Nvidia 3080 Ti & Ampere FULL Leak: 7nm EUV & 18Gbps needed to Defeat AMD RDNA 2

[ENTERPRISE]

 

This is starting to get me ''Amped''

[J7SC_Orion]

 

This is starting to get me ''Amped''

 

 

...things are definitely heating up in the 'pre-Ampere' release arena - apart from NVidia's briefing event tomorrow which may be more focused on next-gen pro cards. Also, similar info from TweakTown a few days ago.

 

If all this is true, NVidia seems to re-establish a clearer line between their consumer and pro models: GA100 *Specs* - 8192 CUDA cores, 1024 Tensor Cores, 256 RT Cores. That one might tidy me over until Hopper

[ENTERPRISE]

 

 

...things are definitely heating up in the 'pre-Ampere' release arena - apart from NVidia's briefing event tomorrow which may be more focused on next-gen pro cards. Also, similar info from TweakTown a few days ago.

 

If all this is true, NVidia seems to re-establish a clearer line between their consumer and pro models: GA100 *Specs* - 8192 CUDA cores, 1024 Tensor Cores, 256 RT Cores. That one might tidy me over until Hopper

 

So you would got for the Titan RTX (Or whatever they choose to call it for the Ampere series) ?

[J7SC_Orion]

 

So you would got for the Titan RTX (Or whatever they choose to call it for the Ampere series) ?

 

 

...depends on pricing. But with RTX, 2080 Ti has 4352 Cuda / 544 tensor versus Titan RTX, Quadro RTX8000 et al that have 4608 Cuda / 576 tensor (Tesla V100 is a different beast / architecture). So the much 'cheaper' (lol) 2080 Ti RTX is already very close to the full-fat RTX, and with some custom PCBs can match the Titan RTX etc via higher clocks. It only really depends on whether the extra Titan RTX VRAM and select other features are needed by the user.

 

If the Ampere rumours pan out, the new top-spec G100 has 8192 Cuda and 1024 tensor vs. G102 (3080 Ti ?) at 5376 Cuda and (? tensor)...a much more dramatic difference. But Nvidia will undoubtedly charge for this extra oomph. One wildcard re. purchasing decisions is what AMD's Big Navi and Intel's Xe end up bringing to the table and at what price...another one is 'Hopper' lurking

Edited May 13, 2020 by J7SC_Orion

[Bastiaan_NL]

That is some interesting stuff!

Wonder what the 30xx range will do folding wise, but for now I'm happy with the 2080ti.

[ENTERPRISE]

 

 

...depends on pricing. But with RTX, 2080 Ti has 4352 Cuda / 544 tensor versus Titan RTX, Quadro RTX8000 et al that have 4608 Cuda / 576 tensor (Tesla V100 is a different beast / architecture). So the much 'cheaper' (lol) 2080 Ti RTX is already very close to the full-fat RTX, and with some custom PCBs can match the Titan RTX etc via higher clocks. It only really depends on whether the extra Titan RTX VRAM and select other features are needed by the user.

 

If the Ampere rumours pan out, the new top-spec G100 has 8192 Cuda and 1024 tensor vs. G102 (3080 Ti ?) at 5376 Cuda and (? tensor)...a much more dramatic difference. But Nvidia will undoubtedly charge for this extra oomph. One wildcard re. purchasing decisions is what AMD's Big Navi and Intel's Xe end up bringing to the table and at what price...another one is 'Hopper' lurking

 

I think that this round in order to protect the Titan RTX, they really need to distance that SKU from the Ti range, otherwise aside from the additional VRAM, the Titan RTX is just lost. It will be interesting to see what RDNA 2 brings, Intel Xe wont be able to compete IMHO with AMD's and Nvidia's next offering. Not saying they will not be impressive, but last time I looked they were not even really aiming for the gaming market all that much.

[J7SC_Orion]

Well, fresh from Jensen Huang's kitchen per this mornings event: Not much on Ampere 'gaming' yet, but as expected, on data-center versions. Top Ampere data-center version seems to have 6912 Cudas (not 8192 as per earlier leaks) - for now, anyway ;-)

 

That's still a lot of room for various lower-level versions, such as Titan and Ti. It seems the whole generation will run on the same base architecture (thus including 'GeForce'), but obviously w/ different Cuda, tensor counts and VRAM types (HBM2 vs GDDR) and amounts.

 

 

 

[ENTERPRISE]

It is certainly an impressive lineup, lets hope the same for the consumer hardware !