Current Generation Hardware Speculation with a Technical Spin [post GDC 2020] [XBSX, PS5]
- Thread starter Proelite
- Start date
- Status
DavidGraham
Veteran
Machine Learning is supported on the XSX on the shaders, and Microsoft intends to use it.
AbsoluteBeginner
Regular
snc
Veteran
I think you know all his posts so know better
It is 560GB/s total. The 6 GB has lower bandwidth because it's the "upper half" of the 2GB chips, that are only attached to 6 of the 10 32-bit interfaces. The 336GB/s is part of the 560GB/s.
Good to see that the RT implementation is like the RTX one, just a port with NV code worked on XSX. RTX 2080 performance in the XSX nice work.
Just 16GB total, doesn't seem much, as there seems to be no seperate DDR4, like the phil spencer SoC upload indicated, with 10 for graphics.
2.4gb/s NVme PCIE4.0 but can reach 6gb/s?
Just 16GB total, doesn't seem much, as there seems to be no seperate DDR4, like the phil spencer SoC upload indicated, with 10 for graphics.
2.4gb/s NVme PCIE4.0 but can reach 6gb/s?
Great overall specs for the X-series. Nice to see Klee was right on the money.
They probably have some kind of jpeg-like lossy compression system meant for textures, where the decompressor is a hardware part with a ~6GB/s output. It makes a lot of sense have such a thing when you have 2.4GB/s storage and a limited ram pool -- it means you can fetch up to a 100 megs of textures directly from the storage to uncompressed (or, only texture compressed) form in memory per frame.
This in turn means there is no need for an intermediate form of texture storage as compressed form in RAM, like is used today for virtual texturing. Saves a lot of ram, and makes that 16GB last a lot longer.
Panino Manino
Newcomer
Microsoft is quoting "4-bit integer". Is this on the shaders or on some dedicated silicon?
I believe it is stated that the decompression chip is capable of delivering 6GB/s.
On shaders, it was first implemented in Vega 20
I read it as 2.4 GB/s raw bandwidth which with compressed data can achieve equivalent of 6 GB/s bandwidth with uncompressed data.
On the shaders. They have a added a bunch of lower-precision modes.
DF's official specs for the SSD (from the chart) are:
2.4GB/s (Raw), 4.8GB/s (Compressed)
And the quote from the article says
DF said:
It is 4,8 GB of uncompressed data.
On the shaders. Data movement is generally much more expensive than the actual computation, and more so the smaller your data elements are. Having that much extra compute on dedicated silicon would almost double the size of the GPU, whereas just providing more formats at the shaders is dramatically less expensive.
The technical solution they have almost certainly gone for is to make packed formats for 32f, 16f, 8i and 4i available at every SIMD pipe.
Panino Manino
Newcomer
I had no idea that pack math went all the way to 4-bit.
Maybe NOW they're feeling more like hiding...
Maybe the PS5 will have more memory, for besides from this I can't see the PS5 being stronger. I expect Sony to talk less about the hardware and technology and more, much more about all the games they're developing for the launch.
Maybe NOW they're feeling more like hiding...
Maybe the PS5 will have more memory, for besides from this I can't see the PS5 being stronger. I expect Sony to talk less about the hardware and technology and more, much more about all the games they're developing for the launch.
Yea, they said they are bringing DirectStorage to Windows PC as well. Very cool. Wonder what other improvements and changes they have in store for the PC side of the equation?
- Status
Similar threads
- Locked
