How exactly does the infinity fabric for Ryzen work and why is Ryzen considered to have an SoC design?

[ZeRedz]

Title. I was wondering how exactly the infinity fabric for Ryzen works and why Ryzen is considered to have an SoC like design. Also, could someone explain to me exactly what v-cache is?

[mariushm]

Infinity fabric is the internal bus used to move data between components of the processor (cores, peripherals etc)

It's soc like design because besides cores and caches, the cpu also has integrated controllers for sata, usb etc and in theory it could function without a "chipset" like component. Threadripper in fact doesn't have chipset.

 

this is for a cpu with integrated graphics (4 cores, with vega graphics.. ex. ryzen 2200g or something like that)

 

 

as for v-cache ... is it that hard to google "amd v-cache" and read the first results?

 

it's extra cache for the cpu, which they plan to basically solder on top of the actual cpu die. the cache being more simplistic and repetitive a big chunk of memory cells, can be manufactured in a different process than cpu die, potentially cheaper. There's a lot of l1 and l2 and shared cache in a cpu die so that space could be used for more cores or other things, if the cache could be put on top of the die eventually.

 

here's for example you can see how much cache takes up from the cpu die ... all that stuff in the middle is cache the 32 MB of it.

The picture shows a 5600x - the IO stuff ( sata, usb controller, pci-e lanes etc) is on separate die under the cpu heatsink

from https://wccftech.com/amd-ryzen-5000-zen-3-vermeer-undressed-high-res-die-shots-close-ups-pictured-detailed/

 

[ZeRedz]

3 minutes ago, mariushm said:

Infinity fabric is the internal bus used to move data between components of the processor (cores, peripherals etc)

It's soc like design because besides cores and caches, the cpu also has integrated controllers for sata, usb etc and in theory it could function without a "chipset" like component. Threadripper in fact doesn't have chipset.

 

this is for a cpu with integrated graphics (4 cores, with vega graphics.. ex. ryzen 2200g or something like that)

 

 

as for v-cache ... is it that hard to google "amd v-cache" and read the first results?

 

it's extra cache for the cpu, which they plan to basically solder on top of the actual cpu die. the cache being more simplistic and repetitive a big chunk of memory cells, can be manufactured in a different process than cpu die, potentially cheaper. There's a lot of l1 and l2 and shared cache in a cpu die so that space could be used for more cores or other things, if the cache could be put on top of the die eventually.

 

here's for example you can see how much cache takes up from the cpu die ... all that stuff in the middle is cache the 32 MB of it.

The picture shows a 5600x :

 

I did google it and read whatever information I found. I just wanted to confirm my findings. Based on what I read, V-Cache basically involves stacking SRAM cache over the existing L3 cache and then connected them via TSV. Also, I read somewhere that it's because of the infinity fabric that Ryzen processors benefit from higher RAM frequencies. How true is this?

[Nathanpete]

3 minutes ago, mariushm said:

Infinity fabric is the internal bus used to move data between components of the processor (cores, peripherals etc)

It's soc like design because besides cores and caches, the cpu also has integrated controllers for sata, usb etc and in theory it could function without a "chipset" like component. Threadripper in fact doesn't have chipset.

Correction, threadripper does have a chipset, the X399 chipset for Zen and Zen+, and the TRX40 chipset for Zen 2 and hopefully Zen 3. 

 

Epyc, the server counterpart to threadripper, does not have a chipset, and the chipset is in fact integrated into the CPU. 

[ZeRedz]

2 minutes ago, Nathanpete said:

Correction, threadripper does have a chipset, the X399 chipset for Zen and Zen+, and the TRX40 chipset for Zen 2 and hopefully Zen 3. 

 

Epyc, the server counterpart to threadripper, does not have a chipset, and the chipset is in fact integrated into the CPU. 

What exactly is the benefit of integration a chipset onto a CPU and, I always wanted to ask this, is the purpose of even having different chipsets?

[Nathanpete]

  

Just now, ZeRedz said:

I did google it and read whatever information I found. I just wanted to confirm my findings. Based on what I read, V-Cache basically involves stacking SRAM cache over the existing L3 cache and then connected them via TSV. Also, I read somewhere that it's because of the infinity fabric that Ryzen processors benefit from higher RAM frequencies. How true is this?

Very true, up until around 4000MHz RAM. Their is a setting in the BIOS called the fabric clock, or f-clock, which should not be changed, and its number is equel to half of the memory clock, or m-clock, which is the frequency the system RAM is running at. Higher speed RAM equels high speed fabric, which equals less latency between parts of the CPU, which equals more performance. 

 

Do not raise the f-clock yourself as this will cause the f-clock to fall out of sync with the m-clock, resulting in an actual loss of performance. 

 

The 4000 MHz RAM limitation is because at the moment, getting a CPU with the capabilities to run stable with a 2000 MHz f-clock is hard, and most Zen 3 CPU's won't go above 1933 MHz F-clock, aka 3866 MHz RAM, but also, don't buy 3866 MHz RAM, 3600 MHz is much cheaper, and you lose less than 1% performance. 

