Rurouni
Veteran
What I want is for Ryzen to be able to have 32GB 3200 memory! Right now, because of those known limitation, if you want 32GB of memory, you need to settle for lower speed since I don't think you can occupy all 4 slots if you want more than 2666 and I don't think they sell 16GB single rank RAM (at least I can't find it in my country).
Has it changed? Like can you have 32GB 3200 running on Ryzen? What is currently the highest memory speed that people can get with 32GB of memory?
Sent from my LG-E988 using Tapatalk
Has it changed? Like can you have 32GB 3200 running on Ryzen? What is currently the highest memory speed that people can get with 32GB of memory?
Sent from my LG-E988 using Tapatalk
We've seen that Ryzen reacts favorably to higher RAM speeds in some usage scenarios, more so than other CPUs. I'd also want to game on my PC so sqeezing more single-thread peformance from it is desirable. I upgrade rarely, waiting 4+ years between changes. It's not unreasonable that I want my builds to be optimized as much as possible for various workloads
Yes someone over at Anandtech managed to hit 3200 I believe, but not sure what ram/board combination. Likely CH6 and FlareX.
D
Deleted member 13524
Guest
Why are 3200MT/s memory needed for Ryzen?
I'd get the need for ~50GB/s if these were APUs with a 1TFLOPs GPU, but for CPU cores only, where does this bandwidth make any practical difference?
The ability for Ryzen to shift data over the infinity fabric is tied to the memory controller, thus faster memory is better.
The jump from 2133 -> 3200 can get you +~10 % perf (!) in a few cases ( e.g. http://www.hardware.fr/articles/958-5/jeux-battlefield-1-project-cars.html )
This is very interesting, the 6900k always gains going from dual -> quad channel so theoretically they should continue to scale until you hit a complete GPU bottleneck. Naples with quad channel and more cores will be interesting
I'm currently running 3466CL14 and hoping to get 3600+ after some bios updates, we'll see
Edit: 6900k DC read/write/copy metrics are all lower than Ryzen, but the 6900k also has lower latency. This is probably something they'll have to work on but read/write/copy speeds seem very good!
Last edited:
First sign of the APUs,
http://wccftech.com/amd-raven-ridge-ryzen-apu-vega-gpu-leak/
4/8; 3.0/3.3GHz; 704 Vega cores. And even L3 cache.
http://wccftech.com/amd-raven-ridge-ryzen-apu-vega-gpu-leak/
4/8; 3.0/3.3GHz; 704 Vega cores. And even L3 cache.
Yes, the GPU is pretty close. Xbox One is 12 CUs @ 853Mhz. One extra CU extra and 53MHz lead. But GCN2 -> GCN5 is huge difference for geometry performance. Then there's delta color compression, ROP caches under L2, tiled rasterizer, instruction prefetch, etc. And for future games: double rate fp16 + SDWA, DPP (cross lane ops in SM 6.0 and Vulkan). I would expect Raven Ridge GPU to be slightly faster than Xbox One.
Quad core (8 thread) Ryzen has almost double clocks and almost 2x IPC. Should be roughly 2x faster than Xbox One CPU in multithreaded tasks, and up to 4x faster in single threaded tasks.
Bandwidth is a question mark however. Dual channel DDR4 at 2666 MHz = 42.7 GB/s. That's slower than Xbox One's quad channel 2133 MHz DDR3 main memory (68.3 GB/s) and there's no EDRAM either. DCC and bigger caches help, but can't do miracles.
Even at 45W that would be pretty sweet
Quad core Ryzen is going to be almost twice as fast as eight core Jaguar plowing through integer spaghetti, but about equal on AVX code, right ?
Cheers
Ryzen has FMA, Jaguar doesn't. That doubles theoretical FLOP per clock (when code uses FMA instructions). Also according to Agner Fog's analysis Ryzen can do (128b) 4xFMA + 4xFADD in 3 cycles. That's 12 FLOP in 3 cycles = 4 FLOP / cycle. Equal FLOP to 2xFMA per cycle, but doesn't require two dependent muls and adds. Ryzen also decodes "fastpath double" ops in single cycle. This includes AVX and AVX2 instructions that will be split to two 128 bit halves. Jaguar doesn't have AVX2, so no full width integer ops and no gather. Also Jaguar doesn't have BMI2 instruction set, which includes pdep/pexp instructions (very nice for morton decode/encode).
Jaguar is also often bound by issue rate when executing SSE/AVX code. Only two instructions in total can be decoded per cycle. Address calculation instructions and loop overhead is directly eating AVX cycles. Ryzen is wider, and can run mixed int + AVX better. Ryzen also has uop cache for loops, which can IIRC process 5 instructions per clock (up from default rate of 4, which is 2x Jaguar already). Also as said previously, Ryzen decodes both fastpath single and fastpath double instructions in single cycle, further increasing it's lead over Jaguar. Haven't yet looked at Ryzen AVX instruction latencies. Jaguar had awesome low latencies compared to Bulldozer/Piledriver.
Yes, the GPU is pretty close. Xbox One is 12 CUs @ 853Mhz. One extra CU extra and 53MHz lead. But GCN2 -> GCN5 is huge difference for geometry performance. Then there's delta color compression, ROP caches under L2, tiled rasterizer, instruction prefetch, etc. And for future games: double rate fp16 + SDWA, DPP (cross lane ops in SM 6.0 and Vulkan). I would expect Raven Ridge GPU to be slightly faster than Xbox One.
Quad core (8 thread) Ryzen has almost double clocks and almost 2x IPC. Should be roughly 2x faster than Xbox One CPU in multithreaded tasks, and up to 4x faster in single threaded tasks.
Bandwidth is a question mark however. Dual channel DDR4 at 2666 MHz = 42.7 GB/s. That's slower than Xbox One's quad channel 2133 MHz DDR3 main memory (68.3 GB/s) and there's no EDRAM either. DCC and bigger caches help, but can't do miracles.
Even at 45W that would be pretty sweet
I'm pretty sure that's a cut-down SKU and the full Raven Ridge features 12 CUs or 768 SPs (per Canard PC's Ryzen leak some time ago); clock speeds are harder to guess but there's no reason that desktop variants (supposedly only available in 2018, alas) shouldn't be able to reach 1.4GHz and above. Mobile variants might be around 1GHz or so. Bandwidth, however, remains an issue indeed.
Similar threads
