New STH videos always makes my day better. Of course I absolutely love the main site and all of the excellent articles / write-ups but these videos are great and I'm really glad that you guys have continued to do them.
The point about ethernet vs. Infiniband is really interesting. People sometimes may think that 100GbE or 200 GbE is the same as EDR IB or HDR IB. I can tell you, especially from the latency perspective, they’re not. Ethernet has at minimum, upwards of 3-4% higher latency than IB and that, to these systems and installations, makes a HUGE difference. In fact, I was thinking about this earlier where AMD is the most significant and mature platform right now that can support two HDR IB cards at full speed using only A QUARTER of the total number of PCIe 4.0 lanes that the CPUs provide.
Lenovo's actions make sense from a sales point of view. It is kind of anoying from the point of view of the press and reviewers, because it amounts to manipulation of the data.
There is another "phenomenon" of this kind. Most Internet cafes in china used Nvidia graphics cards, and everytime someone logged in to his steam account, It counted as a separate system with an Nvidia card. So, if you used Steam entries as indicator of how many Nvidia cards were sold (for example for Stock market analysis) you could be wrong by a decent margin.
@@PanduPoluan I'm not talking about Nvidias own supercomputer, I'm talking about the Systems using A100 accelerators. Most use Epyc Rome, since that's the Nvidia reference and was the only PCIe 4.0 option when it was developed.
It's nice to see that AMD can compete with Intel in a market that has been Intel dominated for a long time. Maybe eventually AMD will be able to compete with Nvidia like they are doing with Intel.
I love to see NEC still scoring wins for their VEs. I've been really fascinated by how much memory bandwidth the two Tsubasa generations have and it seems like they're only going to get faster in the future. They really predicted the move to on-CPU HBM
Love this kind of drill down, it's been interesting seeing my little part of the server market start to go AMD over the last year and here's to hoping this will drive down some used server prices so I can upgrade my home lab! It's funny cause the enthusiast side of the market thinks AMD is 'beating' Nvidia not realizing that for as much press as the desktop GPU market gets Nvidia still completely owns the HEDT/enterprise market which not counting mining over the past years is where the big bucks are spent.
I wonder if we're going to see AMD pair away some of the lower core count options over the next generation based on the core count demand shown here - that's shocking that nothing is quite close to the demand of 64c parts.
They already are using Rome parts as the lower-end even with Milan out. This is for HPC so density is a big deal and there is cooling available for higher-end parts.
Suddenly those rumors of 128 core EPYC chips from "Moore'sLawIsDead" are starting to make sense. AMD must have seen the huge demand for 64 core chips and hence they're upping the core count to double. There's more incentive too since ARM chips have 90 cores, and AMD will probably want to compete with that. if the leak is true, 5nm EPYC with 128c/socket are going to massacre Intel.
@@ServeTheHomeVideo That is how the server world works, though. It's slow to change and adopt. The problem with that for Intel is that they have likely almost universally lost those customers for at least a decade. Also with AMD's push to move drivers and software open source it will be that much more difficult to convince them to come back.
The only reason intel even exists on TOP500 list is because AMD cannot get enough chips. The limited capacity of external foundries has really saved intel.
Do you still want to think AMD is making sustainable progress? AMD has a unique sku with 64 but 2021/2022 but Intel makes their announcements at SuperComputing (SC21). I had mixed reviews about this video but you covered your assumptions at the end which was good. 2022 thought will be different for all these vendors. Now to your A100 reviews 😃
AMD achieved so many records for one primary reason, the application suite being benchmarked or synthetic benchmark for the most part fit entirely in Epyc L3 cache. Large L3 is why AMD secured so many records but are these records, benchmarks, indicative of on system / platform performance and that is a prime question. There is no doubt AMD codevelopers are pioneering, leading, in greater than 28 core application development (not counting out ARM server) which historically is what AMD always did for Intel, Intel even allowing it giving AMD compliment room to innovate too watch and learn on AMD's investment in compliment (Intel customer) lead development. mb
ARM or RISC-5 will both give more processing for the same energy as you get with x86. I think ARM is better. This doesn't mean, however, that somebody won't blindside the whole market with a new technology. We are about due for it.