[ZeRedz]

3 minutes ago, Nathanpete said:

  

Very true, up until around 4000MHz RAM. Their is a setting in the BIOS called the fabric clock, or f-clock, which should not be changed, and its number is equel to half of the memory clock, or m-clock, which is the frequency the system RAM is running at. Higher speed RAM equels high speed fabric, which equals less latency between parts of the CPU, which equals more performance. 

 

Do not raise the f-clock yourself as this will cause the f-clock to fall out of sync with the m-clock, resulting in an actual loss of performance. 

 

The 4000 MHz RAM limitation is because at the moment, getting a CPU with the capabilities to run stable with a 2000 MHz f-clock is hard, and most Zen 3 CPU's won't go above 1933 MHz F-clock, aka 3866 MHz RAM, but also, don't buy 3866 MHz RAM, 3600 MHz is much cheaper, and you lose less than 1% performance. 

Why is it set to half the clockspeed of the m-clock? Also, what exactly allows the infinity fabric to take better advantage of high RAM speeds? What exactly is the science behind it?

[mariushm]

1 minute ago, ZeRedz said:

I did google it and read whatever information I found. I just wanted to confirm my findings. Based on what I read, V-Cache basically involves stacking SRAM cache over the existing L3 cache and then connected them via TSV. Also, I read somewhere that it's because of the infinity fabric that Ryzen processors benefit from higher RAM frequencies. How true is this?

well because the infinity fabric can run up to around 1900-2000 mhz, it's best to use ram as close as possible to that maximum frequency, that's why 3600 Mhz memory sticks are recommended - 3600 mhz sticks run at 1800 mhz in reality, the 3600 mhz is just marketing.

when the infinity fabric matches the speed of the ram, everything's in sync and data is available on ram sticks at every tick of the infinity fabric frequency so things run great.

You can configure dividers in bios, where the ram will be accessed once every 2 ticks of infinity fabric or something like that, which keeps the ram synchronized but obviously there will be higher latency.

 

The idea of different chipsets is to offer flexibility... an a520 chipset may be smaller in size use less area on motherboard compared to a b550 chipset. They can make maybe 500 a520 chipsets from a round silicon wafer and sell them at 5$ while costing you 1$, or they may be able to do only 350 b550 chipsets from a round silicon wafer and sell them at 8$ but costing them $1.5-2 to make.

 

If the plan is to make a motherboard for an office pc, with a minimal number of usb connectors, 2-4 sata ports, just two slots of memory, no overclocking then you don't need a b550 chipset and could work with a520 chipset just fine, and company like dell or hp will go for the 5$ chipset instead of paying 8$ for the b550 chipset.

x570 is larger than b550 and offers pci-e 4.0 support,and more things compared to b550 chipset  but the downside is that the pci-e 4.0 controller is more power hungry and makes the chipset need active cooling. So if someone doesn't need extra pci-e 4.0 lanes or doesn't want a fan on chipset, they can use b550.

[Nathanpete]

  

15 minutes ago, ZeRedz said:

What exactly is the benefit of integration a chipset onto a CPU and, I always wanted to ask this, is the purpose of even having different chipsets?

Gonna be honest, even googling that first part came up empty. But it appears that a possible answer is that multiple socket motherboards exist for Epyc, but not ryzen or threadripper, and having two CPU's fight over time communicating with the chipset, which they would share, would massively slow down performance, and the target customer of Epyc, data centers and big corporations, wouldn't like that, so by integrating a chipset into each Epyc CPU, they can just communicate with each other without a slow middle man, massively increasing efficiency in multi-socket configurations. 

 

Again, that was just a theory I found on reddit, no idea if it is true. 

 

Different chipsets arise to introduce new features. For example, Threadripper 1000 and 2000 did not support PCIe gen 4, because the motherboards, but most importantly, the X399 chipset, did not have the hardware to handle PCIe gen 4. Well with the introduction of PCIe Gen 4 with Threadripper 3000, a new chipset was introduced, with new features, with by far the biggest one being PCIe Gen 4, and more lanes in total than X399. 

8 minutes ago, ZeRedz said:

Why is it set to half the clockspeed of the m-clock? Also, what exactly allows the infinity fabric to take better advantage of high RAM speeds?

Well that is because DDR4 actually stands for Double Data Rate, which means 3600 MHz RAM actually is 1800 MHz RAM, but because it can send double the data, it is effectively running at double the speed, so 3600 MHz. But in true reality, it really is 1800 MHz, and for synchronization reasons this is what the F-clock is also set to, the m-clock, which is the real frequency of the RAM. They double the frequency on the box for marketing and to reduce confusion, and it has become the standard to do so for over a decade. 

[Nathanpete]

@ZeRedzI'm sorry for my rude response at first, I have edited it to be more educational. 

[ZeRedz]

1 minute ago, Nathanpete said:

@ZeRedzI'm sorry for my rude response at first, I have edited it to be more educational. 

You're good, dude. I got what you were saying. S'all cool.