Why is the memory bandwidth only about half of what the new Tesla is offering? Does AMD receive bottom of the barrel HBM2 chips, despite being the developer, while Nvidia gets the best ones?

Why is the memory bandwidth only about half of what the new Tesla is offering? Does AMD receive bottom of the barrel HBM2 chips, despite being the developer, while Nvidia gets the best ones?

i think you should read a little about how HBM or gpu memory in general work and you wouldn't ask such questions nor make such assumptions.

i think you should read a little about how HBM or gpu memory in general work and you wouldn't ask such questions nor make such assumptions.

And how does Nvidia's HBM2 differ from AMD's? Enlighten us.

300W TDP Passive Cooling

What?!

What?!

While they are passivle cooled, industrial fans offer high amounts of airflow through the radiators.

What?!

Most servers and supercomputers are passive like that.

What?!

I would just wait for reviews to come then we know where it really stand.

I would just wait for reviews to come then we know where it really stand.

Hes saying "what" to the fact that they have pasive heatsinks that require chassis fans to cool, like most server CPUs and GPUs. These GPUs have some pretty interesting features. My guess is that the largest and most expensive will cost sigificantly less than the $10,000 for P100 and $13,000 for V100. The NVMe controllers are interesting too. 2TB of NVMe flash directly addressable by the GPU could be useful for a variety of things like analytics and checkpointing.

Hes saying "what" to the fact that they have pasive heatsinks that require chassis fans to cool, like most server CPUs and GPUs.

Petty much the same as saying What The **** which I shore you get. Just because it passive or active heat sinker is rated at 300 watt cooling capacity that doesn't means it use that with a Vega, Beside server room are usually very cooled room round about 22°C or 72°F but some an be bit lower or bit higher.

Petty much the same as saying What The **** which I shore you get. Just because it passive or active heat sinker is rated at 300 watt cooling capacity that doesn't means it use that with a Vega, Beside server room are usually very cooled room round about 22°C or 72°F but some an be bit lower or bit higher.

Yeah its just very common to have passive heatsinks for these huge GPUs. I think some people think "passively cooled by ambient air circulation" when they read passive, but in this case there would be a ton of airflow provided by lots of fans. Theyre just not attached to the GPU. In a lot of supercomputers the variable clock and TDP boosts are also disabled for more deterministic performance. Thats changing, so you could have a whole bunch of GPUs running at 300W easily. These accelerators go into supercomputers that consume MW of power.

i think you should read a little about how HBM or gpu memory in general work and you wouldn't ask such questions nor make such assumptions.

He dose have a point maybe you should read the specifications Vega PCIe 24.6 TFLOPS Half Precision (FP16) 12.3 TFLOPS Single Precision (FP32) 768 GFLOPS Double Precision (FP64) 484GB/sec Memory Bandwidth Tesla V100 PCIe 28 TFLOPS Half Precision (FP16) 14 TFLOPS Single Precision (FP32) 7 TFLOPS Double Precision (FP64) 900GB/sec Memory Bandwidth The big diff is in the die size 12nm FFN vs 16nm FFN which if you take in count for that if AMD had done to then it possible they could very well be the same over all. What more interesting is the Double Precision number

He dose have a point maybe you should read the specifications Vega PCIe 24.6 TFLOPS Half Precision (FP16) 12.3 TFLOPS Single Precision (FP32) 768 GFLOPS Double Precision (FP64) 484GB/sec Memory Bandwidth Tesla V100 PCIe 28 TFLOPS Half Precision (FP16) 14 TFLOPS Single Precision (FP32) 7 TFLOPS Double Precision (FP64) 900GB/sec Memory Bandwidth The big diff is in the die size 12nm FFN vs 16nm FFN which if you take in count for that if AMD had done to then it possible they could very well be the same over all. What more interesting is the Double Precision number

Vega is 14nm. 12nm is just a rebrand of an improved 16nm, the density is basically unchanged - but the dies are 800mm2, which is where most of the performance is coming from.

Vega is 14nm. 12nm is just a rebrand of an improved 16nm, the density is basically unchanged - but the dies are 800mm2, which is where most of the performance is coming from.

I was ref to the Tesla P100 die size to Tesla V100 as in if AMD had done same die 12nm then number could match the Tesla V100.

I was ref to the Tesla P100 die size to Tesla V100 as in if AMD had done same die 12nm then number could match the Tesla V100.

Ok then you were right on the first part but I still disagree that they could match V100. 12nm and 14nm are basically the same density. The process numbers themselves are essentially just made up at this point, it doesn't match their physical size anymore. Power consumption might be a little lower and switching speed may be a little faster but it's nothing significant. The real difference is that the reticles that TSMC is using for it's 12nm process are capable of going up to 800mm2. No other company is offering that at the moment - but it's also not something that's specific to 12nm. GF/Samsung could do a 14nm 800mm2 reticle as well. So basically Nvidia can and is building a 30% larger die than AMD is capable of building. AMD's solution to this, will most likely come with NAVI, is going to look more like Threadripper/EPYC design. They'll probably have multiple smaller dies tied together by a second generation infinity fabric. Which is way more cost effective than one large monolithic design. At least that's what I think they'll do. The other difference, at least with the single precision numbers specifically, is Nvidia's Tensor units. AMD doesn't have a comparable hardware feature on their cards, outside of their mixed math operations, which Nvidia has as well.

i think you should read a little about how HBM or gpu memory in general work and you wouldn't ask such questions nor make such assumptions.

Yeah, sure the latter half of my post was technically inappropriate, being nothing but a cheap jab, but the first half is a perfectly valid question. Which you answered with the oh-so-insightful "google it" answer, proving you don't know any more about it than I do. Both cards have the same amount of HBM2, yet the other's memory bandwidth is almost the double. I have no idea why if both cards share the same memory interface of 4096 bits.

AMD's solution to this, will most likely come with NAVI, is going to look more like Threadripper/EPYC design. They'll probably have multiple smaller dies tied together by a second generation infinity fabric. Which is way more cost effective than one large monolithic design. At least that's what I think they'll do

If AMD manages to achieve that in a way that the games/programs in general recognize it as a single GPU and not some kind of crossfire, they will really change the GPU market

Yeah, sure the latter half of my post was technically inappropriate, being nothing but a cheap jab, but the first half is a perfectly valid question. Which you answered with the oh-so-insightful "google it" answer, proving you don't know any more about it than I do. Both cards have the same amount of HBM2, yet the other's memory bandwidth is almost the double. I have no idea why if both cards share the same memory interface of 4096 bits.

There aren't that many factors involved for bandwith computation except Bus Width: frequency.

There aren't that many factors involved for bandwith computation except Bus Width: frequency.

Yeah, that's what I thought, thus wondering if Nvidia is buying all the chips that can take a higher frequency, leaving lesser samples for AMD. But labidas suggested that's not the answer. So, I was left wondering if I'm missing something essential.

yet the other's memory bandwidth is almost the double. I have no idea why if both cards share the same memory interface of 4096 bits.

I think that may have something to do with memory controller it self on how it is interface with GPU and rest of the sub system pipe line.

I think that may have some to do with memory controller it on how it is interface with GPU and rest of the sub system pipe line.

I think so too, additionally one could question whether the architecture profits from higher frequencies in a way that would be worth the drawbacks from them. I don't think this is yield related since the difference is just too big (~950 MHz vs. ~1.4 GHz). But we won't know until the Boss gets his hands on one of these babies 😉