4 ms·

Considering how much money people drop on gaming PCs these days, why is that sorta hardware advantage not available on PC as well? You'd think there would be a

by Drakim 3y ago

Considering how much money people drop on gaming PCs these days, why is that sorta hardware advantage not available on PC as well? You'd think there would be a market for it, if offers such an advantage.

constantly 3y ago
People may drop more money on PC than the average console person drops on their co sole, but the number of people doing this to their PC is vanishingly small. So per capita it’s better, but in aggregate spending a bunch of dev time on such a small market isn’t smart. Put succinctly: even if this hardware advantage were available on PC, game devs wouldn’t optimize for it.
nemothekid 3y ago
As I understand it, you can only get that advantage by having an SoC, where the CPU, RAM and GPU are all one unit (like Apple's M1 chip). If you want a system where to upgrade the RAM, you have to upgrade the whole system - well that already exists, it's called consoles.
- ericmay 3y ago
  Do you think Apple is working on finding a way to be easier to develop for because of their SOC strategy? Like, hey we are just like a game console, easier than PC!
- kllrnohj 3y ago
  The RAM being "on unit" has no impact here. It just requires the CPU & GPU to share an IOMMU & share the memory bus. Whether or not that RAM is soldered or not is totally irrelevant. Laptops have had unified memory for a ~decade as a result, but once you go to a discreet GPU instead of integrated then there's no feasible way to share the IOMMU & memory bus, at which point it's not practical. And then up-gradable graphics or just 400W GPUs like the 3090/4090 behemoths are incompatible without extreme cost.
  - paulmd 3y ago
    > The RAM being "on unit" has no impact here. It just requires the CPU & GPU to share an IOMMU & share the memory bus. Whether or not that RAM is soldered or not is totally irrelevant. Sure, but to get the same memory bandwidth as a M1 Max you'd need 8-channel DDR5 memory, so your socket would look like Epyc and consume 50W+ just for data movement. GDDR allows moving much more data than regular DDR (or even LPDDR) but both GDDR and LPDDR5X require much tighter signaling requirements that preclude socketed memory. And that means much higher watts-per-byte-transferred for socketed memory, in practice. Like dozens and dozens of watts higher for a typical configuration. > Laptops have had unified memory for a ~decade as a result There's unified memory, and there's unified memory. Fusion HSA and Intel iGPUs have always required you to designate memory as either graphics or CPU side so that it knows which cache controller to run it through, and this has implications for memory and cache visibility. ctrl-f 'garlic' and 'onion': https://www.eurogamer.net/digitalfoundry-how-the-crew-was-ported-to-playstation-4 https://www.eurogamer.net/digitalfoundry-how-the-crew-was-po... https://forum.beyond3d.com/threads/amd-kaveri-apu-features-the-onion-bus-like-the-playstation-4.54966/ https://forum.beyond3d.com/threads/amd-kaveri-apu-features-t... Money shot: "One issue we had was that we had some of our shaders allocated in Garlic but the constant writing code actually had to read something from the shaders to understand what it was meant to be writing - and because that was in Garlic memory, that was a very slow read because it's not going through the CPU caches. That was one issue we had to sort out early on, making sure that everything is split into the correct memory regions otherwise that can really slow you down." In contrast, the newer stuff like PS5/Xbox Series family and Apple Silicon family has true zero-copy where it's all just flat memory and all writes are immediately visible to all clients of the memory controller. You can interleave GPU and CPU and Neural Engine tasks without forcing explicit sync into memory to flush the cache/bus. I think most people would say that's meaningfully different. Both are "unified" but one is more unified than the other.
dontlaugh 3y ago
Mostly because PCs are modular on axes that aren’t that useful today but also can’t easily be changed. You can’t have the same memory with the same latency used for both CPUs and GPUs because GPUs don’t have DIMM and CPUs do, for example. You also can’t guarantee a fast enough SSD to be able to delay loading assets until they are needed.
CDSlice 3y ago
It only works because on modern consoles the RAM is soldered directly on the board and so the data doesn’t have to travel across PCIe lanes. Doing this on a normal PC would have horrific performance. ARM Macs do have this architecture but you pay for it with not being able to replace any individual components for upgrades or even repairs.
- blibble 3y ago
  you can have sockets/slots for your RAM without it needing to be on the PCIe bus any PC would be an example
zokier 3y ago
AMD HSA (which afaik current-gen consoles are based on) is available on PC APUs also. Although it is pretty much completely undocumented mess on software side.
reportgunner 3y ago
Gaming consoles have worse performance than PCs.

