Wednesday, October 21st 2020
AMD Radeon RX 6000 Series Specs Leak: RX 6900 XT, RX 6800 XT, RX 6700 Series
AMD's Radeon RX 6000 series graphics cards, based on the RDNA2 graphics architecture, will see the introduction of the company's first DirectX 12 Ultimate graphics cards (featuring features such as real-time raytracing). A VideoCardz report sheds light on the specifications. The 7 nm "Navi 21" and "Navi 22" chips will power the top-end of the lineup. The flagship part is the Radeon RX 6900 XT, followed by the RX 6800 XT and RX 6800; which are all based on the "Navi 21." These are followed by the RX 6700 XT and RX 6700, which are based on the "Navi 22" silicon.
The "Navi 21" silicon physically features 80 RDNA2 compute units, working out to 5,120 stream processors. The RX 6900 XT maxes the chip out, enabling all 80 CUs, and is internally referred to as the "Navi 21 XTX." Besides these, the RX 6900 XT features 16 GB of GDDR6 memory across a 256-bit wide memory interface, and engine clocks boosting beyond 2.30 GHz. The next SKU in AMD's product stack is the RX 6800 XT (Navi 21 XT), featuring 72 out of 80 CUs, working out to 4,608 stream processors, the same 16 GB 256-bit GDDR6 memory configuration as the flagship, while its engine clocks go up to 2.25 GHz.
A notch below the RX 6800 XT is the RX 6800 (Navi 21 XL), which cuts down the "Navi 21" further, giving it 64 compute units or 4,096 stream processors; the very same 16 GB of 256-bit GDDR6 memory interface, and up to 2.15 GHz engine clocks. The RX 6900 XT, along with the RX 6800 series, will be announced in the October 28 presser.
The next chip AMD is designing is the 7 nm "Navi 22" silicon, which features 40 compute units. On paper, this count looks similar to that of the "Navi 10," and it remains to be seen if this is a re-badge or a new silicon based on RDNA2. The RX 6700 XT maxes this chip out, featuring 40 CUs or 2,560 stream processors; while the RX 6700 features fewer CUs (possibly 36). The interesting thing about these two is their memory configuration—12 GB of 192-bit GDDR6.
Source:
VideoCardz
The "Navi 21" silicon physically features 80 RDNA2 compute units, working out to 5,120 stream processors. The RX 6900 XT maxes the chip out, enabling all 80 CUs, and is internally referred to as the "Navi 21 XTX." Besides these, the RX 6900 XT features 16 GB of GDDR6 memory across a 256-bit wide memory interface, and engine clocks boosting beyond 2.30 GHz. The next SKU in AMD's product stack is the RX 6800 XT (Navi 21 XT), featuring 72 out of 80 CUs, working out to 4,608 stream processors, the same 16 GB 256-bit GDDR6 memory configuration as the flagship, while its engine clocks go up to 2.25 GHz.
The next chip AMD is designing is the 7 nm "Navi 22" silicon, which features 40 compute units. On paper, this count looks similar to that of the "Navi 10," and it remains to be seen if this is a re-badge or a new silicon based on RDNA2. The RX 6700 XT maxes this chip out, featuring 40 CUs or 2,560 stream processors; while the RX 6700 features fewer CUs (possibly 36). The interesting thing about these two is their memory configuration—12 GB of 192-bit GDDR6.
191 Comments on AMD Radeon RX 6000 Series Specs Leak: RX 6900 XT, RX 6800 XT, RX 6700 Series
Edit: Do we even know this is a data cache and not an instruction cache?
I don't doubt AMD has done something smart here, but you guys are simply expecting too much from a L1 cache. The cache my be more effective than it is on a CPU, but bridging hundreds of GB/s of bandwidth difference? I have to see it to believe it.
NAVI10 currently have 512 kb of L1 (128 kb per 10 CU block). if AMD keep the same size per block of 10 CU, the available L1 cache will go from 128 kb to 1 MB for full NAVI21 witch is quite a big increase. If they match the cache size Per CU, they could have around 8 MB of L2 cache.
how much that will help, i don't know. No matter what people say, only the benchmark will really reveal the final information. We may be discussing about stuff that have marginal impact on the real performance.
Or had NV priced card that is 20-30% faster than $1200 card at around $999.
I wonder why didn't they do it.
Oh wait, this can't be the reason, can it:
Amd/comments/jg21aiNeither honestly describes what the process really is (e.g. 7nm TSMC means 22x22nm transistor, while 14nm Intel means 24x24nm), but Samsung's 8nm is even further from truth, so there is that.
We could say AMD want to push it to compete with Nvidia, but on the other end, why would they limit the frequency if people are able to OC these cards with custom bios and stuff? They better just sell them with clock as fast as they can do.
So Navi10's 2GHz game clock becomes 2.15GHz without even changing the TDP or accounting for any architectural improvements that always get made to improve clockspeeds between generations.
If its faster, best believe AMD will price it accordingly.
Pricing on NV cards is BS, street price on 3070 is around 700 Euros, 3080 is missing in action on top of that.
NV won't be able to recover with Ampere on Samsung 8nm, it an un-redeemable clusterf*ck.
How do you know that when the 3070 isnt even out yet?
Nvidia will probably counter next year with Super and Ti variants using TSMC 7nm.
Maybe a 3080 super with 20GB and 6% better than regular 3080
That 3090 is a fucking joke for gamers.
I wouldn’t call it a nightmare (it was until April) but I still have some freezes and black screens, from time to time.
Eespecially if these new cards will actually be available (and I’m not entirely sure about that).
AMD is great at hardware production, but not in the software development
Simple.
Leaks are "saying" about 128MB total cache in RNDA2 flagship at least. We cant really confirm it though.
Rough estimation
If we accept the biggest Navi21 die is 536mm² then:
5700XT (40CU + SOC) = 251mm²
From pictures of the die about 150mm² is the 40CUs and the other 100mm² for SOC stuff.
80CUs = 300mm², but with enhanced 7NP node (+10~15% density) this could result a 250~270mm² area for 80CUs + 100~150mm² for SOC = 80CU + SOC = 350~420mm².
This leaving another 115~185mm² up to 536 for cache only.
Remember that Navi21 has the same 256bit bandwidth so it does not need a more complex mem cotroller (no extra space on die from Navi10).
So the 100+MB cache can be a thing on the Navi21 die...
I'm not trying to defend nvidia, but the situation they are in it's not intentional just a result of some bad calls like manufacturing and ddr6x. I'm sure if it was up to them they would like to be able to make and sell more. One for obvious reason $$$ and second the more they sell now the less market is left for AMD. But I highly doubt that AMD availability will be much better. And to remind those who didn't know amd and nvidia have been sued for having agreement between and price fixing. So I don't expect any price undercutting or anything, the most we can hope for is buy at regular price.