Question Samsung HBM2E Flashbolt Launches

[soresu]

Samsung's 3rd gen HBM2 chips just launched. Link here.

According to the article it can achieve up to 4.2 Gbps per pin, or 537.6 Gbps per stack - up from the 3.6 Gbps per pin claim back in august last year.

With that much per pin, even a single stack would have plenty bandwidth and capacity for most common applications.

 

[Arkaign]

The numbers themselves are impressive as usual. I wonder why HBM has been such a bust in consumer GPU :/ I remember the initial hype (Fury 4GB HBM1?), then that turned out to be disappointing. Then extreme hype for Vega HBM2, which while not a complete bust after price drops, definitely didn't show compelling upgrades from the competition.

It's maddening, because it seems like something that should be almost like the leap from HDD to SSD, but in practice we've not seen that yet? Poor implementation?

 

[joesiv]

The numbers themselves are impressive as usual. I wonder why HBM has been such a bust in consumer GPU :/ I remember the initial hype (Fury 4GB HBM1?), then that turned out to be disappointing. Then extreme hype for Vega HBM2, which while not a complete bust after price drops, definitely didn't show compelling upgrades from the competition.

It's maddening, because it seems like something that should be almost like the leap from HDD to SSD, but in practice we've not seen that yet? Poor implementation?

As far as I could tell it was the cost of parts, costs/complexity of implementation (2.5D packaging). The parts costs was high due to low volume and few manufacturings doing it. The packaging that AMD used was cutting edge at the time, and thus expensive to do across the board.

However, it seems that with more players in the HBM market, and advanced packaging techniques becoming the norm, perhaps HBM could get more use. It seems like it's getting quite a bit of traction in the non-consumer and high end workstation market.

I think the cost per GB is still higher than conventional GDDR, but with the power savings, it seems that if the margins are good enough (high end products) it's still a great option.

 

[aigomorla]

I wonder why HBM has been such a bust in consumer GPU

i heard its awesome in computational and workload applications.
Its horrible in gaming tho.

Consumers are mostly focused on gaming, where most of the target audience that typically buys over 700 dollar cards are gamers.
With Bitcoin and mining declining, you lost about 1/3rd of workload sales, leaving them left to content creation, and design.

Hence why it does not handle very well on the consumer side.

Oh but i also wonder how many of the sales also reflects what apple gobbles up, as apple uses AMD gpu's and have ditched Nvidia altogether.

 

[Mopetar]

i heard its awesome in computational and workload applications.
Its horrible in gaming tho.

What about it would be horrible for gaming? The reason it hasn't been seen on consumer cards comes down to price. The interface is different so you specifically have to design your chip with HBM in mind and given the high costs and limited availability it doesn't make sense from the perspective of a consumer product, especially a mass market one.

However, I've never heard anyone claim that it would be bad for gaming for any technical reasons. It hasn't been used in terribly many consumer products for the reasons outlined above, and the AMD cards that it's been most prominently featured in weren't performing poorly due to using HBM as oppossed to GDDR memory. One could probably make a compelling argument that the cards like Fury and Vega would have done even worse in those circumstances.

It obviously has excellent bandwidth and for top-end cards that are already expensive it's not too much of a problem to use additional stacks to ensure that the amount of memory isn't a limiting factor either. Then again, I've not done any deep research into this myself, and could well be wrong, but I don't recall ever seeing someone make a statement about it being "horrible" let alone "not quite as good" in gaming workloads.

 

[Stuka87]

Only reason HBM isn't big in the consumer market is cost. HBM is superior in almost every metric except for cost. Not only are the raw memory chips more expensive, but it also drives up the costs of the final GPU package. And it prevents card builders from possibly getting deals on memory themselves. They are stuck with whatever pricing AMD/nVidia charge for the package.

 

[aigomorla]

What about it would be horrible for gaming? The reason it hasn't been seen on consumer cards comes down to price.

This is exactly why.
The amount of price to performance does not scale well.
This was also shown in the Fury V cards.

Nvidia also tried to play with it but, not every gamer is out in the market for a 3000 dollar Titan V, when you could get a 1080Ti which trailed behind, for 1/3rd the cost.

And again, even tho gamers like to splurg on videocard, they aren't many deep pocketed gamers as there are enterprise where you see HMB more spamed thoughout.

 

[Mopetar]

This is exactly why.
The amount of price to performance does not scale well.

That doesn't make it horrible for gaming though, it just makes it too expensive to use. One could argue that if it weren't as expensive, it would be the ideal choice for gaming. It's like arguing that using LN2 cooling on a GPU for gaming is horrible. It's the best cooling performance you can get, but it's just not financially feasible to use. But if it were financially feasible, everyone would be using it.

However, there is the question of how much return you get for the better performance. Using two stacks of this for ~1Tbps of memory bandwidth isn't going to give you a 100% performance uplift because most games don't have that bad of a bottleneck when it comes to memory. There are certainly a few, and more bandwidth would definitely help if you can't achieve good compression for whatever reason, but just because something is "better" doesn't mean that it's that much better than some other "good enough" solution.

Any new technology always has this kind of ramp up where it's not widely used enough to reach the low prices provided by an economy of scale and because it isn't cheap enough or available in wide quantities there isn't a massive push to adopt it. If NVidia wanted to they could switch over to HBM for all of their cards, but is there enough global supply of HBM to meet NVidia's demands? Probably not, which means that even if they wanted to switch for performance reasons, they probably couldn't.

It would be better for the world if everyone switched over to using electric vehicles. Not just because it's better for the environment, but because they're more efficient and have better performance characteristics in many ways. They're a better vehicle. But we can't do that yet because they're too costly, and even if cost were no objective, there's simply not enough supply to move everyone over to using one even within the next decade. But no one is going to call them horrible cars because they're too expensive.

