The first generation of NVIDIA & AMD GPUs to utilize GDDR7 memory are most likely going to stick with 16 Gb dies for 2 GB VRAM capacities.
First Wave of GDDR7 Memory-Powered GPUs From NVIDIA & AMD To Use 16 Gb Dies, Offering 2 GB Capacity & 32 Gbps Pin Speeds Per Module
The news comes from kopite7kimi, who disclosed that the first iteration of GPUs from NVIDIA and AMD with GDDR7 memory will come with a 16 Gb density, meaning that the minimum capacity we will see for the next-generation standard is 2 GB of VRAM. For comparison, the GDDR6 era kicked off with 8 Gb dies offering 1 GB VRAM capacity per module. We have reiterated this in an earlier post, where JEDEC defined the GDDR7 standard as a whole.
We will use 16Gbit first.
— kopite7kimi (@kopite7kimi) March 7, 2024
For a quick recap on GDDR7, up until now, we have seen solutions by Samsung & Micron, with both aiming for 32 Gbps pin speeds and 16 Gb dies as well. Micron has also teased 24 Gb & 24 Gb+ dies for the future. With 24 Gb modules, you can get a maximum of 3 GB VRAM per die, and with 24 Gb+ (32 Gb or higher), you can get 4 GB+ VRAM capacities. These higher density dies are likely not going to come this early as GDDR7 has yet to roll out and it will take time for the standard to hit maturity level to allow such capabilities but eventually, by 2026-2028, we will see the rollout of higher VRAM offerings for GPUs.
Following is the maximum bandwidth and memory capacity the 16 Gb dies + 32 Gbps pin speeds would offer across multiple bus configurations:
- 128-bit @ 32 Gbps: 512 GB/s (8 GB)
- 192-bit @ 32 Gbps: 768 GB/s (12 GB)
- 256-bit @ 32 Gbps: 1024 GB/s (16 GB)
- 320-bit @ 32 Gbps: 1280 GB/s (20 GB)
- 384-bit @ 32 Gbps: 1536 GB/s (24 GB)
- 512-bit @ 32 Gbps: 2048 GB/s (32 GB)
2 Gb memory densities would mean that the first iteration of GDDR7 GPUs from NVIDIA and AMD will likely stick with the same memory capacities that we got on current-gen chips. For those who were expecting higher capacities, well we would have to wait for future revisions or refreshes. It's still a possibility that each tier gets a boost such as the RTX 4060 Ti which has a 128-bit bus and is available in both 8 and 16 GB VRAM. The higher VRAM variant can become the standard for next-gen offerings.
Now, in terms of who will get the massive adoption in the markets, it isn't difficult to say right now. Still, Micron has a great track record with NVIDIA, as the firm co-developed the GDDR5X standard with Micron with its GeForce 10 series cards & currently uses the fastest GDDR6X modules for its latest RTX 40 graphics cards.
Micron will likely continue to ship optimized versions of the next-gen G7 memory standard with NVIDIA GPUs. Samsung on the other hand is more inclined with AMD and that's where we are going to see the bulk of their memory dies being offered.
Regarding when we could expect GDDR7 products to drop into the markets, NVIDIA's GeForce RTX 50 "Blackwell" GPUs are hinted for a late 2024 launch though nothing is set in stone right now. At Team Red's camp, AMD's next-gen RDNA 4 GPUs are also expected to debut this year but it is so far not confirmed whether we will see the introduction of GDDR7 on such GPUs or will GPU makers stick with GDDR6 (X) for one more round. So, with this, we aren't far from the launches, which suggests we might see more information regarding the standard unveiling in the upcoming period.
GDDR Graphics Memory Evolution:
| GRAPHICS MEMORY | GDDR7 | GDDR6X | GDDR6 | GDDR5X |
|---|---|---|---|---|
| Workload | Gaming / AI | Gaming / AI | Gaming / AI | Gaming |
| Platform (Example) | GeForce RTX 5090 | GeForce RTX 4090 | GeForce RTX 2080 Ti | GeForce GTX 1080 Ti |
| Die Capacity (Gb) | 16-64 | 8-32 | 8-32 | 8-16 |
| Number of Placements | 12 | 12 | 12 | 12 |
| Gb/s/pin | 28-48 | 19-24 | 14-16 | 11.4 |
| GB/s/placement | 112-192 | 76-96 | 56-64 | 45 |
| GB/s/system | 1536-2304 | 912-1152 | 672-768 | 547 |
| Configuration (Example) | 384 IO (12pcs x 32 IO package)? | 384 IO (12pcs x 32 IO package) | 384 IO (12pcs x 32 IO package) | 384 IO (12pcs x 32 IO package) |
| Frame Buffer of Typical System | 24 GB (16 Gb) 36 GB (24 Gb) | 24 GB | 12GB | 12GB |
| Module Package | 266 (BGA) | 180 (BGA) | 180 (BGA) | 190 (BGA) |
| Average Device Power (pJ/bit) | TBD | 7.25 | 7.5 | 8.0 |
| Typical IO Channel | PCB (P2P SM) | PCB (P2P SM) | PCB (P2P SM) | PCB (P2P SM) |
News Source: Kopite7kimi
About the author: Muhammad Zuhair is a hardware and technology reporter for Wccftech, specializing in the semiconductor industry and the complex interplay between technology, manufacturing, and geopolitics. His coverage focuses on the corporate strategies and technological roadmaps of industry giants like TSMC, NVIDIA, Samsung, and Intel. Zuhair's expertise lies in deconstructing complex topics such as fabrication nodes (e.g., 2nm process), the economic impact of policies like the CHIPS Act, and the strategic development of AI infrastructure from NVIDIA, AMD and Intel.
Follow Wccftech on Google to get more of our news coverage in your feeds.
