NVIDIA enjoys almost absolute supremacy in the PC gaming market. AMD has not progressed since the disappointing release of the RDNA 3 range. If the latest rumors are to be believed, Team Radeon has completely abandoned the enthusiast market, leaving the future GeForce RTX 5090 and 5080 without rivals. Speaking of the RTX 5090, NVIDIA has reportedly started testing Samsung's GDDR7 graphics memory for its next-generation (gaming) offerings.
According to the datasheet, GDDR7 memory is up to 40% faster than existing GDDR6X modules, while reducing power consumption by 20%. This result is obtained thanks to PAM3 modulation, compared to PAM4 for GDDR6X memory. Compared to the move from GDDR5 to GDDR6, this is a lighter upgrade, increasing the number of available states from 2 to 3 (50% more than GDDR6).
GDDR7 GDDR6
Like GDDR6 and GDDR6X, GDDR7 also offers quad-rate memory bursts, although signaling is worse than GDDR6X. It compensates for this by offering a peak bandwidth of 1.5 TB/s on a 384-bit bus, compared to 1.1 TB for GDDR6 and GDDR6X.
Our GDDR7 DRAM will help improve user experience in areas that require exceptional graphics performance, such as workstations, PCs and gaming consoles, and is expected to expand to future applications such as AI , high performance computing (HPC) and automotive vehicles. Next-generation graphics DRAM will be released to the market based on industry demand and we plan to maintain our leadership in this area.
Yongcheol Bae, Executive Vice President, Memory Product Planning, Samsung Electronics
Unfortunately, the absolute power consumption of GDDR7 will be higher than that of GDDR6, with a 6-7% increase for 32 Gbps modules (compared to 24 Gbps for GDDR6). The chip density also remains unchanged (16 or 2 GB), at least at the time of publication. Regardless, GDDR7 is expected to be a hit for GPU makers due to its high memory throughput.
NVIDIA RTX 50 series GPUs are expected to arrive in the second half of 2024 or early 2025. Leveraging TSMC's 3nm process node, they will deliver the next major advancement in rendering performance. The chips remain monolithic, with higher density and greater surface area. The RTX 5090 is expected to have 24 to 32 GB of GDDR7 memory paired with a 384-bit bus.
In theory, this should achieve a bandwidth of 1.5 TB/s and an effective bandwidth of over 2 TB/s thanks to L2 cache. That said, don't hold your breath. We're still a long way from Blackwell's launch, and things could change between now and then.
Source: Revegnus.