AMD Talks V-cache, Ryzen 6000 And AM5 - The Full Nerd Special Edition
video description
Date: 2022-03-15
Related videos
Comments and reviews: 10
Philtography
I am hearing some very interesting things about Zen 4 and it explains (in my mind anyway) why AMD made Ryzen 5800X3D, it-s mainly a test production to check everything works correctly for what they actually have planned for 3D cache later this year with Zen 4.
What AMD appears to be planning is to move the L3 cache off the die itself and exclusively use the stacked 3D cache as their only level 3 cache for the CPU. This would free up massive amounts of die space for the ever increasing core sizes and core count and also create room for things like dedicated accelerators and GPU cores. But crucially it also allows AMD the space to add smaller -efficiency cores- (although AMD will call them low power cores) to go under the cache as these won't require as much cooling.
I believe this is why we have heard so many rumours about a 12-core CCX and so a 24 core Zen 4 chip for desktop, it might even be 16-core CCX and a 32 core Ryzen chip. These rumours just didn-t make any sense to me as 16 high performance cores is easily enough for 99% of desktop users and the other 1% can buy Threadripper or Epyc workstations. But now this makes total sense, it's still going to be 16 big cores but another 8 to 16 low powered cores too. This is partially to fight against Intel's marketing when a 24 core Raptor Lake comes out later this year, I know anyone reading this knows that Intel-s chip is actually only an 8 core chip with a load of much slower efficiency cores but Joe Average on the street wouldn't, they just see 24 cores being bigger=better than 16 cores and buy that.
I have also heard that AMD will use the low power cores as -over-flow cores-, so the normal Zen 4 cores will always be used first and handle everything until they get 100% used and then the low power cores will start to be used. This idea means there won-t be any scheduling issues where high performance tasks are wrongly sent to slow efficiency cores but it also means that things that would be much better off going to a low powered core get sent to the high powered cores which increases heat and power usage, so I guess it-s swings and roundabouts as to which way is best.
reply
I am hearing some very interesting things about Zen 4 and it explains (in my mind anyway) why AMD made Ryzen 5800X3D, it-s mainly a test production to check everything works correctly for what they actually have planned for 3D cache later this year with Zen 4.
What AMD appears to be planning is to move the L3 cache off the die itself and exclusively use the stacked 3D cache as their only level 3 cache for the CPU. This would free up massive amounts of die space for the ever increasing core sizes and core count and also create room for things like dedicated accelerators and GPU cores. But crucially it also allows AMD the space to add smaller -efficiency cores- (although AMD will call them low power cores) to go under the cache as these won't require as much cooling.
I believe this is why we have heard so many rumours about a 12-core CCX and so a 24 core Zen 4 chip for desktop, it might even be 16-core CCX and a 32 core Ryzen chip. These rumours just didn-t make any sense to me as 16 high performance cores is easily enough for 99% of desktop users and the other 1% can buy Threadripper or Epyc workstations. But now this makes total sense, it's still going to be 16 big cores but another 8 to 16 low powered cores too. This is partially to fight against Intel's marketing when a 24 core Raptor Lake comes out later this year, I know anyone reading this knows that Intel-s chip is actually only an 8 core chip with a load of much slower efficiency cores but Joe Average on the street wouldn't, they just see 24 cores being bigger=better than 16 cores and buy that.
I have also heard that AMD will use the low power cores as -over-flow cores-, so the normal Zen 4 cores will always be used first and handle everything until they get 100% used and then the low power cores will start to be used. This idea means there won-t be any scheduling issues where high performance tasks are wrongly sent to slow efficiency cores but it also means that things that would be much better off going to a low powered core get sent to the high powered cores which increases heat and power usage, so I guess it-s swings and roundabouts as to which way is best.
