yep simultaneous/constant back and forth read/write will kill the bandwidth dramatically.
I believe the 140 GB/s on that slide was sort of average case of available bandwidth while the game code is going at it. Certainly not worse case scenario
Could next gen consoles focus mainly on CPU?
- Thread starter Theeoo
- Start date
I believe the 140 GB/s on that slide was sort of average case of available bandwidth while the game code is going at it. Certainly not worse case scenario
Yeah, I was vaguely aware of that. My point is irrelevant. There's a real-world performance and then real-world with CPU impact.
Achieving utilization of 70-76% of theoretical max seems doable for some GPUs (random textures to avoid compression enchantment):
https://techreport.com/review/28513/amd-radeon-r9-fury-x-graphics-card-reviewed/4
https://techreport.com/review/30328/amd-radeon-rx-480-graphics-card-reviewed/5
https://techreport.com/review/28513/amd-radeon-r9-fury-x-graphics-card-reviewed/4
https://techreport.com/review/30328/amd-radeon-rx-480-graphics-card-reviewed/5
No.
The 176GB/s is just a theoretical max of 5.5GHz x 32 bytes. The 140GB is the actual maximum bandwidth achieved by the SOC (by the GPU). And it's not even the same gigabytes. The 176 GB is 176 x 10^9, the 140GB is 140 x 2^30 (150 x 10^9).
Yes, the impact on bandwidth is disproportional to the CPU's bandwidth usage. Why is that ? Don't know, but could be:
1. The CPU has smaller, less effective transactions.
2. It might also be CPU memory transactions have priority over GPU memory transactions and thus leaves less room for reordering for open pages.
3. Poorer locality which might interfere with 2.)
That is a result of a memory system built for a GPU with a CPU cluster attached as an afterthought. Will that be so in the future? I doubt it.
Cheers
For some reason this never occurred to me until you pointed this out.
So the SoC has no way to get to the theoretical maximum here? this really just put things into perspective for me
Wait..what?
I can't ever seem to commit ram tech to long term memory since this is only place I ever discuss it and we only delve into the minutiae of the tech every now and again.
But if the PS4 is rated at 176 GBps (10^9) or 5.5 GHz x 32 bytes then as long as the PS4 can reach 5.5 GHz and can move 32 bytes in any one cycle then the actual theoretical max bandwidth can be achieved.
Or am I missing something? Are bits lost every cycle?
Measured over a single (or few) cycles, yes, measured over a second, no.
The PS4's GPU has 18 CUs, each CU has 4 16-wide SIMD execution units. Each of these exec units can issue a load or store per cycle. That's 72 memory transactions per cycle. The memory system consists of 8 1GB x32 modules, each with 16 banks. That means you have a total of 128 banks. If two memory transactions hit the same bank, one will have to wait for the other to complete. The memory controller tries to reorder memory transactions to resolve bank conflicts (as well as optimize for open pages in a given bank), but there will always be conflicts that induce stalls.
That means you will never ever reach the nameplate bandwidth outside of specially engineered Mickey Mouse benchmarks.
I'm guessing CPU transactions aren't reordered with GPU transactions; The GPU can handle memory latency quite well, a CPU, not so much. Raw latency of GDDR5 is on the order of 50-something ns, but the GPU normally sees ~200ns latency for memory operations, - because of the heavy buffering going on. The CPU memory transactions throws a spanner into the bank/open page optimization strategy and the memory bandwidth utilization drops.
HBM2 has 8 channels per stack, two pseudo channels per channel and 16 banks per pseudo channel, or 256 banks per stack,- and normally multiple stacks. With more banks you have fewer conflicts and higher utilization.
Cheers
D
Deleted member 13524
Guest
HBM2 isn't reliable?
and performance?
Vega and Fiji weren't the fastest cards in the world. Might not have to do something with their memory, but still I haven't seen a fast card with HBM 1 or 2 memory
Well, vega and fiji have a problem bringing it's theoretical power to the table.
I think in a console environment the heavy use of highly optimised async would see much greater utilisation of the CUs and much greater throughput. Would possibly better justify the use of such expensive memory... though I still don't think we'll see HBM in a console.
Megadrive1988
Veteran
I think the GPU will be Navi/NextGen, if not just slightly custom Navi.
phoenix_chipset
Regular
The cpus will be more of a focus than they were last time, anyway.
If it were up to me I would put the emphasis on cpu power and memory bandwidth rather than just a huge gpu, the bare minimum cpu and merely adequate bandwidth. I would like 8 full fat zen cores, not zen lite and either gddr6 with eDRAM (96-128mb?) or gddr6 on a massive bus. Gpu could be 8-10 tf and that'd be more than ok with adaptive resolution.
I'd like 100% perfect hardware BC (beefed up jaguar/amd pitcairn) but not at the cost to new games. I don't think they could beef up jaguar enough for it to not hinder the rest of the system.
If it were up to me I would put the emphasis on cpu power and memory bandwidth rather than just a huge gpu, the bare minimum cpu and merely adequate bandwidth. I would like 8 full fat zen cores, not zen lite and either gddr6 with eDRAM (96-128mb?) or gddr6 on a massive bus. Gpu could be 8-10 tf and that'd be more than ok with adaptive resolution.
I'd like 100% perfect hardware BC (beefed up jaguar/amd pitcairn) but not at the cost to new games. I don't think they could beef up jaguar enough for it to not hinder the rest of the system.
Last edited:
