The relevance of Zen 3/3D systems and their DDR4 RAM...

[Thinder]

There is a good reason why I chose to remain with the DDR4 standard, and it's the Zen3/3D Infinity Fabric design and the fact that they were conceived for lower latency, even more so the 3D series.

I don't know how much weight you'll give to this video but it seems to me that in some ways it validate my choice, I keept saying it; DDR5 is not ready for the AMD processors, they need lower Cl to take full advantage of their cache...

From my PoV we will need to wait until some manufacturer launch a DDR5 RAM kit without the stability issue coming with lower latency, higher frequencies doesn't always cut it and before we see such RAM kit in the market, my DDR4 system will have served me well for some years.

 

[SkateZilla]

it's relevant globally,

You can Run them chips at 8,000MHz DDR5,
But it don't mean squat if the CAS Latency is at 20 CL, Meaning the Ram is only Accessed every 20 Clock Cycles...

You can likely take that Same DDR5-6000 CL38 Stick Drop it to 6000Mhz CL30 and get significantly better performance.

ie:
Latency time (ns) = CL x [1000 ÷ (CLOCK CYCLES)] ≒ Delay in NS

So :
Latency time (ns) = 30 x [1000 ÷ (6000 ÷ 2)] ≒ 9.99 ns

Latency time (ns) = 38 x [1000 ÷ (6000 ÷ 2)] ≒ 12.66 ns

 

Edited March 15, 2023 by SkateZilla

[BitMaster]

Not only that, nothing of their greatnesses has become true for mortal Desktop users yet. Bigger modules, lower voltage.

[Thinder]

On 3/15/2023 at 10:07 AM, SkateZilla said:

it's relevant globally,

You can Run them chips at 8,000MHz DDR5,
But it don't mean squat if the CAS Latency is at 20 CL, Meaning the Ram is only Accessed every 20 Clock Cycles...

You can likely take that Same DDR5-6000 CL38 Stick Drop it to 6000Mhz CL30 and get significantly better performance.

ie:
Latency time (ns) = CL x [1000 ÷ (CLOCK CYCLES)] ≒ Delay in NS

So :
Latency time (ns) = 30 x [1000 ÷ (6000 ÷ 2)] ≒ 9.99 ns

Latency time (ns) = 38 x [1000 ÷ (6000 ÷ 2)] ≒ 12.66 ns

 

 

Thank you guys for your insights..

Sorry for my late reply, I've been struggling with a failing motherboard, will get a new Asus TUF X570 within a couple of weeks, the MSI mobo have been a source of problems for a long time now, the PCI_E2 slots is faster than PCI_E1 at 4K and it shouldn't be the case, as much as I was happy with the B450 Gaming, this one is garbage.

Anyway, the rest of the PC performs very well, something I learned is that the 7 5800X 3D doesn't seems to have the same limitations with the number of ranks than the 5600X it replace, it might be to the marginal gain obtained with interleaving ( 4 X 1 rank), since that with double the capacity, the amount of data available is high enough to compensate or the controller is also different and deal with more than 4 ranks under load.

Obviously with this sort of RAM, the problem is the cost, the premium is significant compared to a non-B.die kit but the quality and performances are there, so I can't really complain especially because it works wonder with the 7 5800X 3D cache.

Now if anyone wonder if I regret my decision of not going for DDR5, the answer is a definitive NO, from my PoV the technology involved in the DDR5 die, its stability and the latency needed to squeeze the max performances out of a Ryzen 3D are simply not there yet, it might on the other hand, suit Intel CPUs much better (familiar pattern).

So this system is gonna do for the time being until hopefully some manufacturer comes up with a DDR5-6000 (as example) with half the latency they have today or lower (we can dream and add those on our X-Mass list)...

 

[BitMaster]

Imho, DDR5 as of now mainly serves the enterprise needs where i.e. multiple NVMe's do overwhelm the RAM bandwidth meanwhile.

If your RAM is slower than your PCIe lanes can pump data into it you have a new bottleneck, RAM bandwidth.

With DDR5, 8 (Intel) or 12 (AMD) channels, Bandwidth is not a concern as of now. That's where DDR5 really shines, already today.

Bandwidth Bandwidth BANDWIDTH.... like developers developers DEVELOPERS back in the days...LoL.

