Some Sapphire Radeon RX 550 models now include 640 SPs

[HEXUS]

Quiet update of the series boosts the number of Stream Processors from 512.

Read more.

[scaryjim]

... With their 200MHz slower default Boost Engine Clocks the new 640 SPs RX 550 cards could well offer improved OC headroom though. ...

Seems unlikely to me - I don't see why they'd use a salvage die and not configure it as 512 shaders @ RX 550 clocks unless they couldn't qualify them at those clock speeds. Of course, even if they only overclock as far as the stock clocks or a 512 shader RX 550 the extra shaders will give them a significant advantage...

[Ozaron]

Ah, Voerwatch, my favorite.

[azrael-]

Ah, Voerwatch, my favorite.

Right after that other fan favourite, League of Legneds. Both are, of course, best played with a Corsair K68 keyboard sporting Cheery MX keys.

[Corky34]

Seems unlikely to me - I don't see why they'd use a salvage die and not configure it as 512 shaders @ RX 550 clocks unless they couldn't qualify them at those clock speeds. Of course, even if they only overclock as far as the stock clocks or a 512 shader RX 550 the extra shaders will give them a significant advantage...

I've not looked into it so don't shoot me but could it be because Polaris 21 doesn't allow a reduction down to 512 SP without also disabling something linked to them like the memory channels? I best go find some diagrams showing how Polaris 20/21 are laid out.

[scaryjim]

I've not looked into it so don't shoot me but could it be because Polaris 21 doesn't allow a reduction down to 512 SP without also disabling something linked to them like the memory channels? ...

Seems unlikely - the shaders are (AFAIK) a completely separate block from everything else. That said, IANAE and there could be some other technical reason they can't disable 8 CUs and the clock speed adjustment is purely to balance the performance with the smaller die. A third possibility is that it's TDP-related - AFAIK the larger silicon area will generate more heat/draw more power even when it's inactive, so it might qualify at a lower TDP running 640 shaders at a lower clock speed. However, if it is power-related you still won't get the OC headroom because you'll hit the cooling and power limits of the cards...

[DanceswithUnix]

Don't need a technical reason. The GTX 1050 was a bit faster than the RX 550, adding a few shaders should close that up?

Edit: 550X would have been a better name though.

[Corky34]

Seems unlikely...

I struggled to even identify a stream processor on all the block diagrams i looked at, that's probably me but they don't make it very clear what constitutes an SP, i think it's an individual core but is a core a single CU (compute unit) or is each CU made up of multiple (8 or 4) cores?

[scaryjim]

Edit: 550X would have been a better name though.

FWIW, Going by their own announced naming scheme it should be an RX555 if it's intended to be a different product. With the reduction in both GPU and memory clocks I suspect it's meant to perform near-identically to the 550, which is why it doesn't have a different name.

I struggled to even identify a stream processor on all the block diagrams i looked at, that's probably me but they don't make it very clear what constitutes an SP ...

That's because they're not really independent units. A CU is a block of 64 SPs, and since GCN it appears they've been divided into 4 blocks of 16 SPs. I don't think that's changed since GCN was introduced (compare the CU blocks of the RX 560 (https://wccftech.com/amd-radeon-rx-5...-notebook-gpu/) with the CU diagram from the 7970 (https://hexus.net/tech/reviews/graph...70-3gb/?page=2). The Hexus write up of GCN from the 7970 launch is pretty clear and worth a read, but essentially the move from VLIW meant a switch from lots of narrower processing units to fewer, but much wider, ones. I'm pretty sure they only mention shader counts at all because marketing like big numbers.

With these cards what's being turned off is CUs, not individual shaders. To get down to 640 shaders they turn off 6 CU blocks from the full die. Those will usually be spread across the die depending on where the silicon defects are. If any particular CU block was tied to other functionality that'd be a significant point of failure - if a silicon defect hits that your entire die is junk. If there is no tie between any particular CU block and other functionality, then you should be able to turn off as many CUs as you want and still have an otherwise fully functioning GPU...

[DanceswithUnix]

FWIW, Going by their own announced naming scheme it should be an RX555 if it's intended to be a different product. With the reduction in both GPU and memory clocks I suspect it's meant to perform near-identically to the 550, which is why it doesn't have a different name.

The slower clocked RAM looks interesting. Perhaps they can't get 1750MHz RAM in quantity, the bigger die may well come with more L2 cache which could help cover the 1500MHz GDDR5 deficit.

Or perhaps they have a bunch of these chips kicking around to get shot of and a looming supply of 12nm parts coming up.

Edit:

Ah, Voerwatch, my favorite.

Is that a "DriectX12" title?

[scaryjim]

... perhaps they have a bunch of these chips kicking around to get shot of and a looming supply of 12nm parts coming up. ...

Sounds pretty legitimate to me. Also (iirc) none of the extant Polaris cards run less than 7Gbps GDDR5, but a number of the R9 cards did run slower RAM - it could well be that they've got a punch of spare Polaris chips, a bunch of 6Gbps GDDR5 with no products to put them on, and just gone "ah, just slap 'em together lads, I'm sure we can get something worthwhile out of it". It wouldn't be the first time that a bunch of left over chips from a previous generation were found a use on a "new" Sapphire card (the one that sticks out to me was when they randomly decided to do a HD 4670 with GDDR4 chips left over from the 3870...)

[JaymondoGB]

I think you can actually get these as they arent powerful enough for the miners or speculators.

[chinf]

Sounds pretty legitimate to me. Also (iirc) none of the extant Polaris cards run less than 7Gbps GDDR5, but a number of the R9 cards did run slower RAM - it could well be that they've got a punch of spare Polaris chips, a bunch of 6Gbps GDDR5 with no products to put them on, and just gone "ah, just slap 'em together lads, I'm sure we can get something worthwhile out of it". It wouldn't be the first time that a bunch of left over chips from a previous generation were found a use on a "new" Sapphire card (the one that sticks out to me was when they randomly decided to do a HD 4670 with GDDR4 chips left over from the 3870...)

Sounds likely - from some Q&A in the recent financial results it seems that they're not short of GPU silicon but are constrained by GDDR availability. With lack of availability of all DRAM, it may be possible to keep to price points by trading off lower spec memory for better spec core silicon.

[CAT-THE-FIFTH]

Sounds pretty legitimate to me. Also (iirc) none of the extant Polaris cards run less than 7Gbps GDDR5, but a number of the R9 cards did run slower RAM - it could well be that they've got a punch of spare Polaris chips, a bunch of 6Gbps GDDR5 with no products to put them on, and just gone "ah, just slap 'em together lads, I'm sure we can get something worthwhile out of it". It wouldn't be the first time that a bunch of left over chips from a previous generation were found a use on a "new" Sapphire card (the one that sticks out to me was when they randomly decided to do a HD 4670 with GDDR4 chips left over from the 3870...)

The RX470 used 1650MHZ GDDR5 as standard,the XBox One X uses 1700MHZ GDDR5 and a quick check of the PS4 PRO RAM chips,suggests its running 1750MHZ GDDR5.

I suspect,there is quite strong demand for the 1650MHZ~1750MHZ GDDR5 for the consoles,so it makes more sense for a cheaper card for AMD to use slower RAM which is in less demand.

Another thing is the RX550 has the same memory bandwidth as a RX460 and RX560 which are much faster:

http://www.tomshardware.com/reviews/...-2gb,5034.html

So in theory it has more memory bandwidth than required - the GT1030 uses the same speed RAM as this new SKU,so I suspect it should not be any slower.