This is a rumor at this time about upcoming Nvidia playing cards. These new leaks come from kopte7kimi And speaking about the schematic diagram of the construction of the new era of greens. Picture of a file block diagram AD102 “Ada Lovelace” GPU It will permit us to drop ourselves on the efficiency of the upcoming RTX 40.
RTX 40: An important spec sheet (if true)
For starters, the GPU Ada Lovelace AD102 It will home as much as 12 GPCs (graphics processing clusters). This is a 70% enhance from in GA102 (bigger than present vary) which accommodates solely 7 GPC. Each GPU will include 6 TPCs and a pair of SMs, which is the identical configuration as the present chip. Each SM (Multiprocessor Stream) can have 4 sub-cores, which is additionally the identical as the GA102 GPU. The actual change is the FP32 and INT32 kernel configuration. Each sub middle will include 128 FP32 items, however mixed FP32 + INT32 items will enhance to 192 items. This is as a result of the FP32 modules don’t share the identical subcentre as the IN32 modules. 128 FP32 cores are separated from 64 INT32 cores.
(*40*) A schematic picture of the RTX 40 GPU from Kopte7kimiThe cache must be one other space the place NVIDIA has outgrown present Ampere GPUs. Ada Lovelace GPUs can have 192KB of L1 cache per SM, a 50% enhance over the Ampere. This quantities to a complete of 4.5MB of L1 cache on the top-of-the-line AD102 GPU. The L2 cache will probably be elevated to 96MB, a quantity that is commonly talked about in a number of leaks. This is practically 16 occasions extra in comparison with the Ampere GPU which solely hosts 6MB of L2 cache. The cache will probably be shared on the GPU.
If the leaks are appropriate, we have now an exponential enhance in L2 cache, which will increase to a complete 96 MB to me’ M 102 . Regarding ROPs, there might have been twice as many modules on this structure, 32 from GPC To be precise, giving us a complete 384 OMR For a possible RTX 4090 versus 112 for an RTX 3090… on paper it’s brutal.
But after this orgy of technical information, what features can we actually anticipate?
Obviously it’s nonetheless too early to get a precise concept but when these are confirmed, the technical sheet reveals an enormous distinction in comparison with the Ampere. To summarize:
- X2 GPC (in comparison with amps)
- 50% extra cores (in comparison with amps)
- 50% extra L1 cache (in comparison with amps)
- 16 occasions the L2 cache (in comparison with amps)
- X2 ROP (in comparison with amps)
- 4th era motor and three cores RT
But what can we anticipate when it comes to precise efficiency?
It’s very tough as a result of we’re lacking a key piece of information: the working frequency.
If we speculate a bit about it, we will introduce ourselves to a energy in FP32 from 90 TFLOPS, greater than twice that of the present GA102. But with TFLOPS we will even have surprises. If they provide an concept of uncooked efficiency, they are going to by no means permit prejudgement of ends in ‘everyday’ use. Leaked adverts from x2 to x2.2 in comparison with the RTX 30… There will clearly be features, they usually appear huge. But to decide next, we must wait slightly longer.