3 ms·
Nvidia should be very afraid of AMD. AMDs mastery of chiplet designs is in stark contrast to Nvidias reliance on extremely large and complex dies and pretty la
by jpgvm 2y ago
Nvidia should be very afraid of AMD.
AMDs mastery of chiplet designs is in stark contrast to Nvidias reliance on extremely large and complex dies and pretty lack luster forays into multi-die in the past.
If Nvidia doesn't come up with ways to combat the inherent cost and yield advantages of chiplets they could get quite easily blown away on $/perf.
GPGPU didn't have enough money behind it to justify fully exploiting all hardware vs just going with whatever was easiest (Nvidia+CUDA) and eating the cost but that has changed substantially.
If you can buy AMD at a 10%+ $/perf advantage when you are talking about laying down $100M+ for accelerators on even a moderate sized cluster you can be damn sure the software will get tuned to take advantage of it.
- Rinzler89 2y agoWhat are you talking about?! Nvidia has been using chiplets and silicone interconnects in the data center ML accelerators way before AMD got in the game. It's the same TSMC tech that both Nvidia and AMD use, it's not something proprietary exclusive to one or another since they're both fables. On the contrary, AMD has to be afraid since they have no answer to Nvidia's 800GB/s data center interconnects they got from their aquisition of Mellanox, which enables Nvidia to scale.
- timschmidt 2y agoAMD purchased Xilinx not long ago, and as a result has access to their high speed interconnects, including at least one capable of supporting 800G networking: https://www.xilinx.com/products/technology/high-speed-serial.html https://www.xilinx.com/products/technology/high-speed-serial...
- jpgvm 2y agoI'm as close as you get to an old school Infiniband fanboi. Don't think that Infiniband is somehow an Nvidia moat, if anything because they ate up the last somewhat independent vendor (Mellanox) it's likely it will be replaced by some variant of what is now being called Ultra Ethernet which for the most part incorporates ideas from Infiniband into the Ethernet PHY layer along with bolting into libfabric at the userland level. The people with the brains behind it include the Omni-Path folk (which is descendent from QLogic and SilverStorm before that) which have extensive experience in Infiniband and are now free from the shackles of Intel. However this has absolutely nothing to do with chiplet designs or multi-die architectures (which Nvidia has done before with their dual GPU stuff from the GTX 900 series days etc). The fact of the matter is that Nvidia has no experience whatsoever with chiplets or advanced packaging. AMD has on the other hand absolutely tons from yes, chiplets but also HBM, 3d-vcache etc all of which are ultimately advanced packaging techniques that all build on shared background of expertise. Nvidia needs to rapidly catch up in this area or their future is a lot less certain. This is putting aside what is happening with Tenstorrent/Etched/Groq, etc which threaten to eat up the vast majority of the non-training workloads associated with the current crop of AI/LLM architectures.
- teaearlgraycold 2y agoNvidia seems familiar with HBM. It’s on a lot of their products.
- Rinzler89 2y agoWhat? Nvidia HAS used HBM in several products except only in DC side and not consumer side like AMD