NVidia GeForce RTX 30x0 cards

Daidraco said:

If the 30x0 and 40x0 are just flooded on the market - then they'll just stretch out the release date for the next generation. AMD sounds like its in a similar boat, but not as bad - so it doesnt sound like competition is going to force them to push forward with the following series as fast.

Click to expand...

I believe that depends on multiple factors: If AMD is on par or beats Nvidia slightly with the next generation, they will most likely not have another "Titan" flagship, and will have to scramble to bring something powerful to the market. They want that halo effect of being "performance king."

AMD on the other hand seems to be quite comfortable in the "good performance for a slightly lower price" at the ankles of Nvidia because of their console SoC sales. Also, If they can't sell their GPU wafers they can just make PS5 or XBox chips with them, or CPUs for the server market. Nvidia on the other hand can do not much else with their booked TSMC wafers.

Mist said:

This is an extremely relevant video. EDIT: The title is hyperbolic but the video is extremely interesting. You have to actually watch the whole thing.

Click to expand...

There are a few misconceptions in that video: ARM is not an "open architecture", it's basically just a license for their cores. You can either build your own SoC with everything else around the cores yourself, or license other parts of ARM, or even let them put something together. But ARM still controls the cores.

One example: Apple used the ARM cores and put a PowerVR GPU for the A10 in the 2017 iPad. Yes, that's the PowerVR that was in the Sega Dreamcast. Since then Apple has invested in their own GPU cores.

Also, I would not fellate that much about ARM: Yes, it's more modern, but it's still an architecture from 1985. When talking about "wasted silicon no one needs anymore like x86" and "keeping the architecture alive by stacking extension after extension on top", ARM has those parts, too. For example, just like MIPS, ARM is Bi-Endianned.

What most "tech journalists" fail to realize: Modern CPU architectures are now completely decoupled from the actual instruction set: The executed code is interpreted by a software layer on the CPU and translated into internal micro-operations. Ever heard of the Spectre and Meltdown exploits, and how that triggered a plethora of microcode updates? By then you should've realized that your processor is just another computer. Yes, ARM is a load-store RISC architecture, while x86 is an old school CISC register-memory architecture, but internally the x86 probably works the same as load-store RISC.

You could say the x86 instruction set is basically the "lingua franca" of the computing world, as most code is compiled to that instruction set. So why would people choose ARM, another closed instruction set, as the next instruction set to compile their software to? Good question! This is also why ARM is starting to sweat right now, because just like Linux some people came up with an open source ISA for a modern RISC architecture: RISC-V. Which is starting to gain traction. With x86 being walled off between Intel and AMD, and ARM not being much better, a lot of companies are looking for alternatives. From the Europeans to the Chinese, even Nvidia are testing RISC-V as replacement for their Falcon controller on the GeForce cards.

On another note, you could probably get comparable performance and power usage to the M1/M2 chips from Apple with an x86 architecture, if you would do the same as Apple did:

• Use the latest TSMC node (5 nm, which Apple has booked exclusively for a while)
• Put all I/O and the GPU on one big die, leave out parts you don't need like PCI-Express
• Put all the RAM directly next to the CPU, but make it non-expansible in the process

With all this you have drawbacks: Big dies and modern processes have low yields, making the result expensive (Apple doesn't care, it was always an expensive premium brand), and once baked you can't change the RAM in the chip (Apple doesn't care, they want users to buy new hardware instead of upgrading).

Mist said:

How I understood the video is that components are going to have to get put into a single large SOC, or much more integrated motherboard designs, or something. DRAM and VRAM being separate is almost certainly going to have to go away.

Maybe >64GB of DRAM will be integrated into the CPU and it will have a direct bus to the interface for the GPU.

Click to expand...

This is complicated, because Video RAM needs to be constantly read to ... well, send the current in the Video RAM image to the screen. Do the math yourself: 3840 x 2160 pixels x 4 bytes per pixel (10 bits per color) x 120 times per second.

This gets more complicated when you are using something called "compositor" (which is in most modern operating systems), where different things on screen each have their own framebuffer in Video RAM with their content, and are stacked "on the fly" to make things like transparency and shadows easily work.

So Samsung invented special, (read: more expensive) memory with high bandwidth and short bursts. DDR5 started with 4.8 Gbits/s per pin, while GDDR6X is at up to 24 Gbits/s per pin.

How do APUs get away with "normal" memory? Well, they don't. If you look at the benchmarks, the 5600G is nearly on par with a 5600 without GPU when you don't do anything GPU-heavy, and especially if the computing tasks can fit in the L1 to L3 caches. But as soon as you start a game or anything GPU-heavy, the CPU performance also drops. This is not only because of thermal limits, but because the RAM bandwidth is the limiting factor.

So, is everything fucked up in PC country? Not quite. For one, you could stack GDDR6X instead of "normal" memory. It's just more expensive, and output

The other thing is to move things closer to each other, so that signaling isn't as complicated and power hungry, and can be faster with lower latency. This means ditching PCI express as a bus and moving to other interconnects.

If you paid attention to your favorite companies, you'd know the somewhat recent acquisitions in the market:

• Nvidia bought Mellanox for $6.9 billion, who are the major player in InfiniBand-related things, which is a low-latency interconnect for data centers
• Intel also bought into InfiniBand by buying QLogic's InfiniBand assets for $125 million, and bought another company related to that. But that name eludes me at the moment.
• AMD bought Xilinx in an estimated $50 billion all-stock merger: Xilinx not only makes FPGAs, but also networking things like "network on chip" for SoCs, and has pioneered something called "stacked silicon interconnect" by a silicon interposer. Which they started shipping in 2011.

So, I don't know where the future is heading, but I guess the GPU will be moving closer to the CPU. This is why Intel is also went into the GPU market: To beat Nvidia into submission on that front, not only for gaming-related stuff, but also for data center acceleration. Probably in cooperation with AMD.

If you want to get into the technical details, I recommend "What every programmer should know about memory".

Break said:

There could be a paradigm shift in how CPUs process instructions. Ye olde "if" statements and other conditional code are the bane of parallelism but if some genius ever discovers a way to scale single-threaded conditional code across multi-cores seamlessly, we could see a jump in computing performance like we've never seen before in history. Imagine a CPU can natively run single threaded code across 16 or 32 cores, seamlessly. GPUs have been doing this forever because drawing the screen is just grunt work but CPUs are still stuck in the medieval times.

Click to expand...

We already worked that out. This is what GPUs are, and all those funny extensions for the CPU instructions, starting with MMX in 1997. It's called SIMD (single instruction, multiple data) and is featured heavily in modern CPU designs, e.g. AVX and AVX512 computers.

Thing is, CPUs can't beat GPUs on that front, because GPUs are designed for SIMD execution from the ground up ... but not much else. Not all computing problems can be "divided and conquered" through parallelism, though. The nasty ones (e.g. in number theory) are recursive. Computer science still has problems proving if a given algorithm even ends in finite time. You should look up "computability theory", "computational complexity theory" and "P versus NP" if you want to go into that rabbit hole.

On another note, there is research going on to improve parallelism. For example "lock-free data structures", which are or aren't more performant, depending on your application. Intel wanted to improve parallel transaction speed with their TSX-NI extensions, but like many things from Intel, that cratered and was never seen or heard of again.