NVIDIA Unveils $2,499 USD TITAN RTX

[schmidtbag]

You have nice desktop APU (without increasing a cost) running on DDR4 and it is faster than ever faster than these with DDR3, same will happen once we switch to DDR5 and so on

At the time of DDR5, we would probably have HBM3 too and around 2022. to 2024. also HBM4 will probably appear

Maybe next-gen gaming consoles if they aim for an end of 2020. launch, will start with HBM3... who knows

DDR5 ought to be a big help to APUs since they really seem to be starving for memory bandwidth, though I think AMD might want to focus on implementing bigger L3 caches (whereas instead, they used a smaller one vs the non-APU parts).
I would agree that HBM3 is likely to come out around the same time as DDR5. Right now, nothing really saturates HBM2 but considering the future of GPUs, that might not be the case for long.
Also I may be wrong about this but I think PS5 is supposed to come out in 2019. They've pretty much already got their processor designed at this point, and I'm not sure if they're even using HBM2 (they should, though it might be too costly).

[dungeon]

APUs are actually more bounded by numbers of CPU cores, for many years they are just 4 cores max. On memory boundware DDR4 helped there, DCC too, etc... but this won't stay forever.

For future of APUs likely total redesign is needed, i guess next step would be 16 CPU cores APUs max - well, maybe on 7nm that starting to be possible

Also I may be wrong about this but I think PS5 is supposed to come out in 2019. They've pretty much already got their processor designed at this point, and I'm not sure if they're even using HBM2 (they should, though it might be too costly).

Only HBM3 will have price reduction, also maybe Samsung will feed them with something early 16 cores and HBM3, that should fly

[TemplarGR]

APUs are the future of desktop PCs, this is certain. They offer many advantages and eventually dGPUs will disappear from the market. The thing is, up until now the only company REALLY pushing for this have been AMD, and their efforts were very weak for various reasons.

APUs obviously need adequate memory bandwidth. This wasn't possible for AMD because the market wouldn't support any "high end" attempt using more DDR channels and HBM was expensive. So AMD APUs ended up being budget systems.

Things might change with low cost HBM samsung claims to developing now. APUs don't need 8 or 16 GBs of HBM. They just need 1, 2, or 4. Just as a L4 cache and framebuffer. Assuming this would be "low cost" variant of HBM, it wouldn't be very expensive. It could be viable. DDR5 will help even more.

If that happens, AMD could exploit 7nm, 7nm+ and 5nm processes, buff the TDP up and release monster-APUs. Remember, current Ryzen APUs on 14nm are 65W TDP chips and brandish 1.8 teraflops of performance. Imagine if they tripled that on a 7nm node, by additionally rising the TDP limit. Put on a DDR5 controller and some low cost HBM as a framebuffer, and you suddenly get an APU capable of full 1080p gaming on a TDP of around 125W. This would sell like hotcakes because Ryzen processors are in demand these days and GPUs are expensive.

If that catches on, AMD could even up the TDP more to 200W or even 250W. Why not? There is no "law" that demands consumer cpus to be restricted to 95W... As long as adequate cooling is provided and a suitable motherboard is produced, it is fine.

And after APUs go into the 200W and 250W range, they could even push for dual socket for even more performance...

Remember, cpu cores keep getting smaller and smaller, BUT, there is no point in delivering 32 or 64 cores in consumer/gamer chips. Those will never be exploited by desktop use cases, let's be real. It took them a decade to somewhat exploit 4 cores...

So as cpu cores become smaller and smaller, cpu manufacturers have the choice to either just lower desktop TDPs, increase mainstream core counts to insane numbers, or just turn their cpus into dGPUs with cpu cores attached that go into the motherboard socket... As cpu cores become smaller, the transistor budget allocated to them is getting smaller too...

This will also help push OpenCL and ROCm and gpgpu, since it will eliminate latency. Also remember that APUs have many efficiency benefits vs cpu + dgpu combo.

So, in the end, it is only a matter of time. Only poor memory bandwidth and market forces have prevented this change until now.

[dungeon]

APUs obviously need adequate memory bandwidth.