reply
crazy
As always great show. Just a follow up to what the last question was about. (Gamers Nexus has a lot more info on this if you are looking for it. He was stating that OEMS, Like some mentioned in the video, when they design around an OEM only part, like the -PRO- parts refenced here. They use custom designed mother boards, PSU's and other components. So lets say for whatever reason a customer really wants one of those -PRO- parts but the only way to get it is through an OEM then you have a lot of -E- waste in the end. Why? Well take the custom designed MB and PSU for instance, if one fails but you still really want that -PRO- part for your specific usage scenario then you have to now re-buy the entire system instead of just replacing the 1 bad component (Or pay 3 times what its normal value if it were built to a -normal- massed produced spec) This causes an enormous amount of e-waste all in the name of profits, because you have to either replace the entire system or buy overly expensive part from an OEM. Most customers when they figure this out end up replacing the whole system because it is not worth it to replace one part at a price way higher than market value.
reply
As always great show. Just a follow up to what the last question was about. (Gamers Nexus has a lot more info on this if you are looking for it. He was stating that OEMS, Like some mentioned in the video, when they design around an OEM only part, like the -PRO- parts refenced here. They use custom designed mother boards, PSU's and other components. So lets say for whatever reason a customer really wants one of those -PRO- parts but the only way to get it is through an OEM then you have a lot of -E- waste in the end. Why? Well take the custom designed MB and PSU for instance, if one fails but you still really want that -PRO- part for your specific usage scenario then you have to now re-buy the entire system instead of just replacing the 1 bad component (Or pay 3 times what its normal value if it were built to a -normal- massed produced spec) This causes an enormous amount of e-waste all in the name of profits, because you have to either replace the entire system or buy overly expensive part from an OEM. Most customers when they figure this out end up replacing the whole system because it is not worth it to replace one part at a price way higher than market value.
reply
The
2: 58 3D V-cache and 5800X3D.
6: 45 5800X3D and some decrease in frequency.
8: 57 price of 5800X3D.
10: 05 modern games and Frequency vs Cores vs Cache for it.
11: 44 latency of V-cache.
12: 28 L3 V-cache vs L2 V-cache.
14: 26 5GHz CPU vs World best CPU.
16: 58 about Cinebench.
19: 30 MS Pluton.
21: 06 details about Ryzen6000.
21: 57 V-cache for non-games apps.
22: 58 asymmetric V-cache.
24: 18 V-cache on I/O-die.
25: 00 software issues with V-cache.
26: 15 performance of Ryzen6000.
28: 44 15W vs 28W of Ryzen U-series.
31: 15 new adapting for new mobile Ryzen.
33: 41 LGA for Ryzen7000.
36: 59 supply of DDR5 and LPDDR5 for Ryzen6000.
reply
2: 58 3D V-cache and 5800X3D.
6: 45 5800X3D and some decrease in frequency.
8: 57 price of 5800X3D.
10: 05 modern games and Frequency vs Cores vs Cache for it.
11: 44 latency of V-cache.
12: 28 L3 V-cache vs L2 V-cache.
14: 26 5GHz CPU vs World best CPU.
16: 58 about Cinebench.
19: 30 MS Pluton.
21: 06 details about Ryzen6000.
21: 57 V-cache for non-games apps.
22: 58 asymmetric V-cache.
24: 18 V-cache on I/O-die.
25: 00 software issues with V-cache.
26: 15 performance of Ryzen6000.
28: 44 15W vs 28W of Ryzen U-series.
31: 15 new adapting for new mobile Ryzen.
33: 41 LGA for Ryzen7000.
36: 59 supply of DDR5 and LPDDR5 for Ryzen6000.
reply
Nocturnal101
Well that was a bit disappointing to hear from AMD GPU side, now I am actually worried for them, PCIe 5 has CXL baked into the standard as well as DX support with Win 11, as Chiplet and MXM configs become more parallel across dies the drivers will need to account for that, Intel may not have to launch high end GPUs if they can turn out mainstream level GPUs that you can pair together as co-processors, NVidia I would imagine with NVlink will already be co-adopted which may lead to AMD having mega amounts of trouble if all they stick to is monolithic and single card solutions in the race to 8k and 4k with high refresh rates at native rendering rather than post processing AI cosmetics.
reply
Well that was a bit disappointing to hear from AMD GPU side, now I am actually worried for them, PCIe 5 has CXL baked into the standard as well as DX support with Win 11, as Chiplet and MXM configs become more parallel across dies the drivers will need to account for that, Intel may not have to launch high end GPUs if they can turn out mainstream level GPUs that you can pair together as co-processors, NVidia I would imagine with NVlink will already be co-adopted which may lead to AMD having mega amounts of trouble if all they stick to is monolithic and single card solutions in the race to 8k and 4k with high refresh rates at native rendering rather than post processing AI cosmetics.
reply
itech
I think Frank just Jebaited us. I think the Zen 4 chiplet has 8 high power cores and 8 LTDP cores, with a stacked L3 3D cache. I think AMD has it designed into the chiplet to take the stacked cash for high end models, or not-for standard models.
