The upcoming Nvidia RTX 4090 gave a base clock frequency of 2235MHz, a boost of 2520MHz and an “actual” clock frequency of 2750MHz. Thus, the base clock frequency is 60 percent higher than the current RTX 3090, which runs at 1,395MHz as standard. Boost hour is 49 percent higher.
This refers to the well-known Kopite7kimi. With 16,384 CUDA cores and a “physical” clock frequency of 2750MHz, computational performance of 90 TFlops can be achieved in single-precision workloads. That’s nearly double the amount of the RTX 3090.
Kopite7Kimi also offers the specification for the GPUs used in the RTX 4080 and RTX 4070. They both use different GPUs, AD103 and AD104 respectively. The RTX 4080 appears to have a 256-bit memory bus and 16 GB gddr6x vram, clocked at the same speed as the RTX 4090. On the other hand, the RTX 4070 has a 160-bit bus and 10 GB gddr6 vram.
Nvidia will likely announce the RTX 4000 series in September or October of this year. The RTX 4060 is expected to be announced during CES 2023.
| model | GeForce RTX 4090 | GeForce RTX 4080 | GeForce RTX 4070 |
|---|---|---|---|
| general engineering | Ada (TSMC N4) | Ada (TSMC N4) | Ada (TSMC N4) |
| GPU | M 102 – 300 | M 103 – 300 | M 104 – 275 |
| Text message | 128 | 80 | 56 |
| CUDA cores | 16,384 | 10.24 | 7.168 |
| base clock | 2.235MHz | nnb | nnb |
| increase the clock | 2,520MHz | nnb | nnb |
| Memory speed | 24 GB gddr6x | 18 Gbps | 18 Gbps |
| Memory bus | 384 bit | 256 bit | 160 bit |
| Bandwidth | 1,008 GB/sec | 576 GB/sec | 360 GB/sec |
| TBP | 450 watts | 420 W | 300 watts |
source:
Twitter (@Kopite7Kimi)
« Previous Post Last Next »