And they have adequate memory bandwidth, for 11 CUs max you don't need more.

Market wanna optimal increasment of everything really, but without increasment of price And time will fix that, so just wait

There is as low as 16 CU premium mobile with HBM2, but that is dGPU already of course These need it the most as laptops does not have so fast DDR4 available, so there it comes HBM to fix it - of course not at so affordable price since it is premium

[TemplarGR]

And they have adequate memory bandwidth, for 11 CUs max you don't need more.

Market wanna optimal increasment of everything really, but without increasment of price And time will fix that, so just wait

There is as low as 16 CU premium mobile with HBM2, but that is dGPU already of course These need it the most as laptops does not have so fast DDR4 available, so there it comes HBM to fix it - of course not at so affordable price since it is premium

This is not true. AGAIN, you are wrong. Why don't you just stop posting when you know nothing about a subject?

In order to understand if an APU has adequate bandwidth, you just have to find about comparable dGPUs and what bandwidth they have...

Ryzen 2400G has around 1.8Teraflops. This is 7850/7790/260X territory. Out of the top of my head, 260X, which i owned in the past, had around 105 GB/s of memory bandwidth. 7850 had a lot more than that.

On the other hand, depending on the clocks DDR4 delivers around 50 gb/s, which has to be shared with the cpu part. Of course, the VEGA architecture has some memory compression and other architectural improvements, but still those don't suffice to cover for the lost memory bandwidth.

So, yet again, i am right, you are wrong. Please, for the sake of your reputation, NEVER attempt to oppose my statements again. Contrary to you, i don't post about things i know nothing about.

[L_A_G]

What about double-precision float performance?

We are talking about the exact same silicon so if there are any differences in FP64 performance, their cause is obviously hobbled firmware.

[dungeon]

This is not true. AGAIN, you are wrong. Why don't you just stop posting when you know nothing about a subject?

I don't see why it is wrong . More CUs are better than more memory bandwidth, are you agree on that?

You actually wanna increase both together, not just one. So it does not make sense on low CU parts Just increase resolution enough and bandwidth dissapear as you are again more CU bound than anything else

So, yet again, i am right, you are wrong.

You are not wrong, but i am not wrong too Even Intel put a chunk of eDRAM on premium and high end parts only, but that is not very fast maybe like dual DDR4@4000 bandwidth or something like that, that is not like big GDDR5 still anyway and let alone smaller but still expensive HBM. So, there is no much difference, just comes with additional cost

It is not an optimal nor cheap solution, so that AMD could make it mandatory from the top to the bottom Could be an job for semi-custom solutions, not for mainstream line anyway where price matters

Please, for the sake of your reputation, NEVER attempt to oppose my statements again.

Please, don't put too much lipstick on a pig just because to ask for more $$

[ryao]

APUs are the future of desktop PCs, this is certain. They offer many advantages and eventually dGPUs will disappear from the market. The thing is, up until now the only company REALLY pushing for this have been AMD, and their efforts were very weak for various reasons.

SoCs are the future of personal computing, not APUs. The SoCs are simultaneously more energy efficient and cheaper. We can already see them doing well in mobile devices while early proofs of concept like the RPi are catching on for use as desktops.

Discrete GPUs will never disappear from the market as long as the current electron hole technology used to construct microprocessors is used in PCs. As long as that is the case, PC Gaming will always rely on discrete for the cutting edge and HPC would need them too. Things will likely change after the introduction of optical logical circuits. It becomes hard to predict what happens after that, but that is fairly far in the future.

Also, throwing away an entire machine instead of installing a discrete GPU when you need better graphics is insane. This is the case even for APUs because the memory bandwidth must scale with processing power. Otherwise, the render pipeline will stall. Cache can help mitigate this, but cache can only help so much unless you make it so arbitrarily large that you might as well put dedicated video memory on the APU. Space would be a problem there unless package sizes increase. Then if you are CPU bound, you get to throw away your really expensive on die/package graphics processor when upgrading and effectively pay for the same GPU twice.

[Jabberwocky]

I need one of these to play Minecraft?