News V-Color announces 2TB RDIMM kits for Threadripper Pro 9000 — 256GB modules promise stability at absurdly high RAM capacities

[Admin]

V-Color has announced its newest line of OC RDIMMs purpose-built for the Ryzen Threadripper 9000 Pro line, reaching up to 256GB on a single RDIMM. In some Pro motherboards, these kits are rated for up to 2TB total installed capacity.

V-Color announces 2TB RDIMM kits for Threadripper Pro 9000 — 256GB modules promise stability at absurdly high RAM capacities : Read more

 

[Geef]

The link you showed for the the NEMIX RAM 2TB (8X256GB) DDR4 3200MHZ is DDR4 on Newegg.
Threadripper Pro 9000 uses DDR5 memory so that memory from the story about will probably have a larger price difference when it comes out.

 

[dimar]

Just for a second I through it was RDRAM..

 

[John Nemesh]

So we can get up to 1TB on workstations now, but graphics cards are still shipping with 8GB on them. Nice.

 

[tamalero]

So we can get up to 1TB on workstations now, but graphics cards are still shipping with 8GB on them. Nice.

Well.. AMD tries to innovate and push forward.

 

[usertests]

So we can get up to 1TB on workstations now, but graphics cards are still shipping with 8GB on them. Nice.

If you can afford Threadripper and terabytes of RAM, surely you can open your wallet for:

RTX Pro 6000 Blackwell tested, performs roughly 10-15% faster than a stock RTX 5090

 

[bit_user]

Well.. AMD tries to innovate and push forward.

TBH, I'm having a little troubling deciphering the specs on the product page. The 256 GB DIMMs are listed as having:

IC/Rank: 16Gx4 / 8Rx4
​

Based on the specs of the lower-capacity DIMMs, I think that means 16 GiB/package = 128 Gib, which is a 4-high stack of 32 Gib dies. Then, I think the DIMM is 8-rank with 4 packages per rank.

This is only possible with registered DIMMs, so don't expect to see anything greater than 64 GB DIMMs for desktop boards anytime soon. The reason is that UDIMMs (i.e. regular, non-registered DIMMs) seem limited to a max of 2 ranks and I'm not aware of any examples of them incorporating die-stacked packages. Furthermore, I'm guessing the reason V-Color had to use so many ranks is precisely because that still works out to 16 chips per rank, which is maybe some kind of limit.

BTW, I'm really not sure how much credit AMD deserves for this, but the V-Color product page does indeed say the DIMMs are designed specifically for WRX90 boards.

https://v-color.net/products/ddr5-ocrdimm-amd-wrx90-workstationmemory
​

FWIW, I just ordered some V-Color memory last week. They seem very innovative in niches like ECC RAM, where I had previously been at the mercy of Kingston and Crucial, who've been moving at an absolutely glacial pace in the ECC UDIMM market.

 

[bit_user]

So we can get up to 1TB on workstations now, but graphics cards are still shipping with 8GB on them. Nice.

The cheapest V-Color kit that gets you to 1 TB (8x 128 GB DDR5-5600) costs $12k (see link in above post). If you want to go all the way to 2 TB, it'll cost a cool $24.7k.

For the 1 TB money, you could buy a RTX Pro 6000 with 96 GB of GDDR7 memory, as @usertests pointed out. For just a little more than the 2 TB price, you can get a H200 PCIe card with 141 GB of HBM3E:

https://networkoutlet.com/products/...en-5-0-pci-express-x16?variant=46805840298213
​

It's always good to read beyond the headlines and try to find out what's the catch.

 

[bit_user]

Just for a second I through it was RDRAM..

I think modern DRAM standards are utilizing pretty much all of the tricks that Rambus incorporated into its RDIMMs, and then some. Of course, once Rambus prevailed in its high-profile court cases against DDR memory makers, I'm sure they were only too happy to license these technologies to JEDEC.

 

[Kindaian]

Don't understand why at this age people create non-ecc memory for workstations / server setups.

