There is a rumour going around today regarding coming Nvidia cards. The leaks are coming from the account kopte7kimi. They also have a sketch diagram of the structure for the next generation of greens. Image of a block of data Diagram AD102 “Ada Lovelace” GPU This will let us test the capabilities of the coming RTX 40.
RTX 40: An excellent specification sheet (if the claims are valid)
The first is that the GPU Ada Lovelace AD102 will contain more than twelve GPCs (graphics processing clusters). This is a 70percent increase from the GA102 (more extensive than the current range), with only 7 GPC. Each GPU will comprise 6 TPCs and two as that is the exact configuration of the existing chip. Each SM (Multiprocessor stream) will comprise four sub-cores. This is the same as GA102. GA102 GPU. The only difference lies in that of the FP32 along with the INT32, the kernel’s configuration. Each sub-centre will comprise 128 units of FP32; however, the combined FP32 and teams for INT32 will rise to an additional 192 units. This is because FP32 modules don’t share the same subcenter as the IN32 modules. 128 FP32 cores are distinct from the 64 INT32 cores.
The cache will be another area where NVIDIA has outstripped the current Ampere GPUs. Ada Lovelace GPUs will have L1 caches of 192KB for each SM, which is a 50% more than the Ampere. It’s an overall 4.5MB in the L1 cache for the top of the line AD102 GPU. The L2 cache will increase to 96MB. This is a figure that is often reported in various leaks. This is almost 16 times larger than the Ampere GPU that only has six MB L2 cache. It will share the cache with the GPU.
If leaks prove accurate, the result is an exponential growth in the L2 cache, which is now 96 millibytes to my the M 102. There could be two times as many modules in the GPC architecture concerning ROPs. Thirty-two for GPC, In fact, and giving us three84 OMR In the case of an RTX 4090 as opposed to 112 OMR for the RTX 3090… on paper, it’s pretty brutal.
However, what kind of improvements can we expect after this assortment of technical information?
It’s too early to have an exact picture; however, the technical sheet suggests the vast difference in the Ampere if they are confirmed. To summarize:
- The X2 GPC (compared to amps)
- 50 more cores (compared amps)
- 50 per cent greater the L1 cache (compared to amps)
- 16 times the cache of L2 (compared with amps)
- The X2 ROP (compared to amps)
- 4th generation motor with three cores of RT
What can we anticipate regarding actual performance?
It’s a challenge because we lack a crucial aspect of the data, Operating frequency.
If we think a little about it, we may reveal a potential strength within FP32 of the 90-TFLOPS that is over twice the amount of GA102. However, with TFLOPS, we could also be surprised. If they can indicate their raw performance, they won’t permit prejudgment of the results in the context of everyday usage. The leaked ads range from x2 to x2.2 compared to the 30 RTX… It is clear that there are bound to be significant improvements, and they appear to be huge. To make a final decision shortly, we’ll need to wait longer.