News Micron to Introduce GDDR7 Memory in 1H 2024

[Admin]

Micron expects to produce GDDR7 memory devices in the first half of 2024.

Micron to Introduce GDDR7 Memory in 1H 2024 : Read more

 

[-Fran-]

I'm pretty sure Intel, AMD and nVidia wanted this in late 2022 instead.

Still, maybe we could see a mid-refresh with GDDR7? Maybe?

Regards.

 

[InvalidError]

Normal memory is 2-3W per DIMM, GDDR6X is ~4W per chip, wonder how much power GDDR7 is going to guzzle at 32+Gbps.

At some point, HBM is going to become necessary just to keep the IO power in check.

 

[Deleted member 2731765]

Usage of this next-gen mem type in 'consumer' products seems like a pipe dream for now.

Currently NVIDIA provides the fastest memory solution with it's RTX 40 series GPUs in the form of the GDDR6X which provides up to 22 Gbps pin speeds, and AMD's Radeon RX 7000 series cards on the other hand utilize the standard 20 Gbps GDDR6 tech.

So let's assume GDDR7 would deliver the following bandwidth figures, based on 36 Gbps pin speeds:

• 128-bit @ 36 Gbps: 576 GB/s
• 192-bit @ 36 Gbps: 846 GB/s
• 256-bit @ 36 Gbps: 1152 GB/s
• 320-bit @ 36 Gbps: 1440 GB/s
• 384-bit @ 36 Gbps: 1728 GB/s

 

[gg83]

Usage of this next-gen mem type in 'consumer' products seems like a pipe dream for now.

Currently NVIDIA provides the fastest memory solution with it's RTX 40 series GPUs in the form of the GDDR6X which provides up to 22 Gbps pin speeds, and AMD's Radeon RX 7000 series cards on the other hand utilize the standard 20 Gbps GDDR6 tech.

So let's assume GDDR7 would deliver the following bandwidth figures, based on 36 Gbps pin speeds:

• 128-bit @ 36 Gbps: 576 GB/s
• 192-bit @ 36 Gbps: 846 GB/s
• 256-bit @ 36 Gbps: 1152 GB/s
• 320-bit @ 36 Gbps: 1440 GB/s
• 384-bit @ 36 Gbps: 1728 GB/s

Can you also explain the PAM3 vs PAM4? Is it for power management? Why would ddr6x use pam4 and ddr7 use pam3?

 

[InvalidError]

Can you also explain the PAM3 vs PAM4? Is it for power management? Why would ddr6x use pam4 and ddr7 use pam3?

The rollback from PAM4 to PAM3 was almost certainly because they analyzed results from GDDR6X and concluded that timing jitter and phase noises from pushing higher clocks with PAM3 would be easier to deal with than keeping signals clean enough for PAM4.

 

[YouFilthyHippo]

If Jeedeedeeareseven comes out in the first half of 2024, when does Jeedeedeeareseveneques come out? Any chance we will see that on the 5090Ti?

 

[FunSurfer]

If Jeedeedeeareseven comes out in the first half of 2024, when does Jeedeedeeareseveneques come out? Any chance we will see that on the 5090Ti?

RTX 5000 set for spring 2025 release, so the timing is good for GDDR7 on the cards. As for individual cards:
RTX 5090 Ti chance of sporting GDDR7: 100%
RTX 5090 chance of sporting GDDR7: 100%
RTX 5080 Ti chance of sporting GDDR7: 100%
RTX 5080 chance of sporting GDDR7: 50%
RTX 5070 Ti chance of sporting GDDR7: 50%
RTX 5070 chance of sporting GDDR7: 10%
RTX 5060 Ti chance of sporting GDDR7: 0%
RTX 5060 chance of sporting GDDR7: 0%

 

[bit_user]

As for individual cards:
...
RTX 5080 chance of sporting GDDR7: 50%
RTX 5070 Ti chance of sporting GDDR7: 50%
RTX 5070 chance of sporting GDDR7: 10%
RTX 5060 Ti chance of sporting GDDR7: 0%
RTX 5060 chance of sporting GDDR7: 0%

Why? We've seen how Nvidia is keen on cutting the number of chips and memory channels, on lower-end models. Faster RAM would let them do even more of that.

 

[thestryker]

Normal memory is 2-3W per DIMM, GDDR6X is ~4W per chip, wonder how much power GDDR7 is going to guzzle at 32+Gbps.

At some point, HBM is going to become necessary just to keep the IO power in check.

GDDR6 @ 20gbps is ~4.8W and GDDR6X @ 20gbps is ~4.65W (21gbps is ~4.87W)

I'm really curious how much they've been able to drop the power consumption on GDDR7, because the improvement from GDDR5X to GDDR6X was pretty small. If the difference between GDDR6X and GDDR7 matches that you'd be looking at ~6.65W per chip. The real question client side becomes whether or not higher capacities are available which is something I haven't seen in any documents or press releases yet. If they're able to launch 32Gb capacity that would mean client side they could drop bus width down fairly significantly without losing memory bandwidth over what is available today so long as they use higher end chips.

 

[InvalidError]

Why? We've seen how Nvidia is keen on cutting the number of chips and memory channels, on lower-end models. Faster RAM would let them do even more of that.

Reviews bashing Nvidia and AMD to a lesser degree for insufficient VRAM and VRAM bandwidth might motivate them to rethink their bus shrinkage strategy. Current-gen stuff often runs into scaling issues previous-gen stuff doesn't have when dipping below 256bits/16GB. If next-gen bumps performance up ~50% for a given marketing tier like the 5090 rumors say it might, trimming bus some more with GDDR7 won't be an option. At best, it might make it possible to keep things the same as they are now.

