AMD set to release Radeon RX 6600 XT for $379

[kompukare]

$379 AMD GPU card using a 128 bit connection. UM WHY? Plus i not going to buy a $379 GPU card that only used 128 bit connection. Sound like they cutting the capability of their card in HALF.
Too bad AMD you just lost another customer.


At $379.00 USA dollars and they only using a 128 bit connection. RIP OFF.

The price is poor, but 128 bit hardly the most important metric.

Otherwise if bus width was so important, everyone would have bought AMD's HBM consumer cards.

As well as what DanceswithUnix said, a narrower GDDR bus uses less energy and RDNA2 with the InfinityCache is very concerned about energy usage.

Navi 23 would make a good laptop GPU.

[CAT-THE-FIFTH]

$379 AMD GPU card using a 128 bit connection. UM WHY? Plus i not going to buy a $379 GPU card that only used 128 bit connection. Sound like they cutting the capability of their card in HALF.
Too bad AMD you just lost another customer.


At $379.00 USA dollars and they only using a 128 bit connection. RIP OFF.

With RDNA2 you can't just look at bus width anymore - there is also a large cache in the GPU itself. Its why the 256 bit bus higher end AMD RDNA2 GPUs can compete with the best Nvidia has to offer. Its a role reversal of Nvidia doing the same to AMD,ie, GTX980 vs R9 390.

The bigger issue with the RX6600XT is the reduction in cache amounts. The RX6700XT has 20% more shaders,but 3 times the amount of cache(96MB vs 32MB) so it does quite decently at qHD and even 4K. The issue with only 32MB of cache,is I can see the RX6600XT falling behind at qHD. It will be interesting to see the relative 1080p and qHD results for the RX6600XT vs the RTX3060 and RTX3060TI.

In laptops,the RX6600XT might have an advantage over the RTX3060. The laptop RTX3060 GPUs actually have more shaders,so the performance might be closer than on desktop,but they tend to only ship with 6GB of VRAM. Plus the lower power will most likely be advantageous in thinner designs for the AMD GPU. I would imagine the RX6600XT in laptops is going to very competitive with what Nvidia has. On desktop the pricing really doesn't help it sadly.

[DanceswithUnix]

As well as what DanceswithUnix said, a narrower GDDR bus uses less energy and RDNA2 with the InfinityCache is very concerned about energy usage.

Navi 23 would make a good laptop GPU.

If they cared that much about energy I think they would have put a bigger cache on there to avoid those expensive off chip accesses. I mean, driving fewer ram chips over fewer lines will indeed help with power, but that and the smaller cache really smack of cost reduction to me.

Navi 23 would make a good £200 GPU

Edit: Having said that, if they really are getting on par with a 1080ti performance with a 128 bit bus when that thing has a 384 bit bus, well that's quite impressive really.

[CAT-THE-FIFTH]

If they cared that much about energy I think they would have put a bigger cache on there to avoid those expensive off chip accesses. I mean, driving fewer ram chips over fewer lines will indeed help with power, but that and the smaller cache really smack of cost reduction to me.

It does,because its only 1/3 the amount of cache compared to the RX6700XT. It looks pretty much a 1080p GPU to me,but because of all the issues,AMD has opportunistically pushed pricing up. It was the same with the RX5700XT which was originally branded RX680/RX690(apparently).

[Xlucine]

An APU with v-cache would be really great

If they cared that much about energy I think they would have put a bigger cache on there to avoid those expensive off chip accesses. I mean, driving fewer ram chips over fewer lines will indeed help with power, but that and the smaller cache really smack of cost reduction to me.

Navi 23 would make a good £200 GPU

Edit: Having said that, if they really are getting on par with a 1080ti performance with a 128 bit bus when that thing has a 384 bit bus, well that's quite impressive really.

Part of that is going GDDR5X to GDDR6, for a 45% increase in data rates per pin

[CAT-THE-FIFTH]

An APU with v-cache would be really great



Part of that is going GDDR5X to GDDR6, for a 45% increase in data rates per pin

Its still just over half of the memory bandwidth on the RX6600XT. However,its hard to say whether at qHD and 4K it will fall behind further.

[DanceswithUnix]

An APU with v-cache would be really great

It so would. I'm sure AMD will do it for high end laptop use, I worry they would consider it muddying the desktop market but that would make a really nice small server chip.

Part of that is going GDDR5X to GDDR6, for a 45% increase in data rates per pin

Well yeah, time has moved on. Just we get this feeling that nothing changes in GPUs apart from the price eternally going up, Then you see AMD getting similar performance to an old high end GPU using a third of the memory traces.
But then I look at this compared to the similar $399 Vega 56 launch price some 4 years ago, and it doesn't look so impressive.