 

[newtechldtech]

Don't understand why at this age people create non-ecc memory for workstations / server setups.

The memory itself is not created for "servers" , it is the board that accepts BOTH non-ecc and ECC .

 

[John Nemesh]

The cheapest V-Color kit that gets you to 1 TB (8x 128 GB DDR5-5600) costs $12k (see link in above post). If you want to go all the way to 2 TB, it'll cost a cool $24.7k.

For the 1 TB money, you could buy a RTX Pro 6000 with 96 GB of GDDR7 memory, as @usertests pointed out. For just a little more than the 2 TB price, you can get a H200 PCIe card with 141 GB of HBM3E:

​

https://networkoutlet.com/products/...en-5-0-pci-express-x16?variant=46805840298213​

​

It's always good to read beyond the headlines and try to find out what's the catch.

All of this is true...but it's also true that 8GB cards still exist...which was my point.

 

[John Nemesh]

If you can afford Threadripper and terabytes of RAM, surely you can open your wallet for:

RTX Pro 6000 Blackwell tested, performs roughly 10-15% faster than a stock RTX 5090

You completely and utterly missed my point.

 

[usertests]

You completely and utterly missed my point.

Your point was pointless. These modules are absurdly expensive, probably using stacked dies. GDDR memory is more expensive than DRAM per gigabyte. 128-bit bus limits manufacturers to 8 GB or 16 GB GDDR6, or soon 12 GB or 24 GB GDDR7. 8 GB cards are still fine for some users and games. There are people still using 4-6 GB cards. If you don't like it, don't buy it, and wait for the competition to start using 3 GB GDDR7 modules to put 12 GB at the low-end.

 

[bit_user]

All of this is true...but it's also true that 8GB cards still exist...which was my point.

It seems you're just now discovering the memory wall!

​

Note the distinct flattening over about the past decade. GB/$ was increasing exponentially, until then. Because the Y-axis is logarithmic, a fixed exponent shows up as a straight diagonal line.

So, one way to read what I said in my above posts is that the 1 TB configuration is achieved by way of combining 256 DDR5 dies, which contain 4 GiB (32 Gigabits) each. When DDR5 memory first started shipping, all of the dies were 16 Gigabits (2 Gigabytes), then came 24 Gib (3 GiB) and now 32 Gib (4 GiB) dies. So, there has been some density increase, but it's still not much improved. So, the only way to have lots of RAM is by using tons of chips. That's why it's so expensive.

GDDR6 memory is more complex, being highly optimized for speed, and therefore innately more expensive and slightly lower-density. It only comes in 16 Gigabits (2 GiB) capacity, and the highest number of dies you can address on the same 32-bit channel is 2. So, a 128-bit card is going to have either 4 or 8 chips, which gives you 8 or 16 GiB.

GDDR7 launched at 16 Gigabits (2 GiB) per die, but now there's 24 Gigabit (3 GiB) dies. That's how a 512-bit GPU like the GB102 (as seen on the RTX 5090 and RTX Pro 6000) can be configured for 32 GiB (16x 2 GiB), 48 GiB (16x 3 GiB), 64 GiB (32 x 2 GiB), or 96 GiB (32 x 3 GiB). It takes a big, fat GPU to host that much RAM, due to the limitation on the number of chips per channel. The combination of a big GPU and lots of GDDR memory becomes very expensive and power-hungry.

 

[John Nemesh]

Your point was pointless. These modules are absurdly expensive, probably using stacked dies. GDDR memory is more expensive than DRAM per gigabyte. 128-bit bus limits manufacturers to 8 GB or 16 GB GDDR6, or soon 12 GB or 24 GB GDDR7. 8 GB cards are still fine for some users and games. There are people still using 4-6 GB cards. If you don't like it, don't buy it, and wait for the competition to start using 3 GB GDDR7 modules to put 12 GB at the low-end.

OK bro, nice chat. Bye now.

 

[usertests]

OK bro, nice chat. Bye now.

Until next time when you whine about something unrelated again and get what's coming.