Analyzing what AMD had to decide two years ago when laying out Zen 4, to me they HAD to make the decision to expand the chiplet to 16 cores on 5nm. Why? Because they knew they could get it designed and working and execute it.
reply
I think Frank just Jebaited us. I think the Zen 4 chiplet has 8 high power cores and 8 LTDP cores, with a stacked L3 3D cache. I think AMD has it designed into the chiplet to take the stacked cash for high end models, or not-for standard models.
Analyzing what AMD had to decide two years ago when laying out Zen 4, to me they HAD to make the decision to expand the chiplet to 16 cores on 5nm. Why? Because they knew they could get it designed and working and execute it.
reply
MR
the gpu prices already choking us. we dont have to much room for more expensive memory. seeing tests, the entry level ddr5 are not faster in most cases then a ddr4 3600. a kingston fury 3600 32gb 4 dim kit is $200 a DDR5 5200 32gb 2 dim $350.
the real performance uplift in most cases is just few % but for 75% more price. what would make sense, is the ddr5 6000mhz, that does have more performance over ddr4 3600, but thats $490 per 32gb 2 dim kit.
reply
the gpu prices already choking us. we dont have to much room for more expensive memory. seeing tests, the entry level ddr5 are not faster in most cases then a ddr4 3600. a kingston fury 3600 32gb 4 dim kit is $200 a DDR5 5200 32gb 2 dim $350.
the real performance uplift in most cases is just few % but for 75% more price. what would make sense, is the ddr5 6000mhz, that does have more performance over ddr4 3600, but thats $490 per 32gb 2 dim kit.
reply
pcworld
in 2022, I really hope AMD releases a 4GB video card with 2015 performance without HW Video encoding and a neutured Bus. and topped off with a fake MSRP that is at minimum $80 USD OVER the listed imaginary number, especially for the lowest end cards in the lineup. Wish I may, wish I might.
reply
in 2022, I really hope AMD releases a 4GB video card with 2015 performance without HW Video encoding and a neutured Bus. and topped off with a fake MSRP that is at minimum $80 USD OVER the listed imaginary number, especially for the lowest end cards in the lineup. Wish I may, wish I might.
reply
Osvaldo
On the CPU locking issue: that feature is for security conscious corporations, and these people don't allow much hardware to be recycled. Once some CPU, motherboard, disk or memory stick is decommissioned, it goes to a metal grinder aka -real secure wipe-.
reply
On the CPU locking issue: that feature is for security conscious corporations, and these people don't allow much hardware to be recycled. Once some CPU, motherboard, disk or memory stick is decommissioned, it goes to a metal grinder aka -real secure wipe-.
reply
Robert
FID. FUD. FUD. Robert is full of shit. Painting a false narrative regarding ram speed trumping core speed, meanwhile AMD refuses to acknowledge the issues with over boosting dynamic FCLK power saving features throwing WHEA 19 errors on Zen 3.
F U D
reply
FID. FUD. FUD. Robert is full of shit. Painting a false narrative regarding ram speed trumping core speed, meanwhile AMD refuses to acknowledge the issues with over boosting dynamic FCLK power saving features throwing WHEA 19 errors on Zen 3.
F U D
reply
storm9154
i wanted a Laptop as a back up to my 3700X system and decided on a Lenovo Legion 4800H with a RTX2060 and what i surprise i got, it was super fast so i would get a new 6000 series next
reply
i wanted a Laptop as a back up to my 3700X system and decided on a Lenovo Legion 4800H with a RTX2060 and what i surprise i got, it was super fast so i would get a new 6000 series next
reply
Add a review, comment
Other channel videos
1:8:44