[[GSV]Trig]

$379 AMD GPU card using a 128 bit connection. UM WHY? Plus i not going to buy a $379 GPU card that only used 128 bit connection. Sound like they cutting the capability of their card in HALF.
Too bad AMD you just lost another customer.


At $379.00 USA dollars and they only using a 128 bit connection. RIP OFF.

4 posts in near on 5000 days, and that's the best excuse to not buy a card you can come up with...

[kompukare]

It so would. I'm sure AMD will do it for high end laptop use, I worry they would consider it muddying the desktop market but that would make a really nice small server chip.

Just need to design an APU which has the traces for 3D cache, then produce the main APU in chiplet like quantities and sell fancier ones with 1 or more slices of 3D cache. "Just", but that is where things will eventually head. TSMC better have tons of packaging lines.

[DanceswithUnix]

Just need to design an APU which has the traces for 3D cache, then produce the main APU in chiplet like quantities and sell fancier ones with 1 or more slices of 3D cache. "Just", but that is where things will eventually head. TSMC better have tons of packaging lines.

I'm sure the next APU design will allow for it. OFC, they don't have to actually fit the extra cache when they package the chips up.

[badass]

If they cared that much about energy I think they would have put a bigger cache on there to avoid those expensive off chip accesses. I mean, driving fewer ram chips over fewer lines will indeed help with power, but that and the smaller cache really smack of cost reduction to me.

Navi 23 would make a good £200 GPU

I bet that's what was planned pre pandemic related shortage!

I'm sure the next APU design will allow for it. OFC, they don't have to actually fit the extra cache when they package the chips up.

Just need to design an APU which has the traces for 3D cache, then produce the main APU in chiplet like quantities and sell fancier ones with 1 or more slices of 3D cache. "Just", but that is where things will eventually head. TSMC better have tons of packaging lines.

Sounds perfectly viable. With 64MB SRAM dies being 36mm^2 on TSMC's 7nm process, it looks like they cost less than $20 for 64MB (probably closer to $10)
I doubt that the TSV's will use up much die space. That could make a pretty potent APU as AMD's APU's have always been bandwidth starved. (not to say the OEM's that equip their laptops with only 1 DIMM to make the situation worse!)
It's a shame this tech wasn't viable for the RX 6600XT launch. Forger the on die 32MB Infinity cache - just add 64MB of 3d Cache.

[kompukare]

Sounds perfectly viable. With 64MB SRAM dies being 36mm^2 on TSMC's 7nm process, it looks like they cost less than $20 for 64MB (probably closer to $10)
I doubt that the TSV's will use up much die space. That could make a pretty potent APU as AMD's APU's have always been bandwidth starved. (not to say the OEM's that equip their laptops with only 1 DIMM to make the situation worse!)
It's a shame this tech wasn't viable for the RX 6600XT launch. Forger the on die 32MB Infinity cache - just add 64MB of 3d Cache.

It's actually not just the density, as SRAM requires less steps too. Still everything is expensive at the moment, but less steps should also mean the wafer doesn't spend quite 3 months in the fab.

[CAT-THE-FIFTH]

Sounds perfectly viable. With 64MB SRAM dies being 36mm^2 on TSMC's 7nm process, it looks like they cost less than $20 for 64MB (probably closer to $10)
I doubt that the TSV's will use up much die space. That could make a pretty potent APU as AMD's APU's have always been bandwidth starved. (not to say the OEM's that equip their laptops with only 1 DIMM to make the situation worse!)
It's a shame this tech wasn't viable for the RX 6600XT launch. Forger the on die 32MB Infinity cache - just add 64MB of 3d Cache.

There will be additional packaging derived costs too?? AFAIK,there is no indication anything under a Ryzen 9 will get it,which is a shame.

[DanceswithUnix]

There will be additional packaging derived costs too?? AFAIK,there is no indication anything under a Ryzen 9 will get it,which is a shame.

I suspect you only get additional packaging costs if you stick the cache die to the top. Otherwise, it's just a chip in a carrier as before.

[CAT-THE-FIFTH]

I suspect you only get additional packaging costs if you stick the cache die to the top. Otherwise, it's just a chip in a carrier as before.

I would assume they were talking about that arrangement? Although the APUs,do seem to be engineered for minimal costs - the lack of L3 cache does seem to noticeably affect CPU performance!

[Xlucine]

Eventually intel will catch up in APU GPU performance, and as they're both up against the same DDR bandwidth limitation AMD will need something so they can pull ahead (assuming AMD cares about having the fastest APU that much, and only until intel puts a cache die next to their compute die).

32MB is enough to add v-cache, but on smaller cache sizes (e.g. the current APUs) I can see difficulties. AIUI the TSVs are in the cache on ryzen, so with less cache you either need to squeeze more TSVs per MB of cache or figure out how to squeeze TSVs in amongst the compute stuff