MadFanboy Forum

PS5

This SSD has been analyzed thoroughly, and there is no special "black technology". Everything is the result of a special adaptation for the game console scene. In theory, the cost is not high, and it may even The SSD cost of the Xbox Series X is even lower.

That's right, more than twice the speed, but in theory the cost is still low, why?

Because the SSD's main control is actually very cheap, the main function is to differentiate pricing for manufacturers. In fact, the high-end main control takes the price of the goods and the low-end main control is not much different. However, good horses and good saddles, high-end SSDs on PCs often have to be paired with a small DDR cache to store the address lookup table (LUT), and that's it.

If you want to use a pile of 64GB of flash memory to build a 1TB hard drive, you can use 16 channels, each channel put a piece of flash memory, so the fastest speed, 16x64GB = 1024GB, and then we can be regarded as 16X.

The PS5 is such a design, 12 channels, each channel 64GB, 12x64 = 768GB. Converted from 1024 to 1000 is 825G. The speed is regarded as 12X.

XSX has only 4 channels, but each channel can stack multiple chips, so 4x4x64GB = 1024GB, but only 4X speed. So if you look at the original speed, 5.5GB / s vs 2.4GB / s, which is about 2.3 times the gap. Why not 3 times? Because 64GB of flash memory does not run on a single channel, multiple chips will use the channel bandwidth more completely when stacked.

This is why the 8-channel E16 main control of the group can run to 5GB / s, while the 12-channel main control of the PS5 can only run to 5.5GB / s. In fact, the incomplete utilization of channel bandwidth by a single flash memory causes a certain amount of waste.

Of course, there is another possibility that the MT / s per channel value of the main control of PS5 may be lower. In other words, the upper limit of the speed of each channel on the main control is lower, but this just saves money ...

===================================

The main control of PS5 is DRAM-LESS, which is also critical because it saves money.

Sony uses a large data block to reduce the total amount of block addresses, thereby reducing the size of the LUT table mentioned above from the usual GB level to the current KB level, so it can be installed in a small SRAM cache, which saves I lost 1GB of DRAM. Don't underestimate this 1GB DRAM, the total of zero zero is reflected in the total cost, which can save almost a dozen dollars.

Memory access is simplified by address + offset. The map you use in your daily life will only show the XX building, not the XX room XX in the XX building, right? In the traditional PC SSD, the address table is equivalent to a map accurate to the XX floor, while the PS5 address table is equivalent to a map accurate to the XX cell. So obviously, the map of PS5 will be much smaller and easier to fit in your pocket.

Then, the rest is the offset. For ordinary PC, it is room XX, XX, and for PS5, it is room XX, XX building XX building ... However, this has little effect on the reading speed.

While writing, you can think of it as a fire drill. The writing of the PC is the fire drill by the people of "XX Building XX Building XX", and the writing of the PS5 is the fire drill by the people of "XX Community". In other words, the latter is more laborious. However, the impact on the game console scene is not great.

Therefore, a design like PS5 is not only more expensive than XSX, but it is also likely to be cheaper than XSX SSD. Because the capacity of the hard disk is smaller, you can spend less than 4 flash memory. Four 64GB flash memories are a lot of money, which is much more expensive than a higher-end SSD master ...

====================================

As for how to treat it? Quite simply, Sony's design goal is to save money, not (higher than competitors) high performance. High performance is just a matter of doing it. As mentioned earlier, you have saved 1/4 of the capacity, and the cost is much lower. If you don't add any more, won't you be killed by the player? So I did such a thing. It was more expensive to switch channels, but it wasn't much expensive and it wasn't the main control of the cost.

The bandwidth required by the graphics card for page swapping is under the current GDDR6 bandwidth level. Unless it is deliberately forced, there cannot be a scenario where SSDs above 2GB / s cannot be satisfied. In the normal game scene, most things should be in video memory and memory. Only in extreme cases (such as spaceships performing jumps) will mass exchanges be required. However, even if it is a mass exchange, 5.5GB / s SSD can't open the gap with 2.4GB / s SSD, because the entire memory space used for the game is only about 13GB, and a large amount of it will be stored in memory There is no need to exchange the items in it. Actually, there will be at most swap GB. You can't design a scene for the player to jump back and forth in order to use this 5.5GB / s. Do you want the player to trigger photosensitive flicker epilepsy ((

In addition, Virtual Geometry and Virtual Texture are essentially materials in picture format (VG is a geometric model saved in picture format). There are a lot of very efficient compression algorithms that can be directly addressed. This is also for both Mao PS5 and XSX. A hardware decompression chip is added, which can not only release the CPU, but also realize ultra-low latency material exchange. What do you mean, usually you use WinRAR to decompress the 1024 younger sister files in a package are made with CPU, and now with hardware, not only high bandwidth but low latency (and the delay is estimated and predictable, which is stable The number of frames is very important!) In addition, you can also directly access the corresponding younger sister without having to fully decompress. So Sony said that this is revolutionary for console game development, because rounding this is the virtual memory used by the game royal friends. It can be directly addressed, and the delay can be expected, which means that developers with brains can directly access almost any material they need at any time, and they can do a lot of access in advance, which largely hides the delay (although Not all).

