Soft channel bits are the memory a UE keeps for HARQ. When a transport block fails its CRC, the UE does not discard what it received. It stores the soft values, and it combines them with the retransmission. The size of this memory is a UE category parameter, and 36.212 turns it into a soft buffer size per transport block and then per code block.

I'll start from the category table, then follow one number down to a single code block. On the way you will see why the same UE gets a smaller buffer per transport block when it is configured for spatial multiplexing.

• What is the Total Number of Soft Channel Bits ?
• How is the Soft Buffer per Transport Block calculated ?
• How is the Soft Buffer split into Code Blocks ?
• Reference

What is the Total Number of Soft Channel Bits ?

Let's first see where the number comes from. The UE reports its category, and the category fixes the total soft buffer that the eNB can assume. The eNB needs this value, because the rate matching at the transmitter has to fit what the receiver can store.

The picture below puts the two ends of that chain side by side. At the top is 36.306 Table 4.1-1, with the Total number of soft channel bits column highlighted. An arrow carries this value, N_soft, into the N_IR formula of 36.212 clause 5.1.4.1.2. The notes under the formula list the values that each factor can take.

N_soft comes from the UE category. The formula divides it by K_C, K_MIMO and min(M_{DL_HARQ}, M_limit) to give the soft buffer of one transport block.

• The table rows show 250368 bits for Category 1, 1237248 bits for Category 2 and Category 3, 1827072 bits for Category 4, 3667200 bits for Category 5 and 7308288 bits for Category 12. The row of dots hides Category 6 to Category 11.
• N_IR is the soft buffer size for the transport block, as the label on the left says.
• K_MIMO is 2 for transmission mode 3, 4, 8, 9 and 10, and 1 otherwise.
• M_{DL_HARQ} is 8 for FDD. For TDD it is 4 to 15, from 36.213 Table 7-1.
• M_limit is fixed at 8.
• K_C is listed as 5 for N_soft = 35982720, 3 for N_soft = 3654144 with no more than two spatial layers, and 1 otherwise.

I checked the table part against 36.306 v19.3.0, and the values in the picture are unchanged. The hidden rows are Category 6 and Category 7 with 3654144 bits, Category 8 with 35982720 bits, Category 9 and Category 10 with 5481216 bits, and Category 11 with 7308288 bits.

One value in the K_C note needs a correction. In 36.212 v19.3.0, N_soft = 3654144 with no more than two spatial layers gives K_C = 2, not 3. You can check this with the numbers. 3654144 is exactly twice the Category 4 value of 1827072, so K_C = 2 gives a two-layer Category 6 or Category 7 UE the same per transport block buffer as Category 4.

Also note which category N_soft comes from. A UE usually reports more than one category, for example Category 6 together with Category 4. 36.212 picks the category field by configuration. The UE uses the ue-Category-v1020 value only when it is configured with transmission mode 9 or 10, or with transmission mode 3 or 4 and maxLayersMIMO-r10 set to fourLayers. In other cases it uses the category without suffix, and the newer fields such as ue-CategoryDL-r12 follow the same pattern.

Compare Category 2 and Category 3 in the table. Both have 1237248 soft channel bits, but Category 3 receives twice as many DL-SCH bits per TTI. So a Category 3 UE has about half the soft memory per received bit, and the rate matching limit in the last section matters more for it.

• N_soft is a category value : 36.306 Table 4.1-1 gives it, and the values in the picture still match v19.3.0.
• Several category fields exist : the configured transmission mode and layers decide which one sets N_soft.
• The picture has one stale number : K_C for N_soft = 3654144 with up to two layers is 2 in the current 36.212.

How is the Soft Buffer per Transport Block calculated ?

The total N_soft is shared by all DL HARQ processes. With spatial multiplexing it is also shared by two transport blocks in each process. So 36.212 divides it before any code block sees it, and each factor in the denominator stands for one way of sharing the memory.

For DL-SCH and PCH, the formula is N_IR = floor( N_soft / ( K_C x K_MIMO x min(M_{DL_HARQ}, M_limit) ) ). The current version also multiplies this by a coefficient for short TTI. The coefficient is 1 for subframe duration and 0.5 for slot duration, and subslot duration has its own value.

Let's go through the factors as 36.212 v19.3.0 defines them. The list of K_C cases is much longer than the picture shows, because each new category with 256QAM or 1024QAM added its own case.