 

[Thinder]

3 hours ago, BitMaster said:

Imho, DDR5 as of now mainly serves the enterprise needs where i.e. multiple NVMe's do overwhelm the RAM bandwidth meanwhile.

If your RAM is slower than your PCIe lanes can pump data into it you have a new bottleneck, RAM bandwidth.

With DDR5, 8 (Intel) or 12 (AMD) channels, Bandwidth is not a concern as of now. That's where DDR5 really shines, already today.

Bandwidth Bandwidth BANDWIDTH.... like developers developers DEVELOPERS back in the days...LoL.

 

True but bandwidth is not everything, as SkateZilla pointed out, the number if time your RAM is accessed by your GPU matters and high  latency remains a bottleneck, even more so for a Ryzen CPU.

Ryzen 3/4 cache weren't originally designed for lower latencies and it matter little if AMD increased bandwidth with their new CPU design, they still will miss out on this capability as long as RAM manufacturers will not produce high frequency/low latency sticks as was the case for Zen 3s and Cl14.

In fact, the situation is like not having B.Die RAM today, the Zen 3/3D wouldn't be able to produce the performances they can with Cl14, and under load (4K) it matters a lot, so I felt like I was already taking a lot of risks with a new GPU without having to pay the premium for a new socket/CPU and RAM, they can do their R&D without my money...

Having said that I took the risk of going for a mediocre motherboard design with unstable BIOS, nobody is perfect.

 

[BitMaster]

Sure, regarding gaming you are right. Just that the IT industry has more customers than just us gamers. What matters for them is volume and I could imagine that DC's buy way more DIMMs per year than Gamers do, so the focus is on the enterprise side of things.With 100, 200, 400 and soon 800Gbit/sec networks you need systems that can fetch and store that much data on the fly and current DDR4 systems have come to their limit.

It's not only about Ryzen, VR and DCS.

 

[Thinder]

On 3/22/2023 at 10:34 AM, BitMaster said:

Sure, regarding gaming you are right. Just that the IT industry has more customers than just us gamers. What matters for them is volume and I could imagine that DC's buy way more DIMMs per year than Gamers do, so the focus is on the enterprise side of things.With 100, 200, 400 and soon 800Gbit/sec networks you need systems that can fetch and store that much data on the fly and current DDR4 systems have come to their limit.

It's not only about Ryzen, VR and DCS.

 

Agreed, your analysis is spot-on.

But if I remember well, B.die wasn't necessarily designed for the industry, I used to work as aa CG Techie, using 3D Studio Max, CATIA, (a Dassault-Systems packages for Industrial CAD design) and Fluent for fluid simulation, I never heared of B.Die when all the gamers were using it to O.C their RAM to get the frequencies the Ryzen and Intels RAM didn't support.

If anything we wouldn't have even consider using it simply because we would go for the most stable and safe system as possible rather than O.Cing anything in our system (we built our PC in house btw) and in the case of 3DS Max, lowering the cost was a major goal.

Imagine, from a £35..000/seat Silicon Graphics machine to a network of PII with 256Mb of RAM each, allowing techies to work at different aspects of a project then network-render for a fraction of the price, it was a revolution...

Gaming has another advantage looking at it this way, it pushes the boundaries of technologies and support the top end of R&D for this particular field, it was for gamers that Silicon Graphics decided to go from the specialized Aviation/Military niche to mainstream because they figured they would largely contribute to R&D finances, that gave you all the advances in GPUs and RAM and ended up with simulation at Squadron levels on PC COTS platforms as used by the French Air Force (RAZBAM M-2000).

If it haven't been for us, buying unproven new, more and more competitive GPUs from the local shops, their R&D costs would have stayed as high as they were and it was a very expensive business at the time.

So we can take some credit for the advances in technology we've seen since the mid-80s, while keeping in mind that manufacturers tend to abuse this and get us to pay for premiums just for the stake of lowering their R&D costs with equipment that aren't always worth it, I doubt very much that a company like Dassault-Aviation or Lockheed Martin are going to call Corsair or G.skill for an upgrade of their systems, even their recommended systems for their students aren't coming anywhere close to gaming standards.

Edited March 23, 2023 by Thinder