News Raja Koduri Possibly Hints at June 2020 Intel Xe Graphics Card Release

[kinggremlin]

It just said "next generation", which it is, with respect to GDDR5. Within that market segment, it is the next generation after what Polaris used.

They were probably planning on this which was announced at Hot Chips a few months before the above roadmap.

AMD likely thought HBM would have gained a much better adoption rate than it did. Nvidia never bothered with it for any of their gaming GPU's which killed any chance for decreased costs by economy of scale. HBM2E was announced last year. Still waiting on that one. HBM3, the proposed cheap HBM, that was originally scheduled for a 2019/2020 release isn't even in the rumor mill anymore.

 

[bit_user]

They were probably planning on this which was announced at Hot Chips a few months before the above roadmap.

<slide/>

AMD likely thought HBM would have gained a much better adoption rate than it did. Nvidia never bothered with it for any of their gaming GPU's which killed any chance for decreased costs by economy of scale. HBM2E was announced last year. Still waiting on that one. HBM3, the proposed cheap HBM, that was originally scheduled for a 2019/2020 release isn't even in the rumor mill anymore.

It's funny that you say low-demand killed HBMx for mainstream applications, since this states there was recently higher-than-expected demand:

https://www.anandtech.com/show/11643/samsung-increases-8gb-hbm2-production-volume

Granted, that was 2 years ago, but probably around the time the decision was finalized to go with GDDR6 for RX 5700.

Since then, HBM2 has been getting a lot of use in high-end FPGAs, HPC chips, and AI chips. Even if mainstream GPUs aren't using it, that doesn't mean demand has fallen off.

 

[kinggremlin]

It's funny that you say low-demand killed HBMx for mainstream applications, since this states there was recently higher-than-expected demand:

https://www.anandtech.com/show/11643/samsung-increases-8gb-hbm2-production-volume

Granted, that was 2 years ago, but probably around the time the decision was finalized to go with GDDR6 for RX 5700.

Since then, HBM2 has been getting a lot of use in high-end FPGAs, HPC chips, and AI chips. Even if mainstream GPUs aren't using it, that doesn't mean demand has fallen off.

I didn't say it killed HBM, I said the lower adoption rate prevented it from dropping in price. Which in turn, has prevented it from making any in roads into mainstream gaming. That article you linked to was published during the peak of the cryptocurrency mining boom when you couldn't find a video card. I'm sure demand was off the charts at that point. The market is not remotely like that any more. There are currently no gaming cards on the market that utilize HBM. The last one was the Radeon VII which I think 8 gamers bought. No, I don't mean that literally, I don't need a link saying there were actually 12 gamers that bought a VII. HBM2 has higher peak bandwidth, considerably lower power usage and requires less board space to implement. If it was remotely cost competitive with GDDR, we would see in more consumer targeted products, at the very least in the highend gamer's market.

 

[bit_user]

I didn't say it killed HBM, I said the lower adoption rate prevented it from dropping in price. Which in turn, has prevented it from making any in roads into mainstream gaming.

Yeah, that's essentially the same as saying it "low-demand killed HBMx for mainstream applications".

That article you linked to was published during the peak of the cryptocurrency mining boom when you couldn't find a video card.

Dude, it was published before any gaming GPUs with HBM2 existed. So, no, it cannot have been meaningfully influenced by the crypto boom.

I also don't think Vega quite launched at peak crypto, or else the MSRP wouldn't have been so low. They were awfully worried about how it stacked up against the GTX 1080, and those were still going for around $500, at the time.

The last one was the Radeon VII which I think 8 gamers bought. No, I don't mean that literally, I don't need a link saying there were actually 12 gamers that bought a VII.

Awww, you are a gremlin!

I don't really count Radeon VII, since Vega 20 wasn't originally made for gaming. That said, I'm sure they're not selling them at a loss.

HBM2 has higher peak bandwidth, considerably lower power usage and requires less board space to implement. If it was remotely cost competitive with GDDR, we would see in more consumer targeted products, at the very least in the highend gamer's market.

I get the benefits. I want to see it go mainstream, too. I just don't think you have enough to go on that AMD somehow changed their plans to use GDDR6, after releasing that roadmap.

GDDR6 is the generation after GDDR5, and Navi is their first GPU to use it. Just because you don't like it doesn't mean it's fake news!

Anyway, this is all just a pointless digression from "Big" Navi. Getting back to that, I think one argument against chiplets is the amount of duplication you'd have, if they took a normal Navi die and doubled them. You'd have 2 compression engines, 2 display drivers, and probably some other stuff that doesn't need to get doubled. So, if their strategy was to go chiplets, then the RX 5700 would already have 2 dies and Big Navi would have 3.

I also don't foresee them itching to do another card with 512-bit memory bus, meaning they'd be disabling 1/4 of the memory controllers.