The hacked Nvidia data appears to contain a great deal of information about the next generation of video cards, which is expected to be called Ada (Lovelace). For example, earlier today we wrote about the number of SMS and cuda kernels associated with this. It now appears that RTX 4000 cards will also have a large amount of L2 cache on board.
For example, Twitterer XinoAssassin with a screenshot of Nvidia’s alleged code shows Ada GPUs serving 16MB of L2 cache per 64-bit memory bus, compared to 512KB per 32-bit in the Ampere case. Memory buses for various AD GPUs have also been leaked, allowing the calculation of the amount of L2 cache.
| Ada Lovelace (‘RTX 4000) | |||
| GPU | Cuda nuclei | bus | L2- Cache |
| M 102 | 18,432 | 384 bit | 96 MB |
| M 103 | 10,752 | 256 bit | 64 MB |
| M 104 | 7.680 | 192 bit | 48 megabytes |
| M 106 | 4.608 | 128 bit | 32 MB |
| M 107 | 3.072 | 128 bit | 32 MB |
| Ampere (RTX 3000) | |||
| GPU | Cuda nuclei | bus | L2- Cache |
| GA102 | 10,752 | 384 bit | 6 megabytes |
| GA103 | 7.680 | 256 bit | 4 megabytes |
| GA104 | 6.144 | 256 bit | 4 megabytes |
| GA106 | 3.840 | 192 bit | 3 megabytes |
| GA107 | 2560 | 128 bit | 2 megabytes |
Key specifications of Ada and Ampere.
The AD102 flagship model is said to have 96MB of L2 cache, which is 16 times more than the 6MB of Ampere’s flagship model GA102. The AD106 and AD107 bottom chipset will also provide a much larger L2 cache than the Ampere at 32MB. How Nvidia realizes this leap is not clear, but technology like AMD’s Infinity Cache is likely to be used. Combined with a greater amount of cuda cores, it appears that Ada at least offers a solid performance compared to its predecessor.
sources:
Zeno Killer (Twitter)† video card† Kopite7kimi (Twitter)
« Previous Post Last Next »
“Lifelong zombie fanatic. Hardcore web practitioner. Thinker. Music expert. Unapologetic pop culture scholar.”