I haven't seen in any documents or press releases yet. If they're able to launch 32Gb capacity that would mean client side they could drop bus width down fairly significantly without losing memory bandwidth over what is available today so long as they use higher end chips.

We already have 24Gbits DDR5, 24Gbits GDDR7 should be quite easily feasible. As mentioned in my bit reply above though, I don't think further memory interface trim will be possible as they'll need the extra bandwidth to feed the higher performance cores required to avoid having another "lowest sales in 20+ years" generation on their hands.

 

[bit_user]

As mentioned in my bit reply above though, I don't think further memory interface trim will be possible as they'll need the extra bandwidth to feed the higher performance cores

I was being a bit silly to suggest they would trim interfaces even further, but if they wish to scale performance while at least keeping interface widths the same, then they'll definitely need higher-performance memory (as you say).

So, it sounds like we agree that the lower-tier models could still be viable targets for GDDR7?

 

[InvalidError]

So, it sounds like we agree that the lower-tier models could still be viable targets for GDDR7?

Depends on how much of a premium GDDR7 is going to carry at least initially. Just like DDR5 and GDDR6X, the premium might be too steep for entry-level during the first year and change.

 

[thestryker]

We already have 24Gbits DDR5, 24Gbits GDDR7 should be quite easily feasible.

I sure wouldn't be that confident seeing as we still don't have over 16Gb GDDR6 despite enterprise needing massive VRAM capacity and the spec allowing for 32Gb. I certainly hope the GDDR7 spec has been designed around bigger capacity, but I wouldn't view it as a forgone conclusion.

As mentioned in my bit reply above though, I don't think further memory interface trim will be possible as they'll need the extra bandwidth to feed the higher performance cores required to avoid having another "lowest sales in 20+ years" generation on their hands.

I think it's going to depend on the silicon cost between cache and memory channels. If there's 32Gb capacity they could drop to a 96 bit bus and still have more memory bandwidth than now on 128 bit with 12GB capacity and the lowest end card could drop to 64 bit while being very close to the same bandwidth as 128 bit now and maintain 8GB.

At this point I could really see it going either way, and pricing will probably determine it all.

 

[InvalidError]

I sure wouldn't be that confident seeing as we still don't have over 16Gb GDDR6 despite enterprise needing massive VRAM capacity and the spec allowing for 32Gb.

While the spec may technically allow larger sizes, the command word structure and the way the DRAM matrix operates may not allow exercising that extra hypothetical flexibility due to issues with making memory array rows wider or sense lines longer. DRAM manufacturers would likely prefer to have an extra bank address bit or two.

 

[FunSurfer]

Why? We've seen how Nvidia is keen on cutting the number of chips and memory channels, on lower-end models. Faster RAM would let them do even more of that.

So where is the RTX 4060 with 21, 22.4 or 24 Gbps VRAM?

 

[bit_user]

So where is the RTX 4060 with 21, 22.4 or 24 Gbps VRAM?

Maybe they spent all the remaining budget on extra L2 cache?

 

[thestryker]

So where is the RTX 4060 with 21, 22.4 or 24 Gbps VRAM?

I think you're missing the point entirely. If GDDR7 has 32Gb chips that means nvidia could take a hypothetical 4060, cut it to a 64 bit bus thus using less silicon for the chip and only need to wire up 2 memory chips while maintaining the same capacity and bandwidth. This could be cheaper than using slower memory on a 128 bit bus and if it is you can be sure it's something nvidia would look at doing.

 

[InvalidError]

This could be cheaper than using slower memory on a 128 bit bus and if it is you can be sure it's something nvidia would look at doing.

I doubt GDDR7 will be anywhere as affordable as GDDR6 for the first year or two from launch. Much like DDR5, the 2nd or 3rd wave of GDDR7 chips will likely have much better timings and clock frequencies than the first-gen stuff too for a lower price. Also, if we get 32Gbits chips, they'll likely be at a steep premium for the first couple of years on top of that. I doubt 2x32Gbits GDDR7 chips will be viable at the entry-level for the first 2-3 years from such chips becoming available.

 

[thestryker]

I doubt GDDR7 will be anywhere as affordable as GDDR6 for the first year or two from launch. Much like DDR5, the 2nd or 3rd wave of GDDR7 chips will likely have much better timings and clock frequencies than the first-gen stuff too for a lower price. Also, if we get 32Gbits chips, they'll likely be at a steep premium for the first couple of years on top of that. I doubt 2x32Gbits GDDR7 chips will be viable at the entry-level for the first 2-3 years from such chips becoming available.

Oh the memory itself absolutely won't be cheaper at all. It would entirely come down to the silicon savings of getting more chips per wafer and lowering board complexity and whether those savings at scale make sense to pair with the more expensive memory. It is also dependent on 32Gb being widely available which I doubt it would be even if it launched with GDDR7 due to enterprise demands. Of course if the 2025 rumor for nvidia's next consumer architecture is accurate that could be far enough out since lower end tend to be the last in line.

 

[FunSurfer]

I think you're missing the point entirely. If GDDR7 has 32Gb chips that means nvidia could take a hypothetical 4060, cut it to a 64 bit bus thus using less silicon for the chip and only need to wire up 2 memory chips while maintaining the same capacity and bandwidth. This could be cheaper than using slower memory on a 128 bit bus and if it is you can be sure it's something nvidia would look at doing.

RTX 5060 16GB GDDR6 24Gbps with 128 bit bus will have more memory bandwidth that RTX 5060 16GB GDDR7 36Gbps with 64 bit bus. But according to what you say there is a good chance for an RTX 5090 32GB GDDR7 36Gbps with 256 bit bus.