Infinity Cache is just a huge L2 128MB on gpu cache. I’m guessing it’s L2 because it being L1 would be crazy. This make it so the gpu keeps having stuff to work on while more is continuously streamed off the gpu board ram. In comparison the 3080 has 128kb of L1 cache and 5MB of L2 cache. So if the 3080 would run out of things to do from the cache in 1 ns it would take AMD 25.6 ns. Nvidia runs out of cache so fast it says I need more and has to hurry the more by having a huge bus to keep the GPU from running out of things to do. Whatever is needed to help the GPU have a non stop Q of things to do is good.
The cache can be used to overcome bandwidth and to store ray tracing or do anything. Being on GPU it would be more of a machine code level thing. At its core all a gpu or cpu does is add. It does do a bit more than that now, but lots of adding still. In that machine code could be instructions to ray trace or do anything really. I also think the GPU cache stores the answers calculated from the GPU. If it fills too fast it has to stop because it has no room to store the answer.
I understand the function of a cache and the benefits of a large cache and also it can be used to store whatever data the situation requires.
What I don’t know is whether the BVH is built from the camera perspective every frame and whether this would benefit from a cache. How could BVH data cache be used from frame to frame?
I understand objects and be added and removed from a BVH tree without needing the tree to be rebuilt so it looks like there is an opportunity to cache these data. So one might imagine that a ray tracer could be sufficiently sped up by only having to trace new objects or new lights etc. Maybe it could use a similar technique to Long GOP compression where only changes are updated and there is a regular ‘reference’ frame update.
The rumoured size of the cache is huge but I have no idea if it would be big enough to store BVH data.
The question is only able to be asked because of the hardware, rumoured size of Infinity Cache is a step change compared to previous designs and is possibly a game changer in many areas.
Infinity Cache is a key component of bringing Radeon SSG workflows to the masses with DirectStorage, the effect on content creation could be profound as your PCIe 4 attached NVMe drive is effectively GPU memory.
Do I sense another story where NVIDIA released faulty products?
I still recall their BS with the GTX 970 VRAM story.
Remember why Apple never wanted to us NVIDIA again.
No, they’re not releasing faulty products as far as I know. The 8nm Samsung node just isn’t as mature as TSMC’s 7nm node so there’s more failed parts per wafer.
AMD have just demoed Big Navi with the 5000 series chips and it’s actually quite unimpressive unless they’re incredibly power efficient or they’re giving them away. If this was Biggest Navi it loses to the 3080 in gaming and does not represent disruptive 4k gaming.
I’ll be interested to see creative benchmarks but I was expecting much more and this probably rules out 20GB 3080s anytime soon, if at all.
Yup. Big navi just seems 2 or 3 fps more than the 3080 in the games shown in 4K, and is unknown what card produces it (no raytracing, no extras). We’ll have to wait until the 28th to know more about these videocards and their capabilities.
Lisa Su showed the 3 fan GPU and called it Big Navi. I think these scores are from the top SKU so therefore these absolute frame numbers are disappointing as they will have been set with the world’s best gaming CPU. There is a chance there’s a big reveal on the 28th that these number were not set on the top SKU, Jebaited again!
However, we don’t know pricing and we don’t know the power draw both of which could make Navi a much better choice for gamers. What’s 10% in absolute performance if the Navi GPUs use 1/3 less power compared to some of the outrageous power hogs the AIBs have (or plan to) release? A win in performance/watt is a win, for me anyway. I’m absolutely put off Ampere by its inefficiency.
I hope Navi is good at compute but these are GPUs that are aimed at gaming first and foremost unlike Ampere which is a compute card that’s repurposed for gaming. I hope AMD don’t allow creative workflows to fall down the gap between RDNA2 and CDNA. I hope there is enough compute performance in RDNA2 for rendering and video editing or there’s an affordable cutdown CDNA GPU coming.
Let’s also see what surprises AMD have with architectural advances that might be applicable for creative workflows like DirectStorage support.
It seems that AMD is preparing for the acquisition of Xilinx which is strong for FPGAs this precludes that at the moment it has no strength in the world of machine learning, AI, and consequently Raytracing etc …
From this I would venture that Big Navi will not have advanced machine learning and raytracing functions, and will only be competitive in terms of performance …
But it is inevitable that this gap will soon be filled, AMD is forced to do this if it wants to save itself for how things will go in the future.
Since a few years ago Intel had already taken this path by acquiring ALTERA other strong company with FPGAs, this precludes that Intel is preparing to come up with something revolutionary both in the field of gpu and in the field of cpu and other… Even more because recently it officially released the ONEAPI libraries …
This also makes us think about why NVIDIA has decided to acquire ARM while paying for it at a very high price …
I would venture that NVIDIA risks being wiped out by the X86 market or otherwise becoming marginal … and therefore is preparing to become aggressive in the competitive world of ARM, in mobile, and also in the desktop and server antagonistic to the X86 world
The Future is still to be written… few words for good connoisseurs to understand…
Yes ok, this type of configuration is particularly noisy and probably also energy-intensive (I presume)
He’s a monster, a workstation hammer …
But how much less time will you render?
It would be interesting this type of comparison; energy consumption divided by rendering time.
The best thing is not to buy* these space heaters and wait for nVidia to refresh them on TSMC’s 7nm node to meet the demand. There will be an abundance of buyer’s remorse among early adopters when a much more efficient refresh arrives sooner than many will expect.
nirved (blender.org)