If, say, Intel simply scaled up however many parallel units it uses in a supercomputer cluster-thing (hey, I don't know) wouldn't they move up the list? It's not as if there's any novel core architecture going on here.
This is really encouraging, hopefully Intel will now get off the bench and start producing decent products again instead of the lazy incremental junk they're been fobbing us off with.
Security - anything wifi/blue tooth is not secure... Just look at all the 'smart' devices which form a greater mesh. Low power blue tooth can connect/ relay data up to 500 meters..
New STH videos always makes my day better. Of course I absolutely love the main site and all of the excellent articles / write-ups but these videos are great and I'm really glad that you guys have continued to do them.
Hey guy! He is Patrick from STH!
The point about ethernet vs. Infiniband is really interesting.
People sometimes may think that 100GbE or 200 GbE is the same as EDR IB or HDR IB.
I can tell you, especially from the latency perspective, they’re not.
Ethernet has at minimum, upwards of 3-4% higher latency than IB and that, to these systems and installations, makes a HUGE difference.
In fact, I was thinking about this earlier where AMD is the most significant and mature platform right now that can support two HDR IB cards at full speed using only A QUARTER of the total number of PCIe 4.0 lanes that the CPUs provide.
Lenovo's actions make sense from a sales point of view. It is kind of anoying from the point of view of the press and reviewers, because it amounts to manipulation of the data.
There is another "phenomenon" of this kind. Most Internet cafes in china used Nvidia graphics cards, and everytime someone logged in to his steam account, It counted as a separate system with an Nvidia card. So, if you used Steam entries as indicator of how many Nvidia cards were sold (for example for Stock market analysis) you could be wrong by a decent margin.
Thanks for the awesome video 👍👍
AMD's upcoming HPC accelerator MI200 features a whopping 128GB HBM2E memory.
I would like to see a full detailed video on NEC Vector Engine
If possible compare performance of it with traditional CPUs
I think part of the crazy high numbers for Rome is that the Nvidia systems use Rome instead of moving to Milan.
That's not a new system.
@@PanduPoluan I'm not talking about Nvidias own supercomputer, I'm talking about the Systems using A100 accelerators. Most use Epyc Rome, since that's the Nvidia reference and was the only PCIe 4.0 option when it was developed.
That is at least somewhat true. Many of the HGX A100 systems are now using Milan though
It's nice to see that AMD can compete with Intel in a market that has been Intel dominated for a long time. Maybe eventually AMD will be able to compete with Nvidia like they are doing with Intel.
I love to see NEC still scoring wins for their VEs. I've been really fascinated by how much memory bandwidth the two Tsubasa generations have and it seems like they're only going to get faster in the future. They really predicted the move to on-CPU HBM
Just say now some OEMs are pushing AMD systems really hard. Looking forward to this quarter's market share data.
Love this kind of drill down, it's been interesting seeing my little part of the server market start to go AMD over the last year and here's to hoping this will drive down some used server prices so I can upgrade my home lab! It's funny cause the enthusiast side of the market thinks AMD is 'beating' Nvidia not realizing that for as much press as the desktop GPU market gets Nvidia still completely owns the HEDT/enterprise market which not counting mining over the past years is where the big bucks are spent.
Just a while ago someone on STH claimed no one's buying AMD Epyc because epyc allegedly has serious issues lol
Watching the Top500 try to catch up with Folding@Home. Crazy, man.
Great video, as always!
How far back in time would we need to go before a pc with a 3090 and a threadripper would top the list?
Looking forward to the A100 content!
I just love this channel!
I wonder if we're going to see AMD pair away some of the lower core count options over the next generation based on the core count demand shown here - that's shocking that nothing is quite close to the demand of 64c parts.
They already are using Rome parts as the lower-end even with Milan out. This is for HPC so density is a big deal and there is cooling available for higher-end parts.