=====================================

Next, answer the specific questions of the subject:

Will the memory advancement of the next generation become a bottleneck if it is small?

meeting. This problem will be more obvious in the light chase scene. Whether it is light chase, 4K, 60-120fps or noise reduction, which one is not a big video memory owner? All are memory destroyers. 448GB / s, 560GB / s, to a certain extent, has already set the upper limit of the next generation game memory access capacity. The gain of SSD is only capacity, not bandwidth.

We know that streaming technologies such as megatexture and virtual texture can save a lot of memory, but the bottleneck is hard disk IO, so can hard disk IO with ps5 up to 8-9GB / s solve this bottleneck?

This depends on how you define the bottleneck. The original Mega Texture and the current virtual texture streaming were two dishes in one dish. The bottleneck can be said to be the hard disk IO, not to mention the 8-9GB / s hard disk, even the 1GB / s hard disk is enough. Even if you have a SATA 3 SSD, you wo n’t see anything you ca n’t load in most games. Because the biggest bottleneck of virtual texture streaming is [random read and write speed], this is why games with linear levels generally use virtual textures casually, and you ca n’t see the card loading. Because linear levels are rarely read randomly, and open world games are different, a large number of random reads will instantly drag the entire game. On the other hand, both PC and PS4 use the CPU to decompress and read, so a lot of reading will also drag the CPU and cause unstable frame generation.

Of course, the above-mentioned violence theory is for PC 1440P and the following scenes. When it reaches 4K, it will rise appropriately> 2 times. However, the next generation is starting at 2.4GB / s, and there is hardware decompression, which can be said to have solved the problem perfectly at a certain cost.

Can such a high-speed SSD help improve the performance of the picture, thereby making up for the shortcomings of 15% of the floating point performance of Microsoft's xsx?

No, the picture performance improvement depends on the bandwidth of the entire data path. On the surface, PS5 is only "2TFlops lower" in floating-point performance, but if you look at the essence:

The memory bandwidth is 25% less, and the difference is 112GB / s, which is more scary than 9GB / s. Of course, I am just scaring you (escape
WGP-level caches such as LDS and L0 are 44% less, and L2 is 25% less. L1 is uncertain but should be less (otherwise XSX will easily feed the CU). For scenes like GPUs, the size of the cache is far more important than the frequency. The RTX 2070 Super is 40 SMs, 1.77GHz, but it ’s still the same as lifting 40 CUs, 1.9GHz 5700XT (Of course, this is just an analogy used to scare you. In fact, Turing efficiency is the same as Teraflops There are many factors that are higher than RDNA1. One of the main factors is that the data path of the N card is actually more efficient, and the buffer size allocated to each stream processor is more solid than the RDNA1 graphics card. In the words of lazy people, to a certain extent, the performance gain brought by the increase in cache capacity is much greater than the frequency increase. Because your cache at the same level is faster, at most twice as fast as others? However, the time overhead caused by each slow external access is more than ten times the normal cache access (
Now you should be able to understand why the demo of UE5 can only run to 1440P 30FPS. Because whether it is Nanite or Lumen, it actually eats more video memory bandwidth (rather than video memory capacity), or more, the GPU's ability to use video memory.
Overall, the data path bandwidth of XSX from memory to CU is much higher (between 25-50%, depending on the specific scenario), but even then I do n’t think XSX can run UE5 ’s demo at 4K 60FPS. Because the essence of this Demo is to show "you can see that I can make such a thing without good optimization", not "you see me, this thing is very good than other things." Especially Lumen, Lumen is actually SSGI + coarse mode reflection + sparse voxel tracking GI, Lumen's coarse mode reflection and voxel tracking part is not actually implemented with light chasing acceleration hardware, completely do not understand why, do not know API Not ready or Epic has personality, anyway, DXR can accelerate voxel tracking. Anyway, it's fascinating. Is it better to save the performance and improve the resolution?

Another example is that Nanite does not use Mesh Shader, of course, this is also expected, after all, PS5 is nothing like this. XSX's geometric performance is much higher than 2080Ti under good optimization. After using mesh, 2080Ti can actually render 50M + triangles and people still> 30fps. Nanite said in this demo that it actually renders triangles around 20M and still 30fps. It's embarrassing:


Mesh Shader running at 2080Ti, 45fps, 48M real drawn triangle, 4K resolution

If it is reduced to 20M solid-drawn triangle like UE5 Demo, or to 1440P, it can save a lot of resources for pixel shading and improve performance. Rounding up> 60fps is no problem (

As I said before, I once again confirmed that PS5 does not support Mesh Shader. The geometric performance is still limited by Primitive Shader with insufficient parallelism.

So do n’t have too high expectations for things that save money as the main design purpose, so it is not good for fans or manufacturers. . . If in the end it's really 399, isn't it really fragrant ...

Flex писал(а): 19 май 2020, 00:37 Короче gpu в ps5 такое же отсталое говно как и в ps3

Только в ps3 хотя бы на CELL можно было какие-то задачи с gpu перенести, а в случае с ps5 так не получится