• K_C is 5 for N_soft = 35982720 or 47431680, and 32 for N_soft = 303562752.
• For N_soft = 14616576, K_C is 3/2 with altCQI-Table-r12 and 2 otherwise.
• For N_soft = 19488768, K_C is 8/5 with altCQI-Table-1024QAM-r15, 2 with altCQI-Table-r12, and 8/3 otherwise.
• For N_soft = 7308288 with altCQI-Table-r12, K_C is 3 with up to two layers and 3/2 otherwise. With altCQI-Table-1024QAM-r15 it is 12/5 or 6/5 in the same way.
• For N_soft = 3654144 with up to two layers, K_C is 2. In every other case K_C is 1.
• K_MIMO is 2 for an initial PDSCH transmission of subframe duration in transmission mode 3, 4, 8, 9 or 10. It is 1 otherwise, and it is always 1 for a BL/CE UE and for slot or subslot duration.
• M_{DL_HARQ} is the maximum number of DL HARQ processes from 36.213 clause 7. It is 8 for FDD in the normal case, and 4 to 15 for TDD from Table 7-1.
• M_limit is a constant equal to 8.

Now let's put numbers in. The table below uses a Category 4 UE with N_soft = 1827072, so K_C = 1. Only the transmission mode and the duplex mode change from row to row.

The second row shows the cost of spatial multiplexing. The UE now holds two transport blocks per HARQ process, so each one gets half the buffer. The third row shows what M_limit does. TDD configuration 5 has 15 DL HARQ processes, but the divisor stops at 8. So 15 x N_IR is almost twice N_soft, and the formula does not reserve a full N_IR for every process in this case. In the last row there are only 4 processes, so each transport block gets twice the FDD buffer.

Carrier aggregation adds one more rule. 36.213 v19.4.0 clause 7.1.8 describes how the UE stores soft channel bits when it is configured with more than one serving cell or with a SCG. The storage is then shared over the number of configured serving cells, and the UE gives priority to the soft bits with the lower index k in the circular buffer.

• Three factors share the buffer : K_C for the category case, K_MIMO for two transport blocks, and the HARQ process count.
• M_limit caps the divisor at 8 : a TDD configuration with more processes does not get a smaller N_IR.
• Spatial multiplexing halves N_IR : a Category 4 UE drops from 228384 to 114192 bits per transport block.

How is the Soft Buffer split into Code Blocks ?

N_IR is still a buffer for the whole transport block. The turbo coder works on code blocks, so the last step divides N_IR again. This step decides how much of each code block's circular buffer the eNB is allowed to use.

Each code block of size K goes through the turbo coder at rate 1/3. After the sub-block interleavers, the systematic bits and the two parity streams fill a circular buffer of length K_w = 3 x K_Π. Here K_Π is the interleaver size, which is K + 4 rounded up to a multiple of 32. The size N_cb of the part in use depends on the transport channel.

• For DL-SCH and PCH, N_cb = min( floor(N_IR / C), K_w ), where C is the number of code blocks from 36.212 clause 5.1.2.
• For UL-SCH, N_cb = K_w. The receiver is the eNB, and no UE category limit applies.
• For a Category 0 UE, N_cb is always K_w for DL-SCH with SI-RNTI or RA-RNTI and for PCH.

So the limit only matters when floor(N_IR / C) is smaller than K_w. Let's take the largest Category 4 transport block, 75376 bits, and work it through. With 24 CRC bits per code block and the 6144 bit limit, the block is split into C = 13 code blocks of K = 5824 bits. So K_Π = 5856 and K_w = 17568.

The first row gives an exact match. The per code block share is exactly K_w, so the full rate 1/3 mother code fits in the buffer. In the second row the UE can store only half of it. The eNB then selects bits from the first 8784 positions of each circular buffer, and the four redundancy versions start at points spread over N_cb rather than over K_w.

This is called limited buffer rate matching. With N_cb = 8784 and K = 5824, the lowest code rate the eNB can reach is about 5824 / 8784, or 0.66. A retransmission therefore adds fewer new parity bits than it would with a full buffer. In practice, a Category 4 UE with two large transport blocks gains less from incremental redundancy than the same UE with one transport block.

You can see the same logic in the category table. A higher category usually carries a larger N_soft because it receives more bits per TTI, not because it keeps more redundancy per bit. When you compare HARQ gains between two UEs, check N_cb for the actual transport block size first.

• N_cb is the usable circular buffer : it is the smaller of the per code block share of N_IR and K_w.
• Only the downlink is limited : UL-SCH always uses the full K_w, because the eNB is the receiver.
• A full buffer is not guaranteed : a Category 4 UE in transmission mode 3 or 4 keeps only half of K_w for its largest transport block, which limits the code rate to about 0.66.

Reference

• 3GPP TS 36.306 v19.3.0 - clause 4.1, Table 4.1-1 Downlink physical layer parameter values set by the field ue-Category
• 3GPP TS 36.212 v19.3.0 - clause 5.1.2 Code block segmentation, clause 5.1.4.1 Rate matching for turbo coded transport channels
• 3GPP TS 36.213 v19.4.0 - clause 7 Physical downlink shared channel related procedures, Table 7-1, clause 7.1.8 Storing soft channel bits