Suddenly those rumors of 128 core EPYC chips from "Moore'sLawIsDead" are starting to make sense. AMD must have seen the huge demand for 64 core chips and hence they're upping the core count to double. There's more incentive too since ARM chips have 90 cores, and AMD will probably want to compete with that. if the leak is true, 5nm EPYC with 128c/socket are going to massacre Intel.
Thanks Patrick
Great video
Surprise surprise... Intel is in trouble... it was obvious for anyone with a comon sense that this would eventually happens.
It took much longer than many expected. You can see that comparing the Nov 20 to this June 21 mix of CPUs
@@ServeTheHomeVideo That is how the server world works, though. It's slow to change and adopt. The problem with that for Intel is that they have likely almost universally lost those customers for at least a decade. Also with AMD's push to move drivers and software open source it will be that much more difficult to convince them to come back.
The only reason intel even exists on TOP500 list is because AMD cannot get enough chips.
The limited capacity of external foundries has really saved intel.
Is Big Red 2 on that list? I know they were going AMD.
Awesome
Thanks for the info
I can't wait to get a 64 core cpu
Does anyone know if Fujitsu are working on an update to the A64FX?
Xeon Phi dying really hurt intels chance to fight in HPC vs Milan and rome
Do you still want to think AMD is making sustainable progress? AMD has a unique sku with 64 but 2021/2022 but Intel makes their announcements at SuperComputing (SC21). I had mixed reviews about this video but you covered your assumptions at the end which was good. 2022 thought will be different for all these vendors. Now to your A100 reviews 😃
Shame Oracle killed off Sun MicroSystems/Fujitsu systems
really cool video!
Glad you think so!
Go AMD ftw!
Ty Pat learned heaps
AMD achieved so many records for one primary reason, the application suite being benchmarked or synthetic benchmark for the most part fit entirely in Epyc L3 cache. Large L3 is why AMD secured so many records but are these records, benchmarks, indicative of on system / platform performance and that is a prime question. There is no doubt AMD codevelopers are pioneering, leading, in greater than 28 core application development (not counting out ARM server) which historically is what AMD always did for Intel, Intel even allowing it giving AMD compliment room to innovate too watch and learn on AMD's investment in compliment (Intel customer) lead development. mb
AMD deserves the success, but I am rooting for ARM based processors :)
ARM or RISC-5 will both give more processing for the same energy as you get with x86. I think ARM is better. This doesn't mean, however, that somebody won't blindside the whole market with a new technology. We are about due for it.
It is annoying that AMD created the EPYC chip when Intel/HP created a EPIC chip (Itanium) and I always think back to that.
If, say, Intel simply scaled up however many parallel units it uses in a supercomputer cluster-thing (hey, I don't know) wouldn't they move up the list? It's not as if there's any novel core architecture going on here.
Frontier will use amd gpu and will be the first exascale computer this yearand will be number 1 supercomputer in the world surpassing fugaku arm
Not in this list so holding off talking about that.
This is the end 🎶
The 7V12 is a 64-core CPU!
This is really encouraging, hopefully Intel will now get off the bench and start producing decent products again instead of the lazy incremental junk they're been fobbing us off with.
The Fugaku uses an ARM based CPU. RISC once again proves its superiority. 😎
nice video
I hope AMD is building a good watchest. Intel can't keep being down forever.
New supercomputers based on Skylake? LOL?
Security - anything wifi/blue tooth is not secure... Just look at all the 'smart' devices which form a greater mesh. Low power blue tooth can connect/ relay data up to 500 meters..
its worse when HP HPE for AMDdoubles their prices all of a sudden
AMD may have the better processor, but if they're not on the shelf they're not worth shit.
From the stats in this video it looks like they’re available just fine.
They are mostly avalible and the epyc lines have the highest chiplet priority
Amd still suck dont lie amd is just good for gaming cant even run maya without crashing
AMD is overpriced and is loosing market share by the second...