AMD Onslaught vs. Intel: Zen 4, Zen 5, & Threadripper Details
video description
Date: 2022-06-12
Related videos
Comments and reviews: 10
Joshua
I wouldn't spend the money on Threadripper Pro, even if you can get a tax write-off. Even if it performs wonderfully in the majority of instances, in the case of Dell (pun intended) we just aren't going to get a realistic look at the capability of that platform, when it is almost certainly going to be suffocated by Dell's incompetence. We all know what that picture looks like, no sense buying and rewarding Dell for more e-waste.
I would almost argue you would be better off buying a sWRX8 motherboard and the CPU itself but then it just comes down to the question of if there is enough of an audience that is interested in the platform and Threadripper Pro CPUs to even remotely recoup the cost in ad revenue and push to the GN Store? You folks would have a better understanding of the metrics than I do, but I would be genuinely surprised (in a good way) if the answer was yes. Honestly I had an interest in it only so far as Threadripper non-PRO is dead, what is the next best option? but now that we know new Threadripper is on the horizon on Zen 4 no less, the interest has certainly tapered. Take those thoughts as you will though, I am only speaking for myself.
reply
I wouldn't spend the money on Threadripper Pro, even if you can get a tax write-off. Even if it performs wonderfully in the majority of instances, in the case of Dell (pun intended) we just aren't going to get a realistic look at the capability of that platform, when it is almost certainly going to be suffocated by Dell's incompetence. We all know what that picture looks like, no sense buying and rewarding Dell for more e-waste.
I would almost argue you would be better off buying a sWRX8 motherboard and the CPU itself but then it just comes down to the question of if there is enough of an audience that is interested in the platform and Threadripper Pro CPUs to even remotely recoup the cost in ad revenue and push to the GN Store? You folks would have a better understanding of the metrics than I do, but I would be genuinely surprised (in a good way) if the answer was yes. Honestly I had an interest in it only so far as Threadripper non-PRO is dead, what is the next best option? but now that we know new Threadripper is on the horizon on Zen 4 no less, the interest has certainly tapered. Take those thoughts as you will though, I am only speaking for myself.
reply
James
In genetics, complementary DNA (cDNA) is DNA synthesized from a single-stranded RNA template in a reaction catalyzed by the enzyme reverse transcriptase.
However
The AMD CDNA architecture is built on AMD ROCm open software platform, the industry's only open software platform for GPU computing.
Thusly
AMD CDNA is an. all-new GPU architecture from AMD to drive accelerated computing into the era of exascale computing. The new architecture augments. scalar and vector processing with new Matrix Core Engines and adds Infinity Fabric technology to scale up to larger systems.
Basically = The Computer Can Design The New Architecture It s Self , While Using Open Software Collecting Personal Data to use as its means for the construction thereof , Again = It s Self . Almost , Skynet worthy , and if we Don t Grasp hold of our Own Faults = This Will be part of it s design , sitting at the brink of destruction or fait , the direction is our to make , i think we need to get it together people and stop all this foolish bickering NOW !
reply
In genetics, complementary DNA (cDNA) is DNA synthesized from a single-stranded RNA template in a reaction catalyzed by the enzyme reverse transcriptase.
However
The AMD CDNA architecture is built on AMD ROCm open software platform, the industry's only open software platform for GPU computing.
Thusly
AMD CDNA is an. all-new GPU architecture from AMD to drive accelerated computing into the era of exascale computing. The new architecture augments. scalar and vector processing with new Matrix Core Engines and adds Infinity Fabric technology to scale up to larger systems.
Basically = The Computer Can Design The New Architecture It s Self , While Using Open Software Collecting Personal Data to use as its means for the construction thereof , Again = It s Self . Almost , Skynet worthy , and if we Don t Grasp hold of our Own Faults = This Will be part of it s design , sitting at the brink of destruction or fait , the direction is our to make , i think we need to get it together people and stop all this foolish bickering NOW !
reply
Eru
AI processing demands depend on exactly what you are doing:
- If you are performing inference int8 is usually fine but if you are training, you usually want fp32.
- Training is way more demanding in all aspects than inference.
- For small organizations on a limited budget, the issue tends not to be processing power but not having enough VRAM to train at all.
