We are hearing more and more about NVIDIA's next-gen Ampere and Hopper GPUs, with more rumors flying in the last week than the last few months. Now? We have news that NVIDIA is bursting right into 5nm with Hopper.
According to a new report from DigiTimes, NVIDIA has "already pre-booked TSMC's 5nm production capacity in 2021" while "also in discussion with Samsung for smaller volume orders". The reason? To "prevent AMD from getting any bigger, NVIDIA has decided to catch up, even leapfrog AMD, in adopting TSMC's 7nm and 5nm EUV nodes".
Now remember, there's no one going on the record here so take this all with a grain of salt. There are some big power moves going on, I'm sure we can all see that by now -- but this is all just so very interesting.
'RetiredEngineer' on Twitter tweeted out the DigiTimes report, nothing that NVIDIA "previously underestimated the effect of AMD embracing TSMC, and made some errors in its own plans for migrating to advanced nodes. Moving to Samsung and wavering in its TSMC strategy, only to be go back to TSMC later, NVIDIA was unable to secure enough 7nm capacity, resulting in AMD grabbing the limelight, eroding NVIDIA's brand value, and increase its own market share for GPUs".
The tweet continues that "in order to prevent AMD from getting any bigger, NVIDIA has decided to catch up, even leapfrog AMD" by adopting TSMC's 7nm and upcoming 5nm EUV nodes.
(New) NVIDIA Ampere GA100 Specs
- 8192 CUDA cores @ 1750MHz boost
- 1024 Tensor Cores
- 256 RT Cores
- Unknown amount of GDDR6 @ 16Gbps
- Unknown TDP
- 7nm
More reading on NVIDIA Ampere:
NVIDIA Ampere GPUs
- GA102 - 84 SMs / 5376 CUDA cores / 12GB GDDR6 / 384-bit bus - 40% faster than RTX 2080 Ti
- GA103 - 60 SMs / 3840 CUDA cores / 10GB GDDR6 / 320-bit bus - 10% faster than RTX 2080 Ti
- GA104 - 48 SMs / 3072 CUDA cores / 8GB GDDR6 / 256-bit bus - 5% slower than RTX 2080 Ti
Older NVIDIA Ampere GA100 Specs
- 8192 CUDA cores @ 2GHz (2.2GHz boost)
- 1024 Tensor Cores
- 130 RT Cores
- 48GB of HBM2e memory @ 1.2GHz
- 300W TDP
- TSMC 7nm+
- 36 TFLOPs peak output
I've already written about rumors that NVIDIA's next-gen Ampere GPU architecture would be up to 75% faster than current-gen GPUs such as the Turing architecture, right after rumors that Ampere would offer 50% more performance at half the power of Turing. This is pretty crazy stuff right there.
Not only that, but we've got some rumored specs on the purported GeForce RTX 3080 and GeForce RTX 3070 graphics cards, which will both be powered by NVIDIA's new Ampere GPU architecture.
We've already heard that Ampere would offer 50% more performance at half the power of Turing, which sent the hairs on my neck standing up. Better yet, you can read about the leaked specs on the purported Ampere-based GeForce RTX 3080 and GeForce RTX 3070 right here.
GA103 (GeForce RTX 3080)
- 10/20GB GDDR6
- 320-bit memory interface
- 60 SMs
- 3480 CUDA cores
GA104 (GeForce RTX 3070)
- 8/16GB GDDR6
- 256-bit memory interface
- 48 SMs
- 3072 CUDA cores
AMD also just revealed its next-gen RDNA 2 / Navi 2X plans, something we referred to as Big Navi until just recently. Those specs are monstrous, but they pale in comparison to what NVIDIA has planned for its flagship HPC destined Ampere GA100 beast.
AMD RDNA 2 / Navi 2X / Big Navi Specs
- 7nm node (confirmed)
- RDNA 2 architecture (confirmed)
- 80 compute units
- 5120 stream processors
- 17.5 TFLOPs compute performance
- Hardware ray tracing support (confirmed)