 

[Stuka87]

This is exactly why.
The amount of price to performance does not scale well.
This was also shown in the Fury V cards.

Nvidia also tried to play with it but, not every gamer is out in the market for a 3000 dollar Titan V, when you could get a 1080Ti which trailed behind, for 1/3rd the cost.

And again, even tho gamers like to splurg on videocard, they aren't many deep pocketed gamers as there are enterprise where you see HMB more spamed thoughout.

The Fury card only existed as a means for AMD to push HBM because they co-invented it. The Fury cards were extremely good at compute, and the added bandwidth helped there. But they were not designed to be a gaming card. They were sold as a gaming card because AMD did not have anything else to fit into that segment of the market.

But this didn't mean HBM was bad for gaming. Its just not cost effective for gaming because gaming does not benefit from the extra bandwidth. Where as compute does, which is why you see things like Tesla cards use HBM.

But unless things change from what I mentioned in my first post, we won't see HBM in consumer cards.

 

[Mopetar]

Its just not cost effective for gaming because gaming does not benefit from the extra bandwidth. Where as compute does, which is why you see things like Tesla cards use HBM.

But unless things change from what I mentioned in my first post, we won't see HBM in consumer cards.

It does to some degree depending on the card or no one would bother running memory overclocks. For some cards like Polaris you could practically see more benefit from a memory OC than a core OC so there are cards that will benefit from more bandwidth, but there's a question to what extent that is true.

The value that HBM brings to consumer products is that it uses less power and that it simplifies the board layout, particularly in the case where it's a high-end card with a lot of memory. Take something like the 2080 Ti which has to make room for 11 different 1 GB memory chips. Here's an image for reference:

You could easily get a much more compact card by using HBM. Compare this with something like the Vega Nano:

While not everything has to be compacted to quite that degree, it certainly shows that there are some other obvious benefits to consumer cards beyond just the increased bandwidth.

 

[Guru]

its amazing for professional workloads, but it's just underutilized and too expensive for gaming and casual use.

 

[aigomorla]

I would love a 2080Ti half size.
Watercooling that bottom card would also be a Ton easier.

 

[AtenRa]

It is unfortunate that AMD was not in a financial state to create a Fury NANO successor at 14nm. There is still hope for a 7nm+ EUV NANO card.

 

[Ottonomous]

It is unfortunate that AMD was not in a financial state to create a Fury NANO successor at 14nm. There is still hope for a 7nm+ EUV NANO card.

Has any tech site reviewed the PowerColor 5700 ITX? Practically non-existent besides the inflated amazon.jp entry

 

[NTMBK]

It is unfortunate that AMD was not in a financial state to create a Fury NANO successor at 14nm. There is still hope for a 7nm+ EUV NANO card.

That presumes that the Fury Nano was popular enough to warrant a successor. I think most gamers don't care enough about small cards to pay a premium for them.

 

[guachi]

I put a Red Dragon 5700XT in the relatively small Silverstone SG13 case. I guess if it were half-height I could use a shorter case. But it's not like the SG13 is large to begin with.

 

[AtenRa]

Has any tech site reviewed the PowerColor 5700 ITX? Practically non-existent besides the inflated amazon.jp entry

This product is only sold in JP and no i havent seen any reviews.

 

[AtenRa]

That presumes that the Fury Nano was popular enough to warrant a successor. I think most gamers don't care enough about small cards to pay a premium for them.

The problem with Fury NANO was its small memory buffer for its category (High-End segment of 1440p and 4K), it only had 4GB memory when the competition (GTX980Ti) had 6GB.
For its TDP (175W) and its size it was the fastest GPU you could buy.

I hope we will see a new 7nm EUV RDNA2 HBM Nano card, it could really make a difference this time since it will come with a good Gaming mArch, 8GB HBM2 is not that expensive any more and the Card could be easily priced at $800 segment this time making it a lot easier to get ROI. Also from a marketing point, it could be a new product with no competition at its TDP and size making it something of a premium product.

 

[Ottonomous]

The problem with Fury NANO was its small memory buffer for its category (High-End segment of 1440p and 4K), it only had 4GB memory when the competition (GTX980Ti) had 6GB.
For its TDP (175W) and its size it was the fastest GPU you could buy.

I hope we will see a new 7nm EUV RDNA2 HBM Nano card, it could really make a difference this time since it will come with a good Gaming mArch, 8GB HBM2 is not that expensive any more and the Card could be easily priced at $800 segment this time making it a lot easier to get ROI. Also from a marketing point, it could be a new product with no competition at its TDP and size making it something of a premium product.

Fury would've been a really good beneficiary of HBCC, AMD did well to optimize for its memory usage, better than Polaris/Hawaii. Wouldn't RDNA2 require a costly rework for the HBM memory controller though? Don't think Apple would order navi hbm-based parts to subsidize it

 

[AtenRa]

Fury would've been a really good beneficiary of HBCC, AMD did well to optimize for its memory usage, better than Polaris/Hawaii. Wouldn't RDNA2 require a costly rework for the HBM memory controller though? Don't think Apple would order navi hbm-based parts to subsidize it

RDNA2 is the mArch, you can have whatever implementation you want. And since GPU prices have increased by 150-250 USD at each segment and 8GB of HBM 2 is not that much expensive now , AMD could easily make a $500-600 8GB card.

Example, i could see a 350-400mm2 RDNA2 chip with 8GB HBM2 that AMD could make 2-3 SKUs out of it at $500 to $800 segment.
Or they could only create a single very big 450-500mm2 chip and make 3 SKUs, one with 8GB HBM2 at $500-600 price point and two with 16GB at $800+