by Drakim 3y ago

constantly 3y ago

People may drop more money on PC than the average console person drops on their co sole, but the number of people doing this to their PC is vanishingly small. So per capita it’s better, but in aggregate spending a bunch of dev time on such a small market isn’t smart. Put succinctly: even if this hardware advantage were available on PC, game devs wouldn’t optimize for it.

nemothekid 3y ago

As I understand it, you can only get that advantage by having an SoC, where the CPU, RAM and GPU are all one unit (like Apple's M1 chip). If you want a system where to upgrade the RAM, you have to upgrade the whole system - well that already exists, it's called consoles.

ericmay 3y ago
Do you think Apple is working on finding a way to be easier to develop for because of their SOC strategy? Like, hey we are just like a game console, easier than PC!
kllrnohj 3y ago
The RAM being "on unit" has no impact here. It just requires the CPU & GPU to share an IOMMU & share the memory bus. Whether or not that RAM is soldered or not is totally irrelevant. Laptops have had unified memory for a ~decade as a result, but once you go to a discreet GPU instead of integrated then there's no feasible way to share the IOMMU & memory bus, at which point it's not practical. And then up-gradable graphics or just 400W GPUs like the 3090/4090 behemoths are incompatible without extreme cost.
- paulmd 3y ago
  > The RAM being "on unit" has no impact here. It just requires the CPU & GPU to share an IOMMU & share the memory bus. Whether or not that RAM is soldered or not is totally irrelevant. Sure, but to get the same memory bandwidth as a M1 Max you'd need 8-channel DDR5 memory, so your socket would look like Epyc and consume 50W+ just for data movement. GDDR allows moving much more data than regular DDR (or even LPDDR) but both GDDR and LPDDR5X require much tighter signaling requirements that preclude socketed memory. And that means much higher watts-per-byte-transferred for socketed memory, in practice. Like dozens and dozens of watts higher for a typical configuration. > Laptops have had unified memory for a ~decade as a result There's unified memory, and there's unified memory. Fusion HSA and Intel iGPUs have always required you to designate memory as either graphics or CPU side so that it knows which cache controller to run it through, and this has implications for memory and cache visibility. ctrl-f 'garlic' and 'onion': https://www.eurogamer.net/digitalfoundry-how-the-crew-was-ported-to-playstation-4 https://www.eurogamer.net/digitalfoundry-how-the-crew-was-po... https://forum.beyond3d.com/threads/amd-kaveri-apu-features-the-onion-bus-like-the-playstation-4.54966/ https://forum.beyond3d.com/threads/amd-kaveri-apu-features-t... Money shot: "One issue we had was that we had some of our shaders allocated in Garlic but the constant writing code actually had to read something from the shaders to understand what it was meant to be writing - and because that was in Garlic memory, that was a very slow read because it's not going through the CPU caches. That was one issue we had to sort out early on, making sure that everything is split into the correct memory regions otherwise that can really slow you down." In contrast, the newer stuff like PS5/Xbox Series family and Apple Silicon family has true zero-copy where it's all just flat memory and all writes are immediately visible to all clients of the memory controller. You can interleave GPU and CPU and Neural Engine tasks without forcing explicit sync into memory to flush the cache/bus. I think most people would say that's meaningfully different. Both are "unified" but one is more unified than the other.

dontlaugh 3y ago

Mostly because PCs are modular on axes that aren’t that useful today but also can’t easily be changed. You can’t have the same memory with the same latency used for both CPUs and GPUs because GPUs don’t have DIMM and CPUs do, for example. You also can’t guarantee a fast enough SSD to be able to delay loading assets until they are needed.

CDSlice 3y ago

It only works because on modern consoles the RAM is soldered directly on the board and so the data doesn’t have to travel across PCIe lanes. Doing this on a normal PC would have horrific performance. ARM Macs do have this architecture but you pay for it with not being able to replace any individual components for upgrades or even repairs.

blibble 3y ago
you can have sockets/slots for your RAM without it needing to be on the PCIe bus any PC would be an example

zokier 3y ago

AMD HSA (which afaik current-gen consoles are based on) is available on PC APUs also. Although it is pretty much completely undocumented mess on software side.

reportgunner 3y ago

Gaming consoles have worse performance than PCs.