Micron has printed its newest roadmap which particulars its upcoming reminiscence applied sciences together with GDDR7 for next-gen NVIDIA GPUs.
Micron To Be A Main Provider of DRAM For Subsequent-Gen NVIDIA GPUs, GDDR7 In 2H 2024 & GDDR7+ In 1H 2026
NVIDIA has been deciding on Micron’s GDDR options for its graphics playing cards for fairly a while. The inexperienced group first co-developed the GDDR5X commonplace with Micron with its GeForce 10 sequence playing cards & presently makes use of the quickest GDDR6X modules for its newest RTX 40 graphics playing cards. The corporate additionally makes use of commonplace GDDR6 chips in its budget-friendly & mainstream designs. Now, Micron has launched its newest roadmap which supplies us a glimpse of what to anticipate sooner or later.
Beginning with the preliminary GDDR7 Graphics DRAM from Micron, we’re taking a look at 16Gb & 24Gb DRAM dies. A 24 Gb module will give us 3 GB VRAM capability per die versus the two GB VRAM capability on a single 16Gb die. The reminiscence modules will even function at larger clock speeds of 32 Gbps, providing as much as 128 GB/s bandwidth per module. A lowly 128-bit resolution will have the ability to output 512 GB/s of bandwidth. For comparability, the GeForce RTX 4060 with a 128-bit bus GDDR6 interface outputs 272 GB/s of bandwidth so GDDR7 presents almost twice as a lot bandwidth.
The primary technology GDDR7 reminiscence modules are anticipated by the mid of 2024 which additionally hints that we are able to count on NVIDIA’s next-generation lineup across the identical time. The corporate plans on refreshing its GeForce RTX 40 portfolio with the brand new “SUPER” household with an unveiling deliberate for CES 2024. These may additionally profit from the upper speeds provided by the present GDDR6X DRAM (as much as 24 Gbps).
Following is the bandwidth the 32 Gbps pin speeds would provide throughout a number of bus configurations:
- 128-bit @ 32 Gbps: 512 GB/s
- 192-bit @ 32 Gbps: 768 GB/s
- 256-bit @ 32 Gbps: 1024 GB/s
- 320-bit @ 32 Gbps: 1280 GB/s
- 384-bit @ 32 Gbps: 1536 GB/s
- 512-bit @ 32 Gbps: 2048 GB/s
Shifting ahead, Micron will additional improve its GDDR7 reminiscence with even larger speeds and extra capacities. This can be made attainable with 24Gb+ reminiscence dies so 32 Gb options are a risk. That will allow 4 GB of VRAM capacities on a single DRAM module. Plus, the pin speeds will even get a lift to 36 Gbps, providing as much as 144 GB/s bandwidth per module. The reminiscence is anticipated by early 2026 and will even launch beneath the “GDDR7X” guise.
- 128-bit @ 36 Gbps: 576 GB/s
- 192-bit @ 36 Gbps: 846 GB/s
- 256-bit @ 36 Gbps: 1152 GB/s
- 320-bit @ 36 Gbps: 1440 GB/s
- 384-bit @ 36 Gbps: 1728 GB/s
- 512-bit @ 36 Gbps: 2304 GB/s
The VRAM capacities utilizing 24 Gb modules will look as follows:
- 128-bit 24 Gb (4 ICs) = 12 GB VRAM
- 192-bit 24 Gb (6 ICs) = 18 GB VRAM
- 256-bit 24 Gb (8 ICs) = 24 GB VRAM
- 320-bit 24 Gb (10 ICs) = 30 GB VRAM
- 384-bit 24 Gb (12 ICs) = 36 GB VRAM
- 512-bit 24 Gb (16 ICs) = 48 GB VRAM
The VRAM capacities utilizing 32 Gb modules will look as follows:
- 128-bit 32 Gb (4 ICs) = 16 GB VRAM
- 192-bit 32 Gb (6 ICs) = 24 GB VRAM
- 256-bit 32 Gb (8 ICs) = 32 GB VRAM
- 320-bit 32 Gb(10 ICs) = 40 GB VRAM
- 384-bit 32 Gb (12 ICs) = 48 GB VRAM
- 512-bit 32 Gb (16 ICs) = 64 GB VRAM
Micron is within the race with Samsung and SK hynix to supply the world’s first GDDR7 reminiscence. Experiences have acknowledged that Samsung has already sampled its GDDR7 reminiscence to NVIDIA for its next-gen GPUs & SK hynix can also be subsequent in line. Though NVIDIA has talked about 2025 as a proper launch of its next-gen gaming lineup, they may fast-track it to the tip of 2024 which can mark two years because the RTX 40 introduction. In the meantime, AMD’s next-gen RDNA 4 GPUs are additionally in line to make the most of the next-gen DRAM commonplace once they debut subsequent 12 months.
Apart from GDDR, the Micron roadmap additionally lists down HBM4 in 2025 and 2027, respectively. The HBM4 reminiscence is anticipated to function 12/16-Hello dies with as much as 36/48 GB capacities and provide 1.5 TB/s bandwidth whereas HBM4E will provide as much as 48/64 GB capacities with bandwidth of as much as 2.0 TB/s.
The corporate can also be anticipating LPCAMM2 with as much as 128 GB capacities and 8533 MT/s speeds in 2024 adopted by a complicated model that includes 192 GB+ capacities and 9600 MT/s speeds. The standard DDR reminiscence will even see a lift with as much as 256 GB+ modules and DDR5-12800 MT/s speeds by 2025.
GDDR Graphics Reminiscence Evolution:
| GRAPHICS MEMORY | GDDR5X | GDDR6 | GDDR6X | GDDR7 |
|---|---|---|---|---|
| Workload | Gaming | Gaming / AI | Gaming / AI | Gaming / AI |
| Platform (Instance) | GeForce GTX 1080 Ti | GeForce RTX 2080 Ti | GeForce RTX 4090 | GeForce RTX 5090? |
| Variety of Placements | 12 | 12 | 12 | 12? |
| Gb/s/pin | 11.4 | 14-16 | 19-24 | 32-36 |
| GB/s/placement | 45 | 56-64 | 76-96 | 128-144 |
| GB/s/system | 547 | 672-768 | 912-1152 | 1536-1728 |
| Configuration (Instance) | 384 IO (12pcs x 32 IO bundle) | 384 IO (12pcs x 32 IO bundle) | 384 IO (12pcs x 32 IO bundle) | 384 IO (12pcs x 32 IO bundle)? |
| Body Buffer of Typical System | 12GB | 12GB | 24 GB | 24 GB? |
| Common Gadget Energy (pJ/bit) | 8.0 | 7.5 | 7.25 | TBD |
| Typical IO Channel | PCB (P2P SM) | PCB (P2P SM) | PCB (P2P SM) | PCB (P2P SM) |
