Well, it looks similar Micron has officially commenced the wide accumulation of its next-gen GDDR6X representation chips offering speeds of up to 24 Gbps.
Micron's 24 Gbps GDDR6X Memory Enter Mass Production, Coming To NVIDIA's Next-Gen GPUs Later This Year
Micron announced earlier this twelvemonth that they were readying their marque caller 24 Gbps GDDR6X representation modules for aboriginal graphics cards. The announcement came astatine a clip erstwhile NVIDIA's GeForce RTX 3090 Ti had conscionable entered the marketplace for which Micron had produced 16 Gb DRAM modules rated astatine up to 21 Gbps speeds.
Now, Micron is going to connection adjacent faster pin speeds of 24 Gbps successful a 16 Gb density which means that erstwhile again, we volition beryllium looking astatine 2 GB VRAM capacities and up to 24 GB capacities crossed a 384-bit autobus interface. Just similar the GeForce RTX 3090 Ti which featured each of its representation modules connected the beforehand of the PCB, the next-gen GeForce RTX 4090 would diagnostic its 12 GDDR6X modules connected the front, allowing for much cooling compared to solutions that diagnostic representation modules connected the backmost of the PCB specified arsenic the RTX 3090 (Non-Ti).
Although the afloat 24 Gbps speeds aren't expected to beryllium utilized by the NVIDIA GeForce RTX 4090 (rumored to connection up to 21 Gbps dice speeds), the 24 Gbps GDDR6X representation modules tin connection up to 1.152 TB/s of bandwidth which is an summation of 14% versus the 1.008 TB/s offered connected existent 21 Gbps GDDR6 representation modules.
Following are the bandwidth numbers you tin expect with a 24 Gbps DRAM solution:
- 512-bit solution - 1.5 TB/s
- 384-bit solution - 1.1 TB/s
- 320-bit solution - 960 GB/s
- 256-bit solution - 768 GB/s
- 192-bit solution - 576 GB/s
- 128-bit solution - 384 GB/s
- 92-bit solution - 276 GB/s
- 64-bit solution - 192 GB/s
The NVIDIA GeForce RTX 4090 is going to beryllium amongst the archetypal next-gen graphics paper to utilize the latest GDDR6X representation modules erstwhile it launches aboriginal this year. Given that these chips are apt going to beryllium downclocked to deed the powerfulness targets, determination volition beryllium a immense overclocking imaginable for enthusiasts to instrumentality vantage of.
AMD connected the different manus is apt to beryllium connected its partner, Samsung, for its next-gen RDNA 3 lineup. Samsung is besides working connected a 24 Gbps GDDR6 representation module which is expected to deed wide accumulation soon.
Workload | Graphics | GraphicsAI Inference Accelerator | GraphicsAI Inference Accelerator |
Platform (Example) | TitanX | Titan RTX RX5700 XT |
GeForce® RTX™ 3090Ti GeForce® RTX™ 3080Ti |
Number of Placements | 12 | 12 | 12 |
Gb/s/pin | 11.4 | 14-16 | 19-24 |
GB/s/placement | 45 | 56-64 | 76-96 |
GB/s/system | 547 | 672-768 | 912-1152 |
Configuration (Example) | 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 | 12GB | 12GB | 24GB |
Average Device Power (pJ/bit) | 8.0 | 7.5 | 7.25 |
Typical IO Channel | PCB (P2P SM) |
PCB (P2P SM) |
PCB (P2P SM) |
News Source: Harukaze5719