- No matter how good is the hardware, competing with cuda is really hard.
- Depending on the architecture of the neural network it may be compute bound , memory bound or somewhere in between.
- Most of the compute in ANNs consists of multiplicating a bunch of numbers and adding the result of all multiplications.
I sincerely hope that somehow AMD pulls it off, NVIDIA needs competition in that field. I would also love to see cuda replaced by an open standard but I don't see it happening unless AMD manages some kind of crazy advantage like 10x the performance for the same cost.
reply
AI processing demands depend on exactly what you are doing:
- If you are performing inference int8 is usually fine but if you are training, you usually want fp32.
- Training is way more demanding in all aspects than inference.
- For small organizations on a limited budget, the issue tends not to be processing power but not having enough VRAM to train at all.
- No matter how good is the hardware, competing with cuda is really hard.
- Depending on the architecture of the neural network it may be compute bound , memory bound or somewhere in between.
- Most of the compute in ANNs consists of multiplicating a bunch of numbers and adding the result of all multiplications.
I sincerely hope that somehow AMD pulls it off, NVIDIA needs competition in that field. I would also love to see cuda replaced by an open standard but I don't see it happening unless AMD manages some kind of crazy advantage like 10x the performance for the same cost.
reply
Mac
I dont know what they exactly mean by metaverse. But the future is 100% in VR and RealTime Rendering incl. Raytracing and all the other demanding sh!t - and that all in a Big VR World. That will be the next World of Warcraft, in the sense of Success and Big Player Community - but much much greater in a realistic rendered 3d world. I am already waiting to invest money into that. And thats maybe what they mean by the word metaverse : A great demand on compute power for realistic rendered VR world. And for this you really need more Gigaflops ... However, it was really fun, I like your humour!
My english is not the best, but imagine the Simulators in Star Trek and what they probably would need on gigaflops with all the AI and stuff .... thats what I mean in which direction its goin'.
And even today: Think on the game/project Star Citizen ... they clearly would profit from more and more and more gigaflops .
reply
I dont know what they exactly mean by metaverse. But the future is 100% in VR and RealTime Rendering incl. Raytracing and all the other demanding sh!t - and that all in a Big VR World. That will be the next World of Warcraft, in the sense of Success and Big Player Community - but much much greater in a realistic rendered 3d world. I am already waiting to invest money into that. And thats maybe what they mean by the word metaverse : A great demand on compute power for realistic rendered VR world. And for this you really need more Gigaflops ... However, it was really fun, I like your humour!
My english is not the best, but imagine the Simulators in Star Trek and what they probably would need on gigaflops with all the AI and stuff .... thats what I mean in which direction its goin'.
And even today: Think on the game/project Star Citizen ... they clearly would profit from more and more and more gigaflops .
reply
Arnab
AI processing is different from digital processes where it consumes more power to do simple calculations. Have heard about using RAM modules to store a said amount of electric charge then comparing them using machine learning algorithm to get an output. So instead of using binary signals, it is using each signals with different weightage. Say instead of going 001110101 it goes and adds the amount of charges in two separate modules to do simple calculations like determining the shape or colour of a simple object. If AMD can achieve AI processes as such with less amount of charges(obviously after changing the parameters) in those modules given that they have greater sensing and retention capacity, it would be a big leap. But the day when they can work close to the efficiency of biological brains is far off.
reply
AI processing is different from digital processes where it consumes more power to do simple calculations. Have heard about using RAM modules to store a said amount of electric charge then comparing them using machine learning algorithm to get an output. So instead of using binary signals, it is using each signals with different weightage. Say instead of going 001110101 it goes and adds the amount of charges in two separate modules to do simple calculations like determining the shape or colour of a simple object. If AMD can achieve AI processes as such with less amount of charges(obviously after changing the parameters) in those modules given that they have greater sensing and retention capacity, it would be a big leap. But the day when they can work close to the efficiency of biological brains is far off.
reply
Zegar
15:15 - AI performance would probably mean basically floating point computation efficiency. This is stuff normally done on GPUs rather than CPUs so I guess this architecture, similarly like NVIDIA's competing products are basically dedicated chips optimised for energy efficient floating point calculations. From my point of view that's interesting as a potential password breaker - hash reverser. Normally when in the process of penetration testing you encounter a hash you'd like to reverse ( break ) you resort to tools like hashcat running on one or many GPUs. I would be interested to see how CDNA compete against 3090 in this field...
reply
15:15 - AI performance would probably mean basically floating point computation efficiency. This is stuff normally done on GPUs rather than CPUs so I guess this architecture, similarly like NVIDIA's competing products are basically dedicated chips optimised for energy efficient floating point calculations. From my point of view that's interesting as a potential password breaker - hash reverser. Normally when in the process of penetration testing you encounter a hash you'd like to reverse ( break ) you resort to tools like hashcat running on one or many GPUs. I would be interested to see how CDNA compete against 3090 in this field...
reply
Sylvania
they have a working driver for RDNA3, Phoenix, and Strix rn... not sure what the retail Desktop APUs gonna use yet.. hw spec wise the RDNA3 is capable handling double the primitive than RDNA1/2 does as its seen on how the driver is configured for it.
My Take on it is that its should hover around 1.5x - 2x performance compared to prev gen (RDNA2), as for AI Learning.. its most likely bfloat16 or plain fp16 which are also used on Modern APIs for games..
Edit: on some research, higher precision is used instead but bruteforce through lower precision is also a thing
reply
they have a working driver for RDNA3, Phoenix, and Strix rn... not sure what the retail Desktop APUs gonna use yet.. hw spec wise the RDNA3 is capable handling double the primitive than RDNA1/2 does as its seen on how the driver is configured for it.
My Take on it is that its should hover around 1.5x - 2x performance compared to prev gen (RDNA2), as for AI Learning.. its most likely bfloat16 or plain fp16 which are also used on Modern APIs for games..
Edit: on some research, higher precision is used instead but bruteforce through lower precision is also a thing
reply
Daily
I am constantly distracted by the coasters and the crystal ball, I wish you would move them off the table Steve; it's clutter my shoulder demon is shouting See people love clutter, we should always have clutter! .
I don't think I have ever seen you put a drink on a coaster near to your work table. I feel like the table should have things you will need on there like a real work table. Good discipline is keep your food away from the work space, coasters encourage poor discipline Steve! Lol.
reply
I am constantly distracted by the coasters and the crystal ball, I wish you would move them off the table Steve; it's clutter my shoulder demon is shouting See people love clutter, we should always have clutter! .
I don't think I have ever seen you put a drink on a coaster near to your work table. I feel like the table should have things you will need on there like a real work table. Good discipline is keep your food away from the work space, coasters encourage poor discipline Steve! Lol.
reply
Melvin
Ai processing usually means matrix/vector multiplikation and addition. Usually with low precision. 32bit is the standard for floating point, Ai can use 16bit without too much/any sacrifices. There also exists 8bit floats now. But I'm sceptical If this still works. and I don't have hardware that is new enough to support 8bit floats so I can't test it.
If cdna 3 jede 8bit floats and they compare it to cdna2 with 32 bit floats you have an increase of x4 right there
reply
Ai processing usually means matrix/vector multiplikation and addition. Usually with low precision. 32bit is the standard for floating point, Ai can use 16bit without too much/any sacrifices. There also exists 8bit floats now. But I'm sceptical If this still works. and I don't have hardware that is new enough to support 8bit floats so I can't test it.
If cdna 3 jede 8bit floats and they compare it to cdna2 with 32 bit floats you have an increase of x4 right there
reply
CMDR
Those guys inbetween that needs the PCI-E lanes
Yup! I am one of those guys and it sucks!
What do I need PCI-E lanes for you ask? CONTROLLERS! RAID / IT controllers for storage.
While harddrives do the bulk storage, having ports for SSDs is very beneficial, but the only way to achieve this now is port expanders... But that is like sharing a SATA interface between several drives and not a good long term solution!
reply
Those guys inbetween that needs the PCI-E lanes
Yup! I am one of those guys and it sucks!
What do I need PCI-E lanes for you ask? CONTROLLERS! RAID / IT controllers for storage.
While harddrives do the bulk storage, having ports for SSDs is very beneficial, but the only way to achieve this now is port expanders... But that is like sharing a SATA interface between several drives and not a good long term solution!
reply
Add a review, comment
Other channel videos
