4 and NRZ currently used in GDDR6X. By going with PAM3, which only features three signal levels rather than four, Samsung is able to increase signal-to-noise ratio and increase the transmission rate, thus leading to greater memory bandwidth. GDDR7 also improves efficiency by using four different read clock modes, with two independent command rates.
In a high-end GPU, GDDR7 would also mean significant increases to total memory bandwidth. In a theoretical RTX 5090 with a 384-bit memory bus, 37Gbps memory chips would equate to a whopping 1.78TB/s of memory bandwidth. To put that into perspective, that's 78% higher than an RTX 4090, which has 1.0TB/s of memory bandwidth.
|
GPU |
Memory Type |
Memory Bus Width |
Transfer Rate |
Memory Bandwidth |
|
Theoretical RTX 5090 |
GDDR7 |
384-bit |
37Gbps |
1.78TB/s (1776GB/s) |
|
RTX 4090 |
GDDR6X |
384-bit |
21Gbps |
1.0TB/s (1008GB/s) |
|
RTX 4080 Super |
GDDR6X |
256-bit |
23Gbps |
736GB/s |
|
RTX 4070 Ti Super |
GDDR6X |
256-bit |
21Gbps |
672GB/s |
|
RTX 4070 Super |
GDDR6X |
192-bit |
21Gbps |
504GB/s |
|
RTX 3090 |
GDDR6X |
384-bit |
19.5Gbps |
936GB/s |
|
RTX 3080 |
GDDR6X |
320-bit |
19Gbps |
760GB/s |
|
Radeon 7900XTX |
GDDR6 |
384-bit |
20Gbps |
960GB/s |
|
Radeon 7900XT |
GDDR6 |
320-bit |
20Gbps |
800GB/s |
Generally speaking, memory bandwidth has been limited on modern high-end graphics cards such as Nvidia's 4000-series, given the dramatic increase in compute power without a commensurate increase in memory bandwidth. With the arrival of the GDDR7 at the end of 2024 or the beginning of 2025, this pain point could finally be resolved leading to greater performance when gaming at high resolutions such as 4K and beyond. GDDR7's release is expected to coincide with the next generation of GPUs from Nvidia and AMD, but exact dates for next-gen GPUs have not yet been determined. When GDDR7 launches, it will most likely be featured in both Nvidia Blackwell (5000-series) and AMD RDNA